AI and science frontiers · 2026-11-11

The Embodied First Exam connects general coding agents to 47 simulated robots and lets each agent choose direct control, code to control, or training a policy, while the real-robot-like Standard Setting supplies only sensor observations and, by default, no resets. Among eight agents, GPT-6 Astra in Codex solves 34% of the Standard tests on 456 core tasks, while on the same 391 tasks it reaches 77% in the Privileged Setting, which exposes the full simulator state and allows unlimited resets. The exam also labels the stage at which failed attempts go wrong, so failures can be compared across settings and agents.

The Embodied First Exam connects general coding agents to 47 simulated robots and lets each agent choose direct control, code to control, or training a policy, while the real-robot-like Standard Setting supplies only sensor observations and, by default, no resets. Among eight agents, GPT-6 Astra in Codex solves 34% of the Standard tests on 456 core tasks, while on the same 391 tasks it reaches 77% in the Privileged Setting, which exposes the full simulator state and allows unlimited resets. The exam also labels the stage at which failed attempts go wrong, so failures can be compared across settings and agents.

Large-model training corpora often cover public benchmarks, so evaluators can no longer assume a fresh test set. This work formalizes evaluability without assuming the test set was held out and without constraining the model or its training algorithm, showing that accurate evaluation from model outputs on the test set alone is feasible if and only if the distribution is close to small support, with sample-complexity bounds tight up to logarithmic factors; allowing extra queries to the model, via self-correctors from program testing, widens the evaluable distributions.

Large-model training corpora often cover public benchmarks, so evaluators can no longer assume a fresh test set. This work formalizes evaluability without assuming the test set was held out and without constraining the model or its training algorithm, showing that accurate evaluation from model outputs on the test set alone is feasible if and only if the distribution is close to small support, with sample-complexity bounds tight up to logarithmic factors; allowing extra queries to the model, via self-correctors from program testing, widens the evaluable distributions.

Reward guidance steers a frozen diffusion or flow model at sampling time by following the gradient of a value function, but both common estimates of that gradient, differentiating a one-step denoised estimate and a learned value model, are least accurate at high noise, where the coarse structure of a sample is decided. VaLoP evaluates the value at a less noisy lookahead state one sampler step ahead and pulls its gradient back with one vector-Jacobian product; across SD3.5-Medium, SD3.5-Large and FLUX.1-dev with three rewards, VaLoP and its amortized variant VaLoP-Fast raise the target reward above StitchVM and DPS, and VaLoP-Fast attains the highest success rates on LIBERO and RoboCasa.

Reward guidance steers a frozen diffusion or flow model at sampling time by following the gradient of a value function, but both common estimates of that gradient, differentiating a one-step denoised estimate and a learned value model, are least accurate at high noise, where the coarse structure of a sample is decided. VaLoP evaluates the value at a less noisy lookahead state one sampler step ahead and pulls its gradient back with one vector-Jacobian product; across SD3.5-Medium, SD3.5-Large and FLUX.1-dev with three rewards, VaLoP and its amortized variant VaLoP-Fast raise the target reward above StitchVM and DPS, and VaLoP-Fast attains the highest success rates on LIBERO and RoboCasa.

Molecular photochemistry needs entire ground- and excited-state potential energy surfaces, yet solving the electronic Schrödinger equation point by point is expensive, and cheap single-reference methods fail at conical intersections. Transferable deep quantum Monte Carlo optimizes one neural-network wavefunction jointly across many molecular geometries and electronic states, sharing parameters between states and ordering them dynamically by per-geometry energy. On ethylene, the carbon dimer, the methylenimmonium cation and a 96-electron rubredoxin active-site model, it reproduces tailored quantum-chemistry references while cutting cost by up to about two orders of magnitude relative to single-point calculations.

Molecular photochemistry needs entire ground- and excited-state potential energy surfaces, yet solving the electronic Schrödinger equation point by point is expensive, and cheap single-reference methods fail at conical intersections. Transferable deep quantum Monte Carlo optimizes one neural-network wavefunction jointly across many molecular geometries and electronic states, sharing parameters between states and ordering them dynamically by per-geometry energy. On ethylene, the carbon dimer, the methylenimmonium cation and a 96-electron rubredoxin active-site model, it reproduces tailored quantum-chemistry references while cutting cost by up to about two orders of magnitude relative to single-point calculations.

When a camera moves or a controller is replaced, the same command moves the robot in a different direction or by a different magnitude, while a frozen world model still plans under its training-time action convention, so every plan is expressed in the wrong action frame. Library of Hands (LoH) trains nothing and leaves the world model frozen: it uses the frozen predictor to infer the applied action from two successive camera images, estimates the map from commands to applied actions, and stores each identified map in a library for reuse. In closed-loop tests where hands change and come back, LoH raises a frozen planner's success from 28.6% to 62.3%, and loses only 0.6 percentage points when nothing changes.

When a camera moves or a controller is replaced, the same command moves the robot in a different direction or by a different magnitude, while a frozen world model still plans under its training-time action convention, so every plan is expressed in the wrong action frame. Library of Hands (LoH) trains nothing and leaves the world model frozen: it uses the frozen predictor to infer the applied action from two successive camera images, estimates the map from commands to applied actions, and stores each identified map in a library for reuse. In closed-loop tests where hands change and come back, LoH raises a frozen planner's success from 28.6% to 62.3%, and loses only 0.6 percentage points when nothing changes.

Recursive self-improvement had only been closed on post-training, where each trial costs GPU-hours and returns a verifiable score; a frontier pre-training run costs hundreds of thousands of GPU-hours, cannot be repeated, and has no single-score objective. A newly formed team at Shanghai Innovation Institute and Xinxu AI split pre-training into smaller closable loops, using agents for debugging, tiny-SFT and related measures to estimate post-training headroom, and quantified workflows for data tasks, producing the 30B-A3B base model 801 in three months. On 13 base benchmarks it averages 73.13, behind only the larger Qwen3.5-35B-A3B at 76.67.

Recursive self-improvement had only been closed on post-training, where each trial costs GPU-hours and returns a verifiable score; a frontier pre-training run costs hundreds of thousands of GPU-hours, cannot be repeated, and has no single-score objective. A newly formed team at Shanghai Innovation Institute and Xinxu AI split pre-training into smaller closable loops, using agents for debugging, tiny-SFT and related measures to estimate post-training headroom, and quantified workflows for data tasks, producing the 30B-A3B base model 801 in three months. On 13 base benchmarks it averages 73.13, behind only the larger Qwen3.5-35B-A3B at 76.67.

For coding, math and computer-use agent workloads, Evolving Infra decouples compute allocation between prefill and decode from KV cache management: a node switches phase without a restart, and its KV stays readable from a shared pool. On up to 256 GB300 GPUs it cuts mean end-to-end latency and P95 time to first token by up to 4.1 and 14 relative to NVIDIA Dynamo, and a phase switch takes 6.2 seconds on average against 754 seconds for a Dynamo relaunch.

For coding, math and computer-use agent workloads, Evolving Infra decouples compute allocation between prefill and decode from KV cache management: a node switches phase without a restart, and its KV stays readable from a shared pool. On up to 256 GB300 GPUs it cuts mean end-to-end latency and P95 time to first token by up to 4.1 and 14 relative to NVIDIA Dynamo, and a phase switch takes 6.2 seconds on average against 754 seconds for a Dynamo relaunch.

The macroscopic constitutive law of an inelastic heterogeneous material depends on both loading history and microstructure, and computational homogenization must repeatedly solve a unit-cell problem at every macroscopic quadrature point. For Kelvin–Voigt viscoelasticity, a recurrent neural operator model takes the microstructure as input, and it is proved to approximate the homogenized constitutive law uniformly in time; the sufficient internal-variable dimension grows logarithmically with the inverse relative stress error and with each material contrast. Crystal-plasticity experiments show accuracy improving and then plateauing as that dimension grows.

The macroscopic constitutive law of an inelastic heterogeneous material depends on both loading history and microstructure, and computational homogenization must repeatedly solve a unit-cell problem at every macroscopic quadrature point. For Kelvin–Voigt viscoelasticity, a recurrent neural operator model takes the microstructure as input, and it is proved to approximate the homogenized constitutive law uniformly in time; the sufficient internal-variable dimension grows logarithmically with the inverse relative stress error and with each material contrast. Crystal-plasticity experiments show accuracy improving and then plateauing as that dimension grows.

Remote sensing models have long been built by hand, with experts designing a training recipe for each new task and tuning it against a benchmark. RSI4EO casts a task as two parts: a harness that fixes the data, pretrained weights and compute an agent may use, and a sealed verifier that returns only a score. A general-purpose coding agent then writes the training code, trains models and improves them recursively. On four benchmarks with published results, spanning optical, SAR and multispectral time-series imagery, the agent-built models surpass the state of the art, and with about 1% of the training labels they keep most of that accuracy.

Remote sensing models have long been built by hand, with experts designing a training recipe for each new task and tuning it against a benchmark. RSI4EO casts a task as two parts: a harness that fixes the data, pretrained weights and compute an agent may use, and a sealed verifier that returns only a score. A general-purpose coding agent then writes the training code, trains models and improves them recursively. On four benchmarks with published results, spanning optical, SAR and multispectral time-series imagery, the agent-built models surpass the state of the art, and with about 1% of the training labels they keep most of that accuracy.

Generating symmetric protein complexes previously required a forward pass over the entire assembly, so time and memory grew with assembly size for higher-order symmetry, while training data on symmetric structures stayed sparse. ZeUS recasts symmetric assembly generation as sampling a single canonical subunit in a quotient space, using phantom subunits to preserve the model's internal operations, turning any SE(3)-equivariant backbone generative model into a symmetric assembly generator without retraining. On FrameDiff, FoldFlow and RFdiffusion, design quality matches full-complex inference while forward-pass time and memory fall with the order of the symmetry group, and it generates symmetric complexes over 10,000 residues on a single A100.

Generating symmetric protein complexes previously required a forward pass over the entire assembly, so time and memory grew with assembly size for higher-order symmetry, while training data on symmetric structures stayed sparse. ZeUS recasts symmetric assembly generation as sampling a single canonical subunit in a quotient space, using phantom subunits to preserve the model's internal operations, turning any SE(3)-equivariant backbone generative model into a symmetric assembly generator without retraining. On FrameDiff, FoldFlow and RFdiffusion, design quality matches full-complex inference while forward-pass time and memory fall with the order of the symmetry group, and it generates symmetric complexes over 10,000 residues on a single A100.

Excited-state energies, unlike ground-state energies, are not sums of atomic contributions, so multistate readouts that pool atomic terms drift when molecules grow larger or two molecules appear together. LEH-MACE instead assembles a symmetric matrix from one diagonal block per atom and coupling blocks for atom pairs, and takes its eigenvalues as the singlet excitation energies, so the spectrum of two molecules beyond the cutoff is exactly the union of their spectra. On TD-DFT data for about 303,000 geometries of 106,933 molecules, the advantage grew away from the training distribution, and only this construction kept single-molecule accuracy for molecule pairs.

Excited-state energies, unlike ground-state energies, are not sums of atomic contributions, so multistate readouts that pool atomic terms drift when molecules grow larger or two molecules appear together. LEH-MACE instead assembles a symmetric matrix from one diagonal block per atom and coupling blocks for atom pairs, and takes its eigenvalues as the singlet excitation energies, so the spectrum of two molecules beyond the cutoff is exactly the union of their spectra. On TD-DFT data for about 303,000 geometries of 106,933 molecules, the advantage grew away from the training distribution, and only this construction kept single-molecule accuracy for molecule pairs.

Interleaved multimodal reasoning inserts continuous visual latent blocks between text spans, but direct perturbation shows that dropping, replacing or shuffling these blocks barely changes the model's own answer probability, while the same perturbations to the textual reasoning trace hurt far more. The authors call this mental simulation collapse and introduce AnchorSim, which during training contrasts predictions with the visual pathway against predictions with it masked, forcing the latents to carry answer-relevant information. Built on Qwen2.5-VL-7B, AnchorSim-7B averages 62.4 across nine visual reasoning benchmarks, 13.5 points above the base model, and visual perturbations become consequential again.

Interleaved multimodal reasoning inserts continuous visual latent blocks between text spans, but direct perturbation shows that dropping, replacing or shuffling these blocks barely changes the model's own answer probability, while the same perturbations to the textual reasoning trace hurt far more. The authors call this mental simulation collapse and introduce AnchorSim, which during training contrasts predictions with the visual pathway against predictions with it masked, forcing the latents to carry answer-relevant information. Built on Qwen2.5-VL-7B, AnchorSim-7B averages 62.4 across nine visual reasoning benchmarks, 13.5 points above the base model, and visual perturbations become consequential again.

Meta released Triton Low-Level Extensions (TLX), which lets engineers explicitly orchestrate warp groups, asynchronous data movement and cluster execution while keeping Triton's programming model. Modern GPU features such as warp specialization, tensor memory and cluster execution were previously either hidden behind compiler abstractions or required large amounts of architecture-specific CUDA code. TLX is already in production across Meta's recommendation training, recommendation inference and large-scale LLM pretraining, where GEM training reached 20-25% end-to-end MFU on several thousand NVIDIA Blackwell GPUs.

Meta released Triton Low-Level Extensions (TLX), which lets engineers explicitly orchestrate warp groups, asynchronous data movement and cluster execution while keeping Triton's programming model. Modern GPU features such as warp specialization, tensor memory and cluster execution were previously either hidden behind compiler abstractions or required large amounts of architecture-specific CUDA code. TLX is already in production across Meta's recommendation training, recommendation inference and large-scale LLM pretraining, where GEM training reached 20-25% end-to-end MFU on several thousand NVIDIA Blackwell GPUs.

Anthropic had 1,000 agents running an early version of Claude Fable 5.2 redo the formalization of Fermat's Last Theorem without the Prove2Me platform and without a researcher directing the work, coordinating only through Git and a shared SQL database. The team finished in 28.7 hours, using about one-third the output tokens of the earlier formalization, which took a researcher and dozens of agents 11 days, and wrote a proof about one-seventh as long. The authors call it a toy experiment suggesting such agent teams can accomplish long-horizon goals within carefully designed safety boundaries.

Anthropic had 1,000 agents running an early version of Claude Fable 5.2 redo the formalization of Fermat's Last Theorem without the Prove2Me platform and without a researcher directing the work, coordinating only through Git and a shared SQL database. The team finished in 28.7 hours, using about one-third the output tokens of the earlier formalization, which took a researcher and dozens of agents 11 days, and wrote a proof about one-seventh as long. The authors call it a toy experiment suggesting such agent teams can accomplish long-horizon goals within carefully designed safety boundaries.

Individually compliant language model agents can still jointly support a policy violation through messages that are harmless one by one. Collective Dependence Analysis (CDA) measures each message's necessity and sufficiency by removing and retaining it, cutting violations from 93% to 6% in controlled conversations and from 98% to 9% in agent-generated operational ones. It returns both an explanation and a minimal intervention set, though jointly required groups resist exact recovery and removals can also disrupt correct decisions.

Individually compliant language model agents can still jointly support a policy violation through messages that are harmless one by one. Collective Dependence Analysis (CDA) measures each message's necessity and sufficiency by removing and retaining it, cutting violations from 93% to 6% in controlled conversations and from 98% to 9% in agent-generated operational ones. It returns both an explanation and a minimal intervention set, though jointly required groups resist exact recovery and removals can also disrupt correct decisions.

Shared prefix caching makes a cache-hit request reach its first token much faster than a miss, and an attacker can exploit that gap to guess another tenant's prompt token by token. Mirage stops isolating caches by content and instead counts new candidate branches under the same sensitive prefix at the gateway; once the budget is spent, it shifts the whole output stream so a hit looks like a natural miss. On SGLang across Qwen, Llama and DeepSeek models and three timing attacks, attack success falls near the random-guessing floor while median first-token latency on public prefixes stays essentially unchanged.

Shared prefix caching makes a cache-hit request reach its first token much faster than a miss, and an attacker can exploit that gap to guess another tenant's prompt token by token. Mirage stops isolating caches by content and instead counts new candidate branches under the same sensitive prefix at the gateway; once the budget is spent, it shifts the whole output stream so a hit looks like a natural miss. On SGLang across Qwen, Llama and DeepSeek models and three timing attacks, attack success falls near the random-guessing floor while median first-token latency on public prefixes stays essentially unchanged.

Undesired behaviors in language models can re-emerge during later fine-tuning, and standard post-hoc suppression had not been compared on immediacy, durability and capability together. PEARL adds adversarial prefixes, continuations from intermediate states of cached misaligned trajectories, to GRPO's on-policy rollouts so the policy practices recovering from partly wrong states. Across Qwen3-4B and GPT-OSS-20B and two organism types, PEARL suppresses strongly while preserving task accuracy, with the clearest gains on the broad-trigger organism and GPT-OSS-20B; on Qwen3-4B with the narrow-trigger organism, SFT is more durable under heavy targeted retraining.

Undesired behaviors in language models can re-emerge during later fine-tuning, and standard post-hoc suppression had not been compared on immediacy, durability and capability together. PEARL adds adversarial prefixes, continuations from intermediate states of cached misaligned trajectories, to GRPO's on-policy rollouts so the policy practices recovering from partly wrong states. Across Qwen3-4B and GPT-OSS-20B and two organism types, PEARL suppresses strongly while preserving task accuracy, with the clearest gains on the broad-trigger organism and GPT-OSS-20B; on Qwen3-4B with the narrow-trigger organism, SFT is more durable under heavy targeted retraining.

Enzyme engineering has relied on directed evolution, which screens thousands of variants even though fewer than 5% of single mutations are beneficial. REXzyme recasts enzyme design as translation: an encoder reads a reaction written in SMILES and a decoder generates an amino acid sequence token by token, after training on about 21 million reaction-enzyme pairs. Steered toward the reverse reaction of formate dehydrogenase, the reduction of CO2 to formate, all 20 designs generated after family-specific fine-tuning and reinforcement-learning alignment expressed as soluble proteins and showed activity, and design C20 reached a specific activity of 80.8 mU/mg, 2.5 times that of wild-type CtFDH.

Enzyme engineering has relied on directed evolution, which screens thousands of variants even though fewer than 5% of single mutations are beneficial. REXzyme recasts enzyme design as translation: an encoder reads a reaction written in SMILES and a decoder generates an amino acid sequence token by token, after training on about 21 million reaction-enzyme pairs. Steered toward the reverse reaction of formate dehydrogenase, the reduction of CO2 to formate, all 20 designs generated after family-specific fine-tuning and reinforcement-learning alignment expressed as soluble proteins and showed activity, and design C20 reached a specific activity of 80.8 mU/mg, 2.5 times that of wild-type CtFDH.

Humanoids asked to sit down or walk to a named object need paired supervision mapping pixels and free-form language to whole-body joint commands, and that data has so far come from teleoperation or wearable capture, where an hour of data costs at least an hour of human time. HALO moves data generation into simulation, synthesizing egocentric vision-language-motion episodes across a thousand procedurally generated homes, then uses a layered stack that splits a free-form request into subtasks, generates five-second whole-body motion plans, and executes them with a perceptive tracker. Without seeing a real home in training, it transfers zero-shot to six unscanned, unmapped real homes, succeeding in 91% of 360 trials.

Humanoids asked to sit down or walk to a named object need paired supervision mapping pixels and free-form language to whole-body joint commands, and that data has so far come from teleoperation or wearable capture, where an hour of data costs at least an hour of human time. HALO moves data generation into simulation, synthesizing egocentric vision-language-motion episodes across a thousand procedurally generated homes, then uses a layered stack that splits a free-form request into subtasks, generates five-second whole-body motion plans, and executes them with a perceptive tracker. Without seeing a real home in training, it transfers zero-shot to six unscanned, unmapped real homes, succeeding in 91% of 360 trials.

Few-step diffusion distillation has long relied on estimating the student score, yet at intermediate steps that score is hard to track and the estimated distribution can differ from the one used to update the student. VA-TDM instead supervises velocity fields built from the student's own sampling process, uses a learned critic to compare student and teacher fields and turns their disagreement into generator feedback, unifying score distillation, Fisher matching and moment matching. On four-step SD3.5-Medium it raises GenEval from TDM's 0.61 to 0.67; on 33B MiniMax-H3 at the same 4-NFE budget it wins 61.33% of human comparisons against Fast-H3 and reaches a 45.53% preference rate against the 100-NFE teacher, versus 24.67% for Fast-H3.

Few-step diffusion distillation has long relied on estimating the student score, yet at intermediate steps that score is hard to track and the estimated distribution can differ from the one used to update the student. VA-TDM instead supervises velocity fields built from the student's own sampling process, uses a learned critic to compare student and teacher fields and turns their disagreement into generator feedback, unifying score distillation, Fisher matching and moment matching. On four-step SD3.5-Medium it raises GenEval from TDM's 0.61 to 0.67; on 33B MiniMax-H3 at the same 4-NFE budget it wins 61.33% of human comparisons against Fast-H3 and reaches a 45.53% preference rate against the 100-NFE teacher, versus 24.67% for Fast-H3.

Reaction figures in the chemical literature often pack several experiments into one shared template, substituent table and product grid, and recovering concrete reaction records from them has meant manual transcription. RxnMiner combines an expert-informed extraction procedure with a chemistry toolbox in one multimodal agent: it reads structures, binds evidence to each experimental entry, and instantiates undrawn reactants with a deterministic R-group engine. On ChemRxn-Bench, 603 reaction figures with 8,547 annotated records, it reaches 71.0% stereochemistry-aware reaction macro-F1 with a Qwen3.8-Flash-Next backbone, against 21.0% for direct prediction and 25.0% for the strongest existing extraction system.

Reaction figures in the chemical literature often pack several experiments into one shared template, substituent table and product grid, and recovering concrete reaction records from them has meant manual transcription. RxnMiner combines an expert-informed extraction procedure with a chemistry toolbox in one multimodal agent: it reads structures, binds evidence to each experimental entry, and instantiates undrawn reactants with a deterministic R-group engine. On ChemRxn-Bench, 603 reaction figures with 8,547 annotated records, it reaches 71.0% stereochemistry-aware reaction macro-F1 with a Qwen3.8-Flash-Next backbone, against 21.0% for direct prediction and 25.0% for the strongest existing extraction system.

DataFarm, from Peking University and NVIDIA, hands the teleoperator's job to a coding agent: given one calibration clip and a data request, the agent rebuilds the scene in simulation, writes an expert controller and a success judge, and a video model conditioned on real images renders multi-view video with action labels. On five real bimanual tasks, policies trained only on farm data reached 93 successes in 100 trials in the original scene, against 84 for an equal number of teleoperated episodes, and kept 87 under a tablecloth and coloured light absent from both training sets, where teleoperation-trained policies fell to 46.

DataFarm, from Peking University and NVIDIA, hands the teleoperator's job to a coding agent: given one calibration clip and a data request, the agent rebuilds the scene in simulation, writes an expert controller and a success judge, and a video model conditioned on real images renders multi-view video with action labels. On five real bimanual tasks, policies trained only on farm data reached 93 successes in 100 trials in the original scene, against 84 for an equal number of teleoperated episodes, and kept 87 under a tablecloth and coloured light absent from both training sets, where teleoperation-trained policies fell to 46.

AMD introduced Fused-Softmax Attention, a training-free algorithm for the CDNA 4 architecture that folds the dominant softmax operations into attention's matrix multiplies so low-precision matrix throughput is actually used. On an MI355X GPU the MXFP8, MXFP6 and MXFP4 kernels reach about 3.0, 4.0 and 4.6 PFLOP/s, while the added error stays below the error floor set by quantizing the GEMM operands. On Wan2.2-T2V-A14B a hybrid MXFP6/4 schedule delivers up to end-to-end speedup while preserving visual quality, and the kernels ship in AMD's open-source AITER library.

AMD introduced Fused-Softmax Attention, a training-free algorithm for the CDNA 4 architecture that folds the dominant softmax operations into attention's matrix multiplies so low-precision matrix throughput is actually used. On an MI355X GPU the MXFP8, MXFP6 and MXFP4 kernels reach about 3.0, 4.0 and 4.6 PFLOP/s, while the added error stays below the error floor set by quantizing the GEMM operands. On Wan2.2-T2V-A14B a hybrid MXFP6/4 schedule delivers up to end-to-end speedup while preserving visual quality, and the kernels ship in AMD's open-source AITER library.

MaskNGly treats protein language model embeddings as a readable source of coevolutionary signal, using a mask-and-probe readout to decide directly whether an N-X-S/T sequon will be N-glycosylated, without training on annotated data. On the N-GlyDE test set it reaches an AUC of 0.940 and an MCC of 0.808, comparable to published supervised neural-network predictors; in independent glycoproteomic data, 106 of 111 previously unreported sites scored above the 0.5 threshold.

MaskNGly treats protein language model embeddings as a readable source of coevolutionary signal, using a mask-and-probe readout to decide directly whether an N-X-S/T sequon will be N-glycosylated, without training on annotated data. On the N-GlyDE test set it reaches an AUC of 0.940 and an MCC of 0.808, comparable to published supervised neural-network predictors; in independent glycoproteomic data, 106 of 111 previously unreported sites scored above the 0.5 threshold.

Continuous diffusion language models generate a whole sentence in parallel without drawing a word at intermediate steps, and the cost is a collapsed lexical tail: ELF nearly matches human text on common words, yet its rarest words appear at only a fraction of the human rate, a gap that persists from 105M to 652M parameters and under every sampler setting tested. GRAFT reads the decoding head mid-generation, commits the rare words the average is about to discard, and regenerates from noise, raising the rare-word share from 0.42–0.77 to 0.82–1.25 of the human rate across three ELF scales and the independent LangFlow implementation, at the cost of a small rise in non-word rate on some models.

Continuous diffusion language models generate a whole sentence in parallel without drawing a word at intermediate steps, and the cost is a collapsed lexical tail: ELF nearly matches human text on common words, yet its rarest words appear at only a fraction of the human rate, a gap that persists from 105M to 652M parameters and under every sampler setting tested. GRAFT reads the decoding head mid-generation, commits the rare words the average is about to discard, and regenerates from noise, raising the rare-word share from 0.42–0.77 to 0.82–1.25 of the human rate across three ELF scales and the independent LangFlow implementation, at the cost of a small rise in non-word rate on some models.

Google released Gemini 4 Carbon for developers, enterprises, scientists and consumers, pitched at long-horizon agentic workflows and cybersecurity. It ranks first on three independent benchmarks, CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. API pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.

Google released Gemini 4 Carbon for developers, enterprises, scientists and consumers, pitched at long-horizon agentic workflows and cybersecurity. It ranks first on three independent benchmarks, CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. API pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.

KC-maps learns the finite-time transition kernel of a stochastic differential equation directly as a neural network map, trained only on uncorrelated marginal samples plus the known drift and diffusion coefficients rather than on paired simulated trajectories. Boundary conditions are built into the network architecture, and a training objective derived from the Kolmogorov–Chapman equation keeps transitions composed over different time scales consistent in distribution. On alanine dipeptide, tetrapeptide and hexapeptide it replaces up to 70 fs of sequential integration with a single learned transition, reaches lower TICA-2 distances at matched GPU-hour budgets, and on AL6 covers conformational modes absent from the training data.

KC-maps learns the finite-time transition kernel of a stochastic differential equation directly as a neural network map, trained only on uncorrelated marginal samples plus the known drift and diffusion coefficients rather than on paired simulated trajectories. Boundary conditions are built into the network architecture, and a training objective derived from the Kolmogorov–Chapman equation keeps transitions composed over different time scales consistent in distribution. On alanine dipeptide, tetrapeptide and hexapeptide it replaces up to 70 fs of sequential integration with a single learned transition, reaches lower TICA-2 distances at matched GPU-hour budgets, and on AL6 covers conformational modes absent from the training data.

Climate simulations are costly and biased, while correcting them too strongly toward observations risks erasing structural differences among models. A conditional diffusion model trained on seventeen CMIP6 models and ERA5 reanalysis steers each emulated model distribution toward observations at sampling time with an adjustable strength, chosen in perfect-model experiments with withheld simulated futures and decreasing with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature and brings CMIP6's roughly threefold historical overestimate of extreme-heat coverage close to observations.

Climate simulations are costly and biased, while correcting them too strongly toward observations risks erasing structural differences among models. A conditional diffusion model trained on seventeen CMIP6 models and ERA5 reanalysis steers each emulated model distribution toward observations at sampling time with an adjustable strength, chosen in perfect-model experiments with withheld simulated futures and decreasing with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature and brings CMIP6's roughly threefold historical overestimate of extreme-heat coverage close to observations.

Long-context inference makes the KV cache grow linearly with sequence length and often exceed GPU memory. VerKV keeps the full cache in CPU memory and only a Poisson sample of tokens, drawn from importance scores, on the GPU; at each decode step a certificate estimates the attention error and extra entries are fetched from CPU only when the bound exceeds tolerance. Across Qwen3-4B, Qwen3.5-9B and Ministral-3-8B on RULER and LOFT at 32K and 128K context, with 3.13% of the cache resident on GPU, the three retain 94.6%, 84.6% and 79.6% of dense quality, where Fast KVzip and KVzip at the same 3.13% retain only 1.9% to 11.9%.

Long-context inference makes the KV cache grow linearly with sequence length and often exceed GPU memory. VerKV keeps the full cache in CPU memory and only a Poisson sample of tokens, drawn from importance scores, on the GPU; at each decode step a certificate estimates the attention error and extra entries are fetched from CPU only when the bound exceeds tolerance. Across Qwen3-4B, Qwen3.5-9B and Ministral-3-8B on RULER and LOFT at 32K and 128K context, with 3.13% of the cache resident on GPU, the three retain 94.6%, 84.6% and 79.6% of dense quality, where Fast KVzip and KVzip at the same 3.13% retain only 1.9% to 11.9%.

Scientists at US Department of Energy national laboratories set the conditions an unpublished follow-up direction must meet, and nine frontier AI agents each wrote 15 follow-up papers, 135 in total. None passed all three checks of integrity, completeness and sufficiency, and 114 failed the integrity check, most often by claiming work that was never done or reporting numbers that contradict the recorded evidence. Agents readily produce papers that look like follow-ups, but none meets the standard the domain scientist set.

Scientists at US Department of Energy national laboratories set the conditions an unpublished follow-up direction must meet, and nine frontier AI agents each wrote 15 follow-up papers, 135 in total. None passed all three checks of integrity, completeness and sufficiency, and 114 failed the integrity check, most often by claiming work that was never done or reporting numbers that contradict the recorded evidence. Agents readily produce papers that look like follow-ups, but none meets the standard the domain scientist set.

In simulated worlds with a planted, known signal, Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra and Qwen3-32B were asked whether a forecasting model beats a static benchmark. After Gaussian shocks removed a resampling artifact, the three commercial models claimed predictability in 34%, 16% and 44% of 110 no-signal worlds, all above the nominal 5%; once false-positive rates are matched, an in-sample test with a multiple-testing adjustment detects weak signals about as well as the best agent.

In simulated worlds with a planted, known signal, Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra and Qwen3-32B were asked whether a forecasting model beats a static benchmark. After Gaussian shocks removed a resampling artifact, the three commercial models claimed predictability in 34%, 16% and 44% of 110 no-signal worlds, all above the nominal 5%; once false-positive rates are matched, an in-sample test with a multiple-testing adjustment detects weak signals about as well as the best agent.

Cellular responses to perturbation depend on several experimental variables at once, including perturbation identity, cell type, donor and dose, so each measured population occupies one coordinate of a multidimensional condition space. Earlier methods conditioned on explicit representations of those variable values and struggled when a value was absent from training. RELAY instead infers a target from measured control–perturbed pairs that share some variables with it, receiving only which variables are shared and never their values; across chemical, genetic and cytokine perturbation data it beats STACK, scGen and TabPFN on unseen variable values and supervised methods on unseen combinations, reaching 0.495 PDCorr on PBMC-3D against 0.225 for the strongest supervised baseline, STATE.

Cellular responses to perturbation depend on several experimental variables at once, including perturbation identity, cell type, donor and dose, so each measured population occupies one coordinate of a multidimensional condition space. Earlier methods conditioned on explicit representations of those variable values and struggled when a value was absent from training. RELAY instead infers a target from measured control–perturbed pairs that share some variables with it, receiving only which variables are shared and never their values; across chemical, genetic and cytokine perturbation data it beats STACK, scGen and TabPFN on unseen variable values and supervised methods on unseen combinations, reaching 0.495 PDCorr on PBMC-3D against 0.225 for the strongest supervised baseline, STATE.

Industrial robots are usually evaluated in episodes that begin from a reset, while factory work is a chain of jobs whose outcomes shape the next one. HolidayFactory 1.0 combines simulation, world construction and evaluation: Holiday-Sim adds constitutive laws, a semi-implicit primal-dual solver and stateful actuators to Newton's sparse-SDF hydroelastic contact, an agentic pipeline generates and verifies assets and scenes, and 24/7-Bench scores seven industrial tasks without resetting the scene, counting completed, failed and unfinished jobs, throughput, latency and motor energy. Evaluation-generation experiments show that passing executable checks does not establish agreement with task intent.

Industrial robots are usually evaluated in episodes that begin from a reset, while factory work is a chain of jobs whose outcomes shape the next one. HolidayFactory 1.0 combines simulation, world construction and evaluation: Holiday-Sim adds constitutive laws, a semi-implicit primal-dual solver and stateful actuators to Newton's sparse-SDF hydroelastic contact, an agentic pipeline generates and verifies assets and scenes, and 24/7-Bench scores seven industrial tasks without resetting the scene, counting completed, failed and unfinished jobs, throughput, latency and motor energy. Evaluation-generation experiments show that passing executable checks does not establish agreement with task intent.

Video diffusion models lose scene geometry when the camera moves or objects move, because cross-frame attention often spreads its response over several regions instead of locking onto the same scene element. GAD extracts correspondence targets from the attention of a frozen geometry foundation model, VGGT-, on clean videos, and routes supervision by noise timestep to the layer where scene structure is forming, training video attention with pairwise and trackwise objectives. On 15-second long-range generation with RealEstate10K and DL3DV-10K, adding GAD to Wan2.2 5B raises cross-frame matching hit rate by at least 30% relative on both datasets, with no extra inference modules or computation.

Video diffusion models lose scene geometry when the camera moves or objects move, because cross-frame attention often spreads its response over several regions instead of locking onto the same scene element. GAD extracts correspondence targets from the attention of a frozen geometry foundation model, VGGT-, on clean videos, and routes supervision by noise timestep to the layer where scene structure is forming, training video attention with pairwise and trackwise objectives. On 15-second long-range generation with RealEstate10K and DL3DV-10K, adding GAD to Wan2.2 5B raises cross-frame matching hit rate by at least 30% relative on both datasets, with no extra inference modules or computation.

Language-guided segmentation usually compresses the referent's location into a single [SEG] hidden-state vector handed to the decoder, a vector with no spatial index while the decoder never sees the expression. AttnSeg instead passes the [SEG] token's attention over image patches as the decoder's dense prompt and supervises that map directly with an attention-guidance objective in two training stages. On Sa2VA, the 4B model raises ReasonSeg gIoU from 57.70 to 64.11 and GS-Eval from 61.91 to 72.11, with gains concentrated on object parts, multiple objects and single objects, and negligible change on stuff regions.

Language-guided segmentation usually compresses the referent's location into a single [SEG] hidden-state vector handed to the decoder, a vector with no spatial index while the decoder never sees the expression. AttnSeg instead passes the [SEG] token's attention over image patches as the decoder's dense prompt and supervises that map directly with an attention-guidance objective in two training stages. On Sa2VA, the 4B model raises ReasonSeg gIoU from 57.70 to 64.11 and GS-Eval from 61.91 to 72.11, with gains concentrated on object parts, multiple objects and single objects, and negligible change on stuff regions.

In hierarchical Bayesian inversion, putting the forward model inside the likelihood means every posterior evaluation costs one forward solve per cell. A triangular transport-map likelihood trained once per design now serves every cell, and is used to infer contact angle and capillary length from photographs of a hole in a thin liquid film. At a synthetic validation design the map's posteriors agree with direct sampling in the wetting properties and pass simulation-based calibration; a single posterior needs several hundred times fewer forward-model evaluations than a gradient-free ensemble sampler, and further posteriors at the same design add no forward solve.

In hierarchical Bayesian inversion, putting the forward model inside the likelihood means every posterior evaluation costs one forward solve per cell. A triangular transport-map likelihood trained once per design now serves every cell, and is used to infer contact angle and capillary length from photographs of a hole in a thin liquid film. At a synthetic validation design the map's posteriors agree with direct sampling in the wetting properties and pass simulation-based calibration; a single posterior needs several hundred times fewer forward-model evaluations than a gradient-free ensemble sampler, and further posteriors at the same design add no forward solve.

Placing the machine-learning weather model GraphCast inside a real data-assimilation cycle, with the mean still supplied by the physics-based MPAS model and GraphCast evolving only ensemble perturbations, gives ensemble spread and covariance structure comparable to an all-physics baseline over one month; the EnVar background RMSE against GFS analyses is consistently lower below model level 40, while degradation appears near the model top from interpolation and extrapolation.

Placing the machine-learning weather model GraphCast inside a real data-assimilation cycle, with the mean still supplied by the physics-based MPAS model and GraphCast evolving only ensemble perturbations, gives ensemble spread and covariance structure comparable to an all-physics baseline over one month; the EnVar background RMSE against GFS analyses is consistently lower below model level 40, while degradation appears near the model top from interpolation and extrapolation.

Scientific foundation models usually ship with aggregate benchmark scores, while the performance limits that matter in use surface only after deployment through scattered reports. AutoSPECTRA constrains a general-purpose coding agent into a reproducible protocol: define a measurable generalization axis, run controlled comparisons against fixed reference methods, and confirm on data not used to select the axis. Across four biological foundation models, ESMFold2, STATE, CONCH and AlphaMissense, the audits tied performance shifts to protein sequence composition and relatedness to pretraining data, representation of perturbations in training, effective spatial resolution, and residue-level genomic, evolutionary and structural context, and rejected several initially plausible hypotheses.

Scientific foundation models usually ship with aggregate benchmark scores, while the performance limits that matter in use surface only after deployment through scattered reports. AutoSPECTRA constrains a general-purpose coding agent into a reproducible protocol: define a measurable generalization axis, run controlled comparisons against fixed reference methods, and confirm on data not used to select the axis. Across four biological foundation models, ESMFold2, STATE, CONCH and AlphaMissense, the audits tied performance shifts to protein sequence composition and relatedness to pretraining data, representation of perturbations in training, effective spatial resolution, and residue-level genomic, evolutionary and structural context, and rejected several initially plausible hypotheses.

Long-horizon agents accumulate interaction histories that outgrow the active context, so they must learn what to write down and when to read it back. Earlier on-policy distillation of memory control relied on dedicated memory tools or compaction protocols that neither the teacher nor the student had seen, which weakened guidance on the teacher side and left the student sampling too few meaningful memory actions. CAMP-II instead routes memory through ordinary Bash file operations that both models already have a prior for, training one policy jointly across five long-horizon environments to a 64.6% average task success rate against 54.2% for the task-reward reference CAMP-I, with transfer to Terminal-Bench 2.1 and SWE-Bench Pro V2.

Long-horizon agents accumulate interaction histories that outgrow the active context, so they must learn what to write down and when to read it back. Earlier on-policy distillation of memory control relied on dedicated memory tools or compaction protocols that neither the teacher nor the student had seen, which weakened guidance on the teacher side and left the student sampling too few meaningful memory actions. CAMP-II instead routes memory through ordinary Bash file operations that both models already have a prior for, training one policy jointly across five long-horizon environments to a 64.6% average task success rate against 54.2% for the task-reward reference CAMP-I, with transfer to Terminal-Bench 2.1 and SWE-Bench Pro V2.

Transformer models for single-cell perturbation prediction need a position for each gene, yet genes have no natural order, and static co-expression masks cannot reflect the regulatory rewiring a perturbation induces. SPEAR replaces that position with continuous coordinates from the co-expression graph Laplacian and, at every layer, re-derives a scalar position from the perturbation-conditioned hidden state, summing both into one rotation angle. On genetic, pharmacological and cytokine perturbation benchmarks it beats baselines including scDFM on differential-expression direction and magnitude, with the largest gap on held-out combinations.

Transformer models for single-cell perturbation prediction need a position for each gene, yet genes have no natural order, and static co-expression masks cannot reflect the regulatory rewiring a perturbation induces. SPEAR replaces that position with continuous coordinates from the co-expression graph Laplacian and, at every layer, re-derives a scalar position from the perturbation-conditioned hidden state, summing both into one rotation angle. On genetic, pharmacological and cytokine perturbation benchmarks it beats baselines including scDFM on differential-expression direction and magnitude, with the largest gap on held-out combinations.

Behavior control in large language models often costs general capability; CANS instead assigns each feed-forward network neuron a scaling coefficient that accounts both for its contribution to the target behavior and for its disruption of general outputs. Target behavior is proxied by the log-probability of target tokens in model responses, capability preservation by output-distribution KL divergence on general-purpose text, and local approximations of the two yield a closed-form solution with no iterative training and no added inference-time computation. On Qwen3-4B and Llama models, CANS matches or beats SFT, CAA and neuron-level baselines on safety, toxicity and hallucination control while better preserving MMLU, BBH and IFEval performance.

Behavior control in large language models often costs general capability; CANS instead assigns each feed-forward network neuron a scaling coefficient that accounts both for its contribution to the target behavior and for its disruption of general outputs. Target behavior is proxied by the log-probability of target tokens in model responses, capability preservation by output-distribution KL divergence on general-purpose text, and local approximations of the two yield a closed-form solution with no iterative training and no added inference-time computation. On Qwen3-4B and Llama models, CANS matches or beats SFT, CAA and neuron-level baselines on safety, toxicity and hallucination control while better preserving MMLU, BBH and IFEval performance.

On mobile devices the binding constraint for large language models is weight memory, not computation. DictLLM splits weight matrices within grouped layers into one shared dictionary plus layer-specific sparse coefficients, all computed from a pretrained model without retraining. On LLaMA-7B it compresses the model 5.1× with only a 2.9-point rise in WikiText-2 perplexity, a compression ratio about 3.6× that of the state-of-the-art low-rank method.

On mobile devices the binding constraint for large language models is weight memory, not computation. DictLLM splits weight matrices within grouped layers into one shared dictionary plus layer-specific sparse coefficients, all computed from a pretrained model without retraining. On LLaMA-7B it compresses the model 5.1× with only a 2.9-point rise in WikiText-2 perplexity, a compression ratio about 3.6× that of the state-of-the-art low-rank method.

For the rational function field C(x) under the shift σ(x)=x+1, it was known that a Picard–Vessiot group G must have G/G° cyclic, but which groups are actually realizable remained open. It is now shown that every closed subgroup G⊆GLn(C) with cyclic G/G° occurs, with coefficient matrix of the form A(x)=h+O(x⁻²) at infinity for semisimple h, answering the conjecture of van der Put and Singer.

For the rational function field C(x) under the shift σ(x)=x+1, it was known that a Picard–Vessiot group G must have G/G° cyclic, but which groups are actually realizable remained open. It is now shown that every closed subgroup G⊆GLn(C) with cyclic G/G° occurs, with coefficient matrix of the form A(x)=h+O(x⁻²) at infinity for semisimple h, answering the conjecture of van der Put and Singer.

Rescene-ESM freezes an AI weather model trained only for medium-range forecasting as its atmosphere and builds observation-constrained physical ocean, sea-ice, land, river, carbon-cycle and energy-balance components around it. A coupling ladder shows that stochastic SST forcing and the land's moisture flux raise El Niño–Southern Oscillation amplitude and teleconnections step by step toward observed levels while the mean climate barely moves. The full model simulates about 1,000 years per day on one GPU, holds a thousand-year control without drift, and carries emissions scenarios through rivers to population heatwave exposure.

Rescene-ESM freezes an AI weather model trained only for medium-range forecasting as its atmosphere and builds observation-constrained physical ocean, sea-ice, land, river, carbon-cycle and energy-balance components around it. A coupling ladder shows that stochastic SST forcing and the land's moisture flux raise El Niño–Southern Oscillation amplitude and teleconnections step by step toward observed levels while the mean climate barely moves. The full model simulates about 1,000 years per day on one GPU, holds a thousand-year control without drift, and carries emissions scenarios through rivers to population heatwave exposure.

In sheet-forming simulation, Transformer-based neural operators move information efficiently across the mesh but tend to smooth the localised thinning near deep-drawn corners and reinforcing ribs, the regions that often decide whether a part can be formed. BiLoGo couples local node–element message passing on a bipartite graph with a dual-stream Transformer neural operator, compressing the blank state and the tool contact into separate tokens and keeping local features through gated fusion. On a ribbed hot-stamping case, BiLoGo-U raises the thinning smoothing index near the ribs from 0.41–0.56 for Transformer-based operators to 0.94, and cuts mean positional error by 8.3% relative to the most accurate hierarchical graph network, CAtt-BiUGNN.

In sheet-forming simulation, Transformer-based neural operators move information efficiently across the mesh but tend to smooth the localised thinning near deep-drawn corners and reinforcing ribs, the regions that often decide whether a part can be formed. BiLoGo couples local node–element message passing on a bipartite graph with a dual-stream Transformer neural operator, compressing the blank state and the tool contact into separate tokens and keeping local features through gated fusion. On a ribbed hot-stamping case, BiLoGo-U raises the thinning smoothing index near the ribs from 0.41–0.56 for Transformer-based operators to 0.94, and cuts mean positional error by 8.3% relative to the most accurate hierarchical graph network, CAtt-BiUGNN.

The bottleneck in analog layout automation is not any single tool but the missing, revisable coordination between sizing and physical structure. CASCoDe organizes sizing candidates, floorplan form, intra-row refinement and routing attempts into a hierarchical discrete action space, and an LLM agent runs feedback-guided tree search over it using post-extraction electrical results. On 20 SKY130 amplifiers it is the only flow that delivers a DRC- and LVS-clean routed layout for every circuit, cutting area by 59–85% (mean 71%) against APPlace on identical PCells.

The bottleneck in analog layout automation is not any single tool but the missing, revisable coordination between sizing and physical structure. CASCoDe organizes sizing candidates, floorplan form, intra-row refinement and routing attempts into a hierarchical discrete action space, and an LLM agent runs feedback-guided tree search over it using post-extraction electrical results. On 20 SKY130 amplifiers it is the only flow that delivers a DRC- and LVS-clean routed layout for every circuit, cutting area by 59–85% (mean 71%) against APPlace on identical PCells.

When LLM agents operate Web services from the command line, they assume every tool call returns within a bounded time with the requested action finished; but many commands ask a question first, and a harness that closes input can let a program return status 0 without doing the work, while an open input makes it wait until a time limit. Receipt keeps input open, reads from system calls what each process waits for, and returns that state to an unmodified agent; across 5 LLMs it lowers false successes from 34 under closed input to 9 and uses 24.4% fewer tokens than a plain terminal.

When LLM agents operate Web services from the command line, they assume every tool call returns within a bounded time with the requested action finished; but many commands ask a question first, and a harness that closes input can let a program return status 0 without doing the work, while an open input makes it wait until a time limit. Receipt keeps input open, reads from system calls what each process waits for, and returns that state to an unmodified agent; across 5 LLMs it lowers false successes from 34 under closed input to 9 and uses 24.4% fewer tokens than a plain terminal.

LLM pruning usually targets parameter counts, but parameter count does not predict latency, throughput or memory, which depend on architecture, workload, hardware and serving engine. AutoCompress puts latency, throughput, parameter and memory constraints directly into differentiable architecture search, searching global dimensions and per-layer structure together and guided by cost models calibrated by measurement inside the target serving engine. Under the same deployment budgets and matched recovery training as Minitron++, it raises average accuracy across eight benchmarks by 6.57, 2.35 and 0.57 percentage points on RTX 5070, RTX 5080 and H100, using roughly one-half to one-third as many architecture-selection tokens.

LLM pruning usually targets parameter counts, but parameter count does not predict latency, throughput or memory, which depend on architecture, workload, hardware and serving engine. AutoCompress puts latency, throughput, parameter and memory constraints directly into differentiable architecture search, searching global dimensions and per-layer structure together and guided by cost models calibrated by measurement inside the target serving engine. Under the same deployment budgets and matched recovery training as Minitron++, it raises average accuracy across eight benchmarks by 6.57, 2.35 and 0.57 percentage points on RTX 5070, RTX 5080 and H100, using roughly one-half to one-third as many architecture-selection tokens.

Risk in multi-step LLM agents propagates along the workflow, and step-by-step independent verification does not see that propagation. TAS models workflow safety states as a Markov decision process, measures rare but high-cost tail risk with a spectral risk measure, and constrains each step's decision with one budget shared across steps. On Agent Security Bench, TAS reaches average attack success rates of 45.15% under direct prompt injection and 29.73% under indirect prompt injection, below the eleven defense baselines it compares against.

Risk in multi-step LLM agents propagates along the workflow, and step-by-step independent verification does not see that propagation. TAS models workflow safety states as a Markov decision process, measures rare but high-cost tail risk with a spectral risk measure, and constrains each step's decision with one budget shared across steps. On Agent Security Bench, TAS reaches average attack success rates of 45.15% under direct prompt injection and 29.73% under indirect prompt injection, below the eleven defense baselines it compares against.

Pretrained flow-matching models learn only the data distribution and often violate conservation laws and boundary conditions at inference; earlier zero-shot fixes used algebraic projections that hit memory bottlenecks and disturb the probability path on dense, full-domain PDEs. Energy-Guided Transport writes the physical residual as a smooth potential energy and applies a continuous dissipative force to intermediate states during integration, obtaining the gradient implicitly through reverse-mode automatic differentiation. Across the heat equation, reaction-diffusion, Burgers' equation and 2D Navier-Stokes, it holds a 100% success rate under full-PDE constraints, generating a 2D Navier-Stokes sample in about 0.94 s where explicit projection fails.

Pretrained flow-matching models learn only the data distribution and often violate conservation laws and boundary conditions at inference; earlier zero-shot fixes used algebraic projections that hit memory bottlenecks and disturb the probability path on dense, full-domain PDEs. Energy-Guided Transport writes the physical residual as a smooth potential energy and applies a continuous dissipative force to intermediate states during integration, obtaining the gradient implicitly through reverse-mode automatic differentiation. Across the heat equation, reaction-diffusion, Burgers' equation and 2D Navier-Stokes, it holds a 100% success rate under full-PDE constraints, generating a 2D Navier-Stokes sample in about 0.94 s where explicit projection fails.

Clinical EEG reading has long depended on experts paging through recordings, while recent EEG foundation models first pretrain on large unlabeled EEG corpora. This work renders EEG segments as the bipolar montage waveforms or spectrograms clinicians already inspect, pairs each image with a fixed text prompt, and hands it to a general vision-language model, Qwen3-VL, with no EEG pretraining at all. On the 4-second interictal epileptiform discharge task vEpiSet, Qwen3-VL-4B adapted with LoRA reaches 0.835 balanced accuracy against 0.763 for fully fine-tuned BIOT; on the TUEV EEG-Bench split, Qwen3-VL-8B reaches 0.750 against 0.739 for the second-best single-task model.

Clinical EEG reading has long depended on experts paging through recordings, while recent EEG foundation models first pretrain on large unlabeled EEG corpora. This work renders EEG segments as the bipolar montage waveforms or spectrograms clinicians already inspect, pairs each image with a fixed text prompt, and hands it to a general vision-language model, Qwen3-VL, with no EEG pretraining at all. On the 4-second interictal epileptiform discharge task vEpiSet, Qwen3-VL-4B adapted with LoRA reaches 0.835 balanced accuracy against 0.763 for fully fine-tuned BIOT; on the TUEV EEG-Bench split, Qwen3-VL-8B reaches 0.750 against 0.739 for the second-best single-task model.

Vision-language-action models and world models for minimally invasive surgery need large-scale 3D supervision, yet public surgical datasets are largely ex-vivo or simulated and annotate either tissue or instruments, leaving tool-tissue contact regions unmarked. SurgTAP-3D converts uncalibrated monocular surgical videos into pseudo-scale camera poses, depth and 2D/3D point trajectories through contact-region prompting, active-tool segmentation, dynamic bundle adjustment and global 4D refinement, gathering 4,549 action clips, 519,642 frames and about 1.38 million trajectories across robotic and manual procedures. On the SurgT, SurgPose and STIR benchmarks, the pipeline reaches an average AJ3D of 56.57 for 3D tracking, above the compared methods in the table.

Vision-language-action models and world models for minimally invasive surgery need large-scale 3D supervision, yet public surgical datasets are largely ex-vivo or simulated and annotate either tissue or instruments, leaving tool-tissue contact regions unmarked. SurgTAP-3D converts uncalibrated monocular surgical videos into pseudo-scale camera poses, depth and 2D/3D point trajectories through contact-region prompting, active-tool segmentation, dynamic bundle adjustment and global 4D refinement, gathering 4,549 action clips, 519,642 frames and about 1.38 million trajectories across robotic and manual procedures. On the SurgT, SurgPose and STIR benchmarks, the pipeline reaches an average AJ3D of 56.57 for 3D tracking, above the compared methods in the table.

Generalizing randomized-trial findings to a broader target population can yield distorted estimates when the trial sample is misaligned with that population. TRACE decomposes the asymptotic variance of the efficient generalizability estimator into a baseline-mismatch channel and a censoring channel, then uses a tree search to localize underrepresented subgroups and attribute each to its mechanism. In a case study generalizing the JTPA Study to U.S. JTPA-eligible adults, restricting to reliably generalizable subgroups narrows the 95% confidence interval for the refined estimand by 60%.

Generalizing randomized-trial findings to a broader target population can yield distorted estimates when the trial sample is misaligned with that population. TRACE decomposes the asymptotic variance of the efficient generalizability estimator into a baseline-mismatch channel and a censoring channel, then uses a tree search to localize underrepresented subgroups and attribute each to its mechanism. In a case study generalizing the JTPA Study to U.S. JTPA-eligible adults, restricting to reliably generalizable subgroups narrows the 95% confidence interval for the refined estimand by 60%.

LieSPHGP combines port-Hamiltonian structure, Lie-group geometry, Gaussian processes and stochastic differential equations to learn controlled, dissipative rigid-body dynamics from noisy trajectories. Matérn-kernel Gaussian processes model mass, potential energy, dissipation, control and diffusion separately, while exponential-map updates keep rotations on the manifold. On a windy pendulum and a quadrotor benchmark, it improves prediction accuracy, long-horizon stability and manifold-constraint satisfaction over neural dynamics baselines.

LieSPHGP combines port-Hamiltonian structure, Lie-group geometry, Gaussian processes and stochastic differential equations to learn controlled, dissipative rigid-body dynamics from noisy trajectories. Matérn-kernel Gaussian processes model mass, potential energy, dissipation, control and diffusion separately, while exponential-map updates keep rotations on the manifold. On a windy pendulum and a quadrotor benchmark, it improves prediction accuracy, long-horizon stability and manifold-constraint satisfaction over neural dynamics baselines.

FinCosmos opens all three uncertainties of financial agent evaluation at once: a simulated investor reveals a hidden strategy only when asked, the agent must source its own Web data and backtest under a data cutoff, and programs verify the delivery instead of matching a reference answer. Across 150 sessions with four coding agents and a tool-free model, none is accepted, and all 35 sessions the simulated user signed off fail at least one program veto.

FinCosmos opens all three uncertainties of financial agent evaluation at once: a simulated investor reveals a hidden strategy only when asked, the agent must source its own Web data and backtest under a data cutoff, and programs verify the delivery instead of matching a reference answer. Across 150 sessions with four coding agents and a tool-free model, none is accepted, and all 35 sessions the simulated user signed off fail at least one program veto.

Working again with Diode Computers, Claude Fable 5.2 designed three devices—a speakerphone, a voice-assistant hub and a handheld virtual pet—from short prompts, handling board layout, routing, enclosure modeling and firmware itself. Earlier models could not reliably lay out a circuit board and produced boards that were 95% complete yet unusable; on Diode's internal PCB Gym benchmark, Fable 5.2 scored 69% against 25% for Fable 5. All three boards were manufactured directly from the design files without adjustments, but they remain prototypes that have not been through electrical validation, antenna tuning, thermal testing or certification.

Working again with Diode Computers, Claude Fable 5.2 designed three devices—a speakerphone, a voice-assistant hub and a handheld virtual pet—from short prompts, handling board layout, routing, enclosure modeling and firmware itself. Earlier models could not reliably lay out a circuit board and produced boards that were 95% complete yet unusable; on Diode's internal PCB Gym benchmark, Fable 5.2 scored 69% against 25% for Fable 5. All three boards were manufactured directly from the design files without adjustments, but they remain prototypes that have not been through electrical validation, antenna tuning, thermal testing or certification.

On 2D halide perovskite bandgap prediction, Motif moves the large language model from emitting numbers to generating chemical taxonomies offline, which then define environments for invariant feature selection. Across six leave-one-spacer-family-out folds, mean OOD MAE falls from the kNN+GBT baseline's 0.122 eV to 0.111 eV; under a frozen dev/test protocol the deployable configuration lowers electronic bandgap error from 0.123 to 0.117 eV and optical bandgap error from 0.181 to 0.173 eV.

On 2D halide perovskite bandgap prediction, Motif moves the large language model from emitting numbers to generating chemical taxonomies offline, which then define environments for invariant feature selection. Across six leave-one-spacer-family-out folds, mean OOD MAE falls from the kNN+GBT baseline's 0.122 eV to 0.111 eV; under a frozen dev/test protocol the deployable configuration lowers electronic bandgap error from 0.123 to 0.117 eV and optical bandgap error from 0.181 to 0.173 eV.

Coding agents resend the same long conversation, so whether the server still holds that history outside GPU memory decides how many tokens each turn must recompute. AMD's open-source ATOM inference engine, which places prefill and decode on separate GPUs, now connects to Mooncake Store to merge both halves of host memory into one KV pool that prefill saves to and restores from. On four AMD Instinct MI350X servers, the shared pool reached about twice the throughput of per-server pools at 96 and 112 concurrent sessions, with KV reuse held at 94–96%.

Coding agents resend the same long conversation, so whether the server still holds that history outside GPU memory decides how many tokens each turn must recompute. AMD's open-source ATOM inference engine, which places prefill and decode on separate GPUs, now connects to Mooncake Store to merge both halves of host memory into one KV pool that prefill saves to and restores from. On four AMD Instinct MI350X servers, the shared pool reached about twice the throughput of per-server pools at 96 and 112 concurrent sessions, with KV reuse held at 94–96%.

When a regulator evaluates a learned policy, the inputs and outputs are usually all that is visible, so it is hard to tell whether value comes from acting on the state or from average behavior. This work shows that, in a finite-horizon Markov decision process, the expected-cost difference between a policy and its open-loop marginal equals exactly the covariance between state and action, and it bounds adaptive value by the mutual information the policy uses. Auditing that quantity under bandit feedback needs a sample size inversely proportional to an average KL divergence, while a deterministic non-constant policy is not identifiable from its own logs.

When a regulator evaluates a learned policy, the inputs and outputs are usually all that is visible, so it is hard to tell whether value comes from acting on the state or from average behavior. This work shows that, in a finite-horizon Markov decision process, the expected-cost difference between a policy and its open-loop marginal equals exactly the covariance between state and action, and it bounds adaptive value by the mutual information the policy uses. Auditing that quantity under bandit feedback needs a sample size inversely proportional to an average KL divergence, while a deterministic non-constant policy is not identifiable from its own logs.

Google Cloud released Gemini for healthcare, an industry configuration of the general Gemini agent it introduced last month, pairing reusable skills, connectors to electronic health records and claims systems, and a single governance layer to carry multi-step clinical and administrative work across systems. It targets a setting where patient records, coverage policies and clinical evidence sit in disconnected records, claims systems and administrative tools, and its outputs are meant to carry traceable citations with human oversight. The offering is available in preview in the U.S. starting today, and Google says it is built to support HIPAA compliance.

Google Cloud released Gemini for healthcare, an industry configuration of the general Gemini agent it introduced last month, pairing reusable skills, connectors to electronic health records and claims systems, and a single governance layer to carry multi-step clinical and administrative work across systems. It targets a setting where patient records, coverage policies and clinical evidence sit in disconnected records, claims systems and administrative tools, and its outputs are meant to carry traceable citations with human oversight. The offering is available in preview in the U.S. starting today, and Google says it is built to support HIPAA compliance.

Periodic trained Periodic Neon on X-ray diffraction data from its own labs, reaching a 55.3% success rate on FrontierXRD, its hardest internal evaluation set, against an initial 2.7% for Kimi K2.6, the open-weight model it post-trained. The evaluation covers 134 samples from its labs that human experts take hours to resolve because of their complexity and ambiguity; Neon surpasses frontier models at a lower cost per analysis and is now deployed in its labs analyzing experiments in the search for better superconductors and magnets.

Periodic trained Periodic Neon on X-ray diffraction data from its own labs, reaching a 55.3% success rate on FrontierXRD, its hardest internal evaluation set, against an initial 2.7% for Kimi K2.6, the open-weight model it post-trained. The evaluation covers 134 samples from its labs that human experts take hours to resolve because of their complexity and ambiguity; Neon surpasses frontier models at a lower cost per analysis and is now deployed in its labs analyzing experiments in the search for better superconductors and magnets.

Wildfire decisions are made in hours, yet most learned spread models predict daily on targets that include already burned area. FireGraphNet recasts roughly 12-hour growth as node classification on directed H3 hexagonal graphs: candidates are unburned cells within five rings of the front, edges carry wind alignment, elevation difference, fuel continuity and barriers, and a graph neural network passes messages along those edges. On five sealed California test fires the five variants reach average precision 1.7–1.9 times the strongest distance baseline, and the GINE variant keeps a 1.62-fold advantage out of state.

Wildfire decisions are made in hours, yet most learned spread models predict daily on targets that include already burned area. FireGraphNet recasts roughly 12-hour growth as node classification on directed H3 hexagonal graphs: candidates are unburned cells within five rings of the front, edges carry wind alignment, elevation difference, fuel continuity and barriers, and a graph neural network passes messages along those edges. On five sealed California test fires the five variants reach average precision 1.7–1.9 times the strongest distance baseline, and the GINE variant keeps a 1.62-fold advantage out of state.

Spatial transcriptomics often measures RNA without matching protein, and RNA abundance does not fully reflect protein abundance. MorphoProt uses each spot's RNA expression as a query over five H&E fields of view, then aggregates physical and transcriptomic neighborhoods through gated pathways before jointly decoding the protein panel. On held-out sections of human tonsil, mouse spleen and human liver it reached mean protein-wise Spearman correlations of 0.652, 0.604 and 0.372, above RNA-only, graph-based, HIPI and SPRINT comparisons in all three tissues.

Spatial transcriptomics often measures RNA without matching protein, and RNA abundance does not fully reflect protein abundance. MorphoProt uses each spot's RNA expression as a query over five H&E fields of view, then aggregates physical and transcriptomic neighborhoods through gated pathways before jointly decoding the protein panel. On held-out sections of human tonsil, mouse spleen and human liver it reached mean protein-wise Spearman correlations of 0.652, 0.604 and 0.372, above RNA-only, graph-based, HIPI and SPRINT comparisons in all three tissues.

Dynamic PET must pick a kinetic model for each voxel's tracer curve, and the standard route fits every candidate by nonlinear least squares and compares AIC or BIC, which is slow and favors simple models under noise. A consistency model folds the continuous kinetic parameters and the discrete model index into one target vector conditioned on the measured time-activity curve and input function, so one forward sampling pass yields the joint posterior. On 7,000 simulated FDG total-body PET curves it reached 97.97% selection accuracy across seven compartment models versus 58.67% for NLS+AIC, and 91.4% versus 73.5% for NLS+BIC in simulated brain dopamine-release detection.

Dynamic PET must pick a kinetic model for each voxel's tracer curve, and the standard route fits every candidate by nonlinear least squares and compares AIC or BIC, which is slow and favors simple models under noise. A consistency model folds the continuous kinetic parameters and the discrete model index into one target vector conditioned on the measured time-activity curve and input function, so one forward sampling pass yields the joint posterior. On 7,000 simulated FDG total-body PET curves it reached 97.97% selection accuracy across seven compartment models versus 58.67% for NLS+AIC, and 91.4% versus 73.5% for NLS+BIC in simulated brain dopamine-release detection.

Hospitals use readmission-risk predictions to decide who leaves first, but outcomes are recorded only for discharged patients, so retraining a predictor on that data can lock in censoring bias. Coupling a queueing-based Markov decision process with a periodically retrained risk predictor, the work finds that a discharge-up-to policy, which sets a target census and discharges by priority only when occupancy exceeds it, spreads observed lengths of stay through natural arrival variability. In synthetic experiments and calibrated simulations built on patient-level data from a partner hospital, the policy reduced learning error by 24%–25% relative to thresholding on risk and shrank baseline-adjusted readmission cost disparities by 38%–63%.

Hospitals use readmission-risk predictions to decide who leaves first, but outcomes are recorded only for discharged patients, so retraining a predictor on that data can lock in censoring bias. Coupling a queueing-based Markov decision process with a periodically retrained risk predictor, the work finds that a discharge-up-to policy, which sets a target census and discharges by priority only when occupancy exceeds it, spreads observed lengths of stay through natural arrival variability. In synthetic experiments and calibrated simulations built on patient-level data from a partner hospital, the policy reduced learning error by 24%–25% relative to thresholding on risk and shrank baseline-adjusted readmission cost disparities by 38%–63%.

Meta's open-weights mixture-of-experts model Muse Spark Open has 816B total parameters and about 42B active per token; serving it on TPU v7 previously took 2.9× the time needed just to stream its weights from HBM for one verify step. SGLang-TorchTPU now writes all 62 layers of a decode or verify step as one Pallas program, issuing weight, KV and cross-device transfers ahead of the compute that needs them. Per output token, a single request drops from 3.78 ms to 1.95 ms, two concurrent requests gain 1.85×, and the gain narrows as concurrency grows.

Meta's open-weights mixture-of-experts model Muse Spark Open has 816B total parameters and about 42B active per token; serving it on TPU v7 previously took 2.9× the time needed just to stream its weights from HBM for one verify step. SGLang-TorchTPU now writes all 62 layers of a decode or verify step as one Pallas program, issuing weight, KV and cross-device transfers ahead of the compute that needs them. Per output token, a single request drops from 3.78 ms to 1.95 ms, two concurrent requests gain 1.85×, and the gain narrows as concurrency grows.

BacLLM injects semantic representations of UniProt annotations, encoded by a large language model, into an ESM-2 sequence encoder through cross-modal contrastive learning, so inference reads amino acid sequences alone. Across six resistance-prediction benchmarks, BacLLM (4B) beat Ankh, ESM-2, ProteinBERT and ProTrek in mean F1 under both logistic regression and an RBF-kernel support vector machine; on the SHV-1 β-lactamase P0AD64 it recovered six of eight experimentally supported resistance-attenuating sites among the top 65 ranked positions, while ESM-2 recovered none. The text branch is removed after training, so these gains come from sequence input, not annotation input.

BacLLM injects semantic representations of UniProt annotations, encoded by a large language model, into an ESM-2 sequence encoder through cross-modal contrastive learning, so inference reads amino acid sequences alone. Across six resistance-prediction benchmarks, BacLLM (4B) beat Ankh, ESM-2, ProteinBERT and ProTrek in mean F1 under both logistic regression and an RBF-kernel support vector machine; on the SHV-1 β-lactamase P0AD64 it recovered six of eight experimentally supported resistance-attenuating sites among the top 65 ranked positions, while ESM-2 recovered none. The text branch is removed after training, so these gains come from sequence input, not annotation input.

Full waveform inversion reconstructs subsurface velocity models from seismic records, but it is highly nonlinear and ill-posed, and an inaccurate initial model or noisy data can push it to a wrong solution. Bayesian full waveform inversion built on annealed Stein variational gradient descent previously constrained its particles with handcrafted Gaussian or smoothing priors that cannot represent complex geology. This work embeds a diffusion model trained on the GeoFWI dataset as an implicit geological projection operator inside the particle update, lowering mean absolute error on the Marmousi synthetic model from 94.80 with the classic method to 92.68, raising structural similarity from 0.69 to 0.72, and concentrating posterior uncertainty near deep structures and high-contrast interfaces.

Full waveform inversion reconstructs subsurface velocity models from seismic records, but it is highly nonlinear and ill-posed, and an inaccurate initial model or noisy data can push it to a wrong solution. Bayesian full waveform inversion built on annealed Stein variational gradient descent previously constrained its particles with handcrafted Gaussian or smoothing priors that cannot represent complex geology. This work embeds a diffusion model trained on the GeoFWI dataset as an implicit geological projection operator inside the particle update, lowering mean absolute error on the Marmousi synthetic model from 94.80 with the classic method to 92.68, raising structural similarity from 0.69 to 0.72, and concentrating posterior uncertainty near deep structures and high-contrast interfaces.

NVIDIA has donated Agent Substrate, an open-source runtime for running large numbers of sandboxed AI agents on Kubernetes, to the CNCF Sandbox. Agents spend most of their time waiting on a prompt, a tool result or a model response; keeping one Pod per session holds memory for idle sessions, while creating Pods on demand pays a multi-second cold start at every step and loses state. Agent Substrate multiplexes many sessions over a small pool of warm worker Pods, checkpointing idle sessions including memory and filesystem and restoring them onto whichever worker is free, and the project reports 10x higher density than standard containers.

NVIDIA has donated Agent Substrate, an open-source runtime for running large numbers of sandboxed AI agents on Kubernetes, to the CNCF Sandbox. Agents spend most of their time waiting on a prompt, a tool result or a model response; keeping one Pod per session holds memory for idle sessions, while creating Pods on demand pays a multi-second cold start at every step and loses state. Agent Substrate multiplexes many sessions over a small pool of warm worker Pods, checkpointing idle sessions including memory and filesystem and restoring them onto whichever worker is free, and the project reports 10x higher density than standard containers.

Microsoft researchers describe Lazarus, a backdoor that needs no retraining: rewriting fewer than 10 rarely firing dormant neurons in a large language model is enough to make an AI agent built on it execute attacker commands when a hidden trigger appears. The trigger can be as mundane as a date change, the model passes standard evaluations, and reliable detection requires diffing weights against the original, so provenance and weight integrity become central.

Microsoft researchers describe Lazarus, a backdoor that needs no retraining: rewriting fewer than 10 rarely firing dormant neurons in a large language model is enough to make an AI agent built on it execute attacker commands when a hidden trigger appears. The trigger can be as mundane as a date change, the model passes standard evaluations, and reliable detection requires diffing weights against the original, so provenance and weight integrity become central.

A 6U CubeSat trained and ran a thermal-infrared segmentation model onboard in low Earth orbit, then fed the model's output into the attitude determination and control system to point at the Moon in closed loop from a live thermal image stream. Training ran on the GPU of an NVIDIA Jetson TX2 NX for about 13 minutes across 30 plus 30 epochs and reached a final mIoU above 0.9; the two tracking experiments on January 26 and 30, 2026 brought mean pointing error down to 7.4 and 13.5 pixels, while earlier runs oscillated visibly.

A 6U CubeSat trained and ran a thermal-infrared segmentation model onboard in low Earth orbit, then fed the model's output into the attitude determination and control system to point at the Moon in closed loop from a live thermal image stream. Training ran on the GPU of an NVIDIA Jetson TX2 NX for about 13 minutes across 30 plus 30 epochs and reached a final mIoU above 0.9; the two tracking experiments on January 26 and 30, 2026 brought mean pointing error down to 7.4 and 13.5 pixels, while earlier runs oscillated visibly.

Weak-lensing shear data from the third year of the Dark Energy Survey (DES Y3) have been analysed cosmologically with the integrated three-point correlation function (i3PCF), a higher-order statistic that correlates local aperture mass with surrounding shear, and for the first time a three-point shear statistic on survey data is treated with fully simulation-based inference. Jointly with the two-point correlation functions, the i3PCF is detected at a signal-to-noise ratio of 12.8, and the figure of merit in the Ωm–S8–w space roughly doubles over two-point statistics alone, with the gain coming from angular scales where baryonic effects are subdominant.

Weak-lensing shear data from the third year of the Dark Energy Survey (DES Y3) have been analysed cosmologically with the integrated three-point correlation function (i3PCF), a higher-order statistic that correlates local aperture mass with surrounding shear, and for the first time a three-point shear statistic on survey data is treated with fully simulation-based inference. Jointly with the two-point correlation functions, the i3PCF is detected at a signal-to-noise ratio of 12.8, and the figure of merit in the Ωm–S8–w space roughly doubles over two-point statistics alone, with the gain coming from angular scales where baryonic effects are subdominant.

In verbal memory tests, the order of recalled words and the pauses between them carry diagnostic information, but a pause means different things inside a semantic category and at a switch between categories. TrajLift treats a recall episode as a timed trajectory on a semantic hierarchy, uses heat diffusion to lift each word onto several hierarchical scales, and lets the inter-item latency modulate the diffusion time so structure and timing enter one operator. On Hong Kong List Learning Test data from 387 participants after excluding 15 with annotation issues, accuracy on the 30-minute delayed-recall trial was 80.38%, 4.5 percentage points above a handcrafted baseline using recall count alone; the paired difference on the 10-minute trial was not significant.

In verbal memory tests, the order of recalled words and the pauses between them carry diagnostic information, but a pause means different things inside a semantic category and at a switch between categories. TrajLift treats a recall episode as a timed trajectory on a semantic hierarchy, uses heat diffusion to lift each word onto several hierarchical scales, and lets the inter-item latency modulate the diffusion time so structure and timing enter one operator. On Hong Kong List Learning Test data from 387 participants after excluding 15 with annotation issues, accuracy on the 30-minute delayed-recall trial was 80.38%, 4.5 percentage points above a handcrafted baseline using recall count alone; the paired difference on the 10-minute trial was not significant.

VeRL-Omni v0.3.0 extends asynchronous execution to multimodal reinforcement-learning training across text, audio input and diffusion models, and makes the V1 trainer built on TransferQueue and ReplayBuffer the default. Previously a strictly alternating loop left generation and training idle in turn. In MiniCPM-o 4.5 audio-visual question answering, separate-async ran about 9% slower per prompt on natural short answers but about 6.7% faster per prompt at forced 1024-token responses, so the execution mode should follow the workload.

VeRL-Omni v0.3.0 extends asynchronous execution to multimodal reinforcement-learning training across text, audio input and diffusion models, and makes the V1 trainer built on TransferQueue and ReplayBuffer the default. Previously a strictly alternating loop left generation and training idle in turn. In MiniCPM-o 4.5 audio-visual question answering, separate-async ran about 9% slower per prompt on natural short answers but about 6.7% faster per prompt at forced 1024-token responses, so the execution mode should follow the workload.

NVIDIA has open-sourced NVIDIA Snapshot, a checkpoint-and-restore component for GPU inference workloads on Kubernetes that captures a fully initialized workload, including CPU and GPU memory, as a portable artifact and restores it on nodes with the same GPU model and driver, skipping weight loading, compilation and graph capture. Across vLLM, SGLang and TensorRT LLM, from a 600-million-parameter model to mixture-of-experts models with 1 TB checkpoints on eight GPUs, every tested configuration restored a replica in under 45 seconds, 8x to 36x faster than a cold start.

NVIDIA has open-sourced NVIDIA Snapshot, a checkpoint-and-restore component for GPU inference workloads on Kubernetes that captures a fully initialized workload, including CPU and GPU memory, as a portable artifact and restores it on nodes with the same GPU model and driver, skipping weight loading, compilation and graph capture. Across vLLM, SGLang and TensorRT LLM, from a 600-million-parameter model to mixture-of-experts models with 1 TB checkpoints on eight GPUs, every tested configuration restored a replica in under 45 seconds, 8x to 36x faster than a cold start.

ExecuTorch adds a WebGPU backend that lets PyTorch models exported once with torch.export to a .pte file run on the GPU directly in WebGPU-enabled browsers, without first converting to another framework's format such as GGUF, ONNX or TFLite. Previously, running a PyTorch model in the browser usually meant doing that conversion first. On Chrome on an Apple M1 Pro, 4-bit quantized Llama 3.2 1B and Qwen3-0.6B ran 4–7× faster on prefill and 2–10× faster on decode than ExecuTorch's own Wasm CPU backend (XNNPACK).

ExecuTorch adds a WebGPU backend that lets PyTorch models exported once with torch.export to a .pte file run on the GPU directly in WebGPU-enabled browsers, without first converting to another framework's format such as GGUF, ONNX or TFLite. Previously, running a PyTorch model in the browser usually meant doing that conversion first. On Chrome on an Apple M1 Pro, 4-bit quantized Llama 3.2 1B and Qwen3-0.6B ran 4–7× faster on prefill and 2–10× faster on decode than ExecuTorch's own Wasm CPU backend (XNNPACK).

Clinical survival models usually emit distribution parameters, abstract quantities that do not map onto the risk factors clinicians recognize. ODDS-DR reparameterizes the network outputs so they add up in log-odds, letting each output neuron be attributed directly to input features. On the SUPPORT2 and aortic coarctation cohorts its discrimination matches or exceeds DeepSurv, DeepHit and DSM, with a C-index of 0.83 for the log-logistic variant on SUPPORT2.

Clinical survival models usually emit distribution parameters, abstract quantities that do not map onto the risk factors clinicians recognize. ODDS-DR reparameterizes the network outputs so they add up in log-odds, letting each output neuron be attributed directly to input features. On the SUPPORT2 and aortic coarctation cohorts its discrimination matches or exceeds DeepSurv, DeepHit and DSM, with a C-index of 0.83 for the log-logistic variant on SUPPORT2.

For kinetic plasmas governed by the Vlasov–Maxwell system, feedback control has either taken the full phase-space perturbation as input or relied on instantaneous field cancellation, which degrades once sensing and actuation delays appear. This controller instead uses macroscopic quantities such as current densities, electromagnetic fields and their spatial gradients as feedback information, and a low-rank neural operator learns a residual correction to the delayed cancellation baseline while the control objective stays kinetic: stabilizing the full distribution toward a prescribed equilibrium. In numerical experiments on two-stream and Weibel-type current-filamentation instabilities, the controller keeps suppressing perturbation growth over long times under substantial feedback delay.

For kinetic plasmas governed by the Vlasov–Maxwell system, feedback control has either taken the full phase-space perturbation as input or relied on instantaneous field cancellation, which degrades once sensing and actuation delays appear. This controller instead uses macroscopic quantities such as current densities, electromagnetic fields and their spatial gradients as feedback information, and a low-rank neural operator learns a residual correction to the delayed cancellation baseline while the control objective stays kinetic: stabilizing the full distribution toward a prescribed equilibrium. In numerical experiments on two-stream and Weibel-type current-filamentation instabilities, the controller keeps suppressing perturbation growth over long times under substantial feedback delay.

ECG foundation models are usually compared on in-domain interpretation tasks, leaving their transfer to longitudinal prognosis across healthcare systems uncertain. PACE-ECG treats two recordings from the same patient within 60 days as a contrastive positive pair and adds waveform- and report-derived multitask supervision, with an encoder of 2.21 million parameters. On MIMIC-IV and two external HEEDB cohorts, its frozen representations predict post-myocardial-infarction all-cause mortality and incident heart failure with higher AUROC than DeepECG-SSL, with the largest gains in external heart failure.

ECG foundation models are usually compared on in-domain interpretation tasks, leaving their transfer to longitudinal prognosis across healthcare systems uncertain. PACE-ECG treats two recordings from the same patient within 60 days as a contrastive positive pair and adds waveform- and report-derived multitask supervision, with an encoder of 2.21 million parameters. On MIMIC-IV and two external HEEDB cohorts, its frozen representations predict post-myocardial-infarction all-cause mortality and incident heart failure with higher AUROC than DeepECG-SSL, with the largest gains in external heart failure.

Clinical multi-turn agents must decide what to ask, when to stop, and what to do from incomplete information. Across three controlled simulated settings, medication refill triage, acute diagnostic work-up and dosing titration, the study compares reward designs: outcome-only GRPO often raises average value by asking fewer questions while producing premature commitment and weaker high-risk performance, whereas extra credit for questions that reduce uncertainty about the hidden state raises held-out value and moderate-case support across three training seeds in diagnostics without the high-risk seed collapse seen under outcome-only training.

Clinical multi-turn agents must decide what to ask, when to stop, and what to do from incomplete information. Across three controlled simulated settings, medication refill triage, acute diagnostic work-up and dosing titration, the study compares reward designs: outcome-only GRPO often raises average value by asking fewer questions while producing premature commitment and weaker high-risk performance, whereas extra credit for questions that reduce uncertainty about the hidden state raises held-out value and moderate-case support across three training seeds in diagnostics without the high-risk seed collapse seen under outcome-only training.

Low-dose, sparse-view and metal-artifact CT degradations have different physics, yet universal restoration models have processed all of them with one shared architecture. AgentCT uses a LoRA-adapted Qwen3-VL-4B controller to identify degradation type and severity, dispatch the image to an image-domain diffusion denoiser or a dual-domain specialist, and re-check the restored output. On nine degradation-severity settings of patient-disjoint Mayo 2020 abdominal CT, routing with ground-truth degradation labels gave an overall PSNR/SSIM of 46.18 dB/0.984, versus 44.93 dB/0.976 for AMIR and 44.29 dB/0.979 for TAMP.

Low-dose, sparse-view and metal-artifact CT degradations have different physics, yet universal restoration models have processed all of them with one shared architecture. AgentCT uses a LoRA-adapted Qwen3-VL-4B controller to identify degradation type and severity, dispatch the image to an image-domain diffusion denoiser or a dual-domain specialist, and re-check the restored output. On nine degradation-severity settings of patient-disjoint Mayo 2020 abdominal CT, routing with ground-truth degradation labels gave an overall PSNR/SSIM of 46.18 dB/0.984, versus 44.93 dB/0.976 for AMIR and 44.29 dB/0.979 for TAMP.

Vercel Labs has released agent-desktop, which lets AI agents operate desktop apps on macOS, Windows and Linux without screenshots or pixel matching. It reads an app's accessibility tree as JSON, assigns a ref to every element the agent can act on, and the agent clicks or types on a ref, then reads the UI again to verify. The tool ships as a CLI and a C library, contains no model, and leaves the observe-act-verify loop in the agent.

Vercel Labs has released agent-desktop, which lets AI agents operate desktop apps on macOS, Windows and Linux without screenshots or pixel matching. It reads an app's accessibility tree as JSON, assigns a ref to every element the agent can act on, and the agent clicks or types on a ref, then reads the UI again to verify. The tool ships as a CLI and a C library, contains no model, and leaves the observe-act-verify loop in the agent.

Complex-disease risk often acts through regulatory programs that appear only after a cell is perturbed, and steady-state molecular atlases capture them poorly. Perturbscape couples multiscale gene programs learned for each perturbation in CRISPR Perturb-seq screens with human genetics, evaluated across 15,859 perturbation-by-context profiles from 7 Perturb-seq systems and 53 diseases and traits. Perturbation-induced programs reached up to 4.3-fold higher heritability informativeness than steady-state multi-tissue gene prioritization and retained significant conditional heritability informativeness in all 34 matched cell system–trait comparisons after accounting for steady-state prioritization.

Complex-disease risk often acts through regulatory programs that appear only after a cell is perturbed, and steady-state molecular atlases capture them poorly. Perturbscape couples multiscale gene programs learned for each perturbation in CRISPR Perturb-seq screens with human genetics, evaluated across 15,859 perturbation-by-context profiles from 7 Perturb-seq systems and 53 diseases and traits. Perturbation-induced programs reached up to 4.3-fold higher heritability informativeness than steady-state multi-tissue gene prioritization and retained significant conditional heritability informativeness in all 34 matched cell system–trait comparisons after accounting for steady-state prioritization.

Antibody screening usually splits specificity, affinity and developability across separate models, and structural information is often compressed into a single confidence score such as ipTM. FoldAbS freezes the open-source folding model Protenix as a shared encoder, trains only lightweight task heads, and supplies specificity supervision with pseudo-negatives built from antigen mismatch, CDRH3 mutation and CDRH3 replacement. With Protenix as backbone, FoldAbS-Px raises macro AUROC on Origin-1 from 0.6690 to 0.7570 and reports the best results on affinity and four developability properties.

Antibody screening usually splits specificity, affinity and developability across separate models, and structural information is often compressed into a single confidence score such as ipTM. FoldAbS freezes the open-source folding model Protenix as a shared encoder, trains only lightweight task heads, and supplies specificity supervision with pseudo-negatives built from antigen mismatch, CDRH3 mutation and CDRH3 replacement. With Protenix as backbone, FoldAbS-Px raises macro AUROC on Origin-1 from 0.6690 to 0.7570 and reports the best results on affinity and four developability properties.

Anthropic has extended interactive connectors from a small live view inside the conversation to full tool interfaces that open beside the chat or go full screen, and Claude can now make changes directly to a page, row, or file the user has open, with each action approved the way connector actions are approved. Until now these connectors in Claude mainly appeared as interactive views within the conversation; Claude now supports more of the MCP Apps extension, including app tools and embedded sites, starting with a small group of tools including Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel, and Webflow.

Anthropic has extended interactive connectors from a small live view inside the conversation to full tool interfaces that open beside the chat or go full screen, and Claude can now make changes directly to a page, row, or file the user has open, with each action approved the way connector actions are approved. Until now these connectors in Claude mainly appeared as interactive views within the conversation; Claude now supports more of the MCP Apps extension, including app tools and embedded sites, starting with a small group of tools including Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel, and Webflow.

AirLLM pushes large-model inference memory down to a few gigabytes on a single card: Kimi K3 (2.8T) measured end to end at 3.72GB of VRAM on one RTX 6000 Ada, Qwen3.8-Flash-Next (125B) at 5.95GB on one RTX 4090, and DeepSeek-V3 (671B) at about 12GB. It uses no quantization, distillation or pruning; instead it streams weights layer by layer, and for MoE models loads only the experts each token actually routes to. The project extends the same idea to training: frozen weights stream in one layer at a time while adapters stay on the GPU, letting a 125B model train under 6GB.

AirLLM pushes large-model inference memory down to a few gigabytes on a single card: Kimi K3 (2.8T) measured end to end at 3.72GB of VRAM on one RTX 6000 Ada, Qwen3.8-Flash-Next (125B) at 5.95GB on one RTX 4090, and DeepSeek-V3 (671B) at about 12GB. It uses no quantization, distillation or pruning; instead it streams weights layer by layer, and for MoE models loads only the experts each token actually routes to. The project extends the same idea to training: frozen weights stream in one layer at a time while adapters stay on the GPU, letting a 125B model train under 6GB.

Whether hydrogen sulphide physisorbs or dissociates on nanoscale refractory oxide surfaces had been hard to settle because the configuration space is too large to enumerate. This work pairs unbiased automated adsorption sampling with an off-the-shelf atomistic foundation model: MACE's MPA screens candidate minima, each survivor is re-optimised with PBE-D2*(N) and HSE06 DFT, and DFT multiconfigurational infrared spectra are compared with in situ spectra at 150 K. The computed spectra reproduce the red-shift order of the H–S stretch across the three materials: silica at 2603 cm⁻¹, forsterite at 2588 cm⁻¹ and MgO at 2504 cm⁻¹.

Whether hydrogen sulphide physisorbs or dissociates on nanoscale refractory oxide surfaces had been hard to settle because the configuration space is too large to enumerate. This work pairs unbiased automated adsorption sampling with an off-the-shelf atomistic foundation model: MACE's MPA screens candidate minima, each survivor is re-optimised with PBE-D2*(N) and HSE06 DFT, and DFT multiconfigurational infrared spectra are compared with in situ spectra at 150 K. The computed spectra reproduce the red-shift order of the H–S stretch across the three materials: silica at 2603 cm⁻¹, forsterite at 2588 cm⁻¹ and MgO at 2504 cm⁻¹.

Bayesian parameter inference for extreme-mass-ratio inspirals (EMRIs) has relied on narrow priors centered on the true parameters or Markov chains initialized near them, information real observations do not provide. A deep-learning-assisted framework first uses a neural network to produce approximate posteriors from the data and shrink the search region, then particle swarm optimization finds a high-likelihood starting point, and staged Markov chain Monte Carlo sampling produces the final posteriors. On two noise-free six-month simulated signals, the resulting posteriors agree closely with reference analyses initialized near the true values, and once the network is trained each signal takes about ten hours.

Bayesian parameter inference for extreme-mass-ratio inspirals (EMRIs) has relied on narrow priors centered on the true parameters or Markov chains initialized near them, information real observations do not provide. A deep-learning-assisted framework first uses a neural network to produce approximate posteriors from the data and shrink the search region, then particle swarm optimization finds a high-likelihood starting point, and staged Markov chain Monte Carlo sampling produces the final posteriors. On two noise-free six-month simulated signals, the resulting posteriors agree closely with reference analyses initialized near the true values, and once the network is trained each signal takes about ten hours.

Compact models for ferroelectric field-effect transistors (FeFETs) have required either deriving physics-based equations term by term or training neural networks on smooth high-order derivatives from numerical solvers. This work trains a neural network on measured drain current instead, using scaled spectral normalization to constrain the network weights, smooth derivatives and filter measurement noise, covering the full range of polarization amplitudes and supporting dynamic polarization switching in transient simulations. Write and read operations of memory arrays up to large scale, including parasitics, complete transient simulation in commercial SPICE, and all converge.

Compact models for ferroelectric field-effect transistors (FeFETs) have required either deriving physics-based equations term by term or training neural networks on smooth high-order derivatives from numerical solvers. This work trains a neural network on measured drain current instead, using scaled spectral normalization to constrain the network weights, smooth derivatives and filter measurement noise, covering the full range of polarization amplitudes and supporting dynamic polarization switching in transient simulations. Write and read operations of memory arrays up to large scale, including parasitics, complete transient simulation in commercial SPICE, and all converge.

A clinical image classifier that returns a set of possible diagnoses does not guarantee that a severe case avoids a harmful action. HC-ROCP adds a harm penalty weighted by predicted severe-class probability to frozen classifier probabilities, tests coverage and severe-action error on disjoint calibration splits, and then certifies, reports descriptively, or refuses each claim. On COVID-19 radiographs across ten calibration seeds, severe-action errors fall from the exact ROCP baseline's 4.68% to 0.54%, at higher modeled loss.

A clinical image classifier that returns a set of possible diagnoses does not guarantee that a severe case avoids a harmful action. HC-ROCP adds a harm penalty weighted by predicted severe-class probability to frozen classifier probabilities, tests coverage and severe-action error on disjoint calibration splits, and then certifies, reports descriptively, or refuses each claim. On COVID-19 radiographs across ten calibration seeds, severe-action errors fall from the exact ROCP baseline's 4.68% to 0.54%, at higher modeled loss.

The Babylonian Astronomical Diaries record the sky for nearly six centuries, and many broken tablets have lost their date formula, which specialists previously dated one tablet at a time by comparing the surviving observations with computed phenomena. The new system turns observations into fixed-field clauses with a rule-based parser, has a language model check them against the text and add missed ones, and tests the clauses in all 701 years from 626 BCE to AD 75. On 120 dated tablets with decisive clauses, the top-ranked group contains the published year in 96% of cases; for 86 of them the group is a single year, the published year in 84.

The Babylonian Astronomical Diaries record the sky for nearly six centuries, and many broken tablets have lost their date formula, which specialists previously dated one tablet at a time by comparing the surviving observations with computed phenomena. The new system turns observations into fixed-field clauses with a rule-based parser, has a language model check them against the text and add missed ones, and tests the clauses in all 701 years from 626 BCE to AD 75. On 120 dated tablets with decisive clauses, the top-ranked group contains the published year in 96% of cases; for 86 of them the group is a single year, the published year in 84.

When a platform or planner changes an incentive rule, it usually has to wait for trajectory data under that rule before knowing how the population will shift. PA-RmD keeps the payoff function as an explicit input and learns only the payoff-independent population response structure from historical trajectories, so it can predict the induced trajectory under a new payoff directly. On held-out payoffs in synthetic zero-sum cyclic and potential congestion games, and on held-out countries in a ten-country YouTube Trending panel, its one-step, full-rollout and final-state errors are lower than SINDy, Neural ODE, LEADS, CoDA and SIAR.

When a platform or planner changes an incentive rule, it usually has to wait for trajectory data under that rule before knowing how the population will shift. PA-RmD keeps the payoff function as an explicit input and learns only the payoff-independent population response structure from historical trajectories, so it can predict the induced trajectory under a new payoff directly. On held-out payoffs in synthetic zero-sum cyclic and potential congestion games, and on held-out countries in a ten-country YouTube Trending panel, its one-step, full-rollout and final-state errors are lower than SINDy, Neural ODE, LEADS, CoDA and SIAR.

Small models specialized for one domain usually inherit the multilingual vocabulary of a general base model, so domain text is split into many pieces and every generated token scores tens of thousands of entries the task never uses. Dharma-AI's VocabX changes the selection criterion from token frequency to inference cost weighted by field, rebuilding sub-word segmentation while merging repeated spans into macro tokens. On a Portuguese essay grader built on Qwen3-4B, with the same base model, data and training recipe, average response length fell from 350 to 101 tokens, measured inference cost dropped 77.0%, and the bench score rose from 0.703 to 0.810.

Small models specialized for one domain usually inherit the multilingual vocabulary of a general base model, so domain text is split into many pieces and every generated token scores tens of thousands of entries the task never uses. Dharma-AI's VocabX changes the selection criterion from token frequency to inference cost weighted by field, rebuilding sub-word segmentation while merging repeated spans into macro tokens. On a Portuguese essay grader built on Qwen3-4B, with the same base model, data and training recipe, average response length fell from 350 to 101 tokens, measured inference cost dropped 77.0%, and the bench score rose from 0.703 to 0.810.

Emergency departments rarely have time for PTSD screening. PREVIA feeds the transcript of a patient's answer to the routine question "What brought you to the ED?" into a clinically informed ensemble of Llama, Mistral and DeepSeek models arranged in three layers: four clinical dimensions are assessed first, synthesized into a structured summary, then combined with age, sex and trauma type into a 0–100 risk score with a short explanation. Among 145 trauma survivors at two New York City level-1 trauma centers, the ensemble reached AUROC 78.49%±5.75% and sensitivity 95.4%±3.99% for one-month PTSD diagnosis; on 30 randomly selected narratives it scored AUROC 84.39% versus 83.86% for an emergency clinician specialized in PTSD.

Emergency departments rarely have time for PTSD screening. PREVIA feeds the transcript of a patient's answer to the routine question "What brought you to the ED?" into a clinically informed ensemble of Llama, Mistral and DeepSeek models arranged in three layers: four clinical dimensions are assessed first, synthesized into a structured summary, then combined with age, sex and trauma type into a 0–100 risk score with a short explanation. Among 145 trauma survivors at two New York City level-1 trauma centers, the ensemble reached AUROC 78.49%±5.75% and sensitivity 95.4%±3.99% for one-month PTSD diagnosis; on 30 randomly selected narratives it scored AUROC 84.39% versus 83.86% for an emergency clinician specialized in PTSD.

Omics-based cancer outcome prediction usually assigns samples to discrete ancestry groups and fits a separate model within each group, yet accuracy falls continuously with genetic distance from the training set. DA-ICL treats ancestry as a continuous variable, projects genotypes into a low-dimensional space with principal component analysis, builds a context from nearby samples at each ancestry point, and lets a pretrained tabular foundation model predict at inference time, with distribution-aware context enhancement adding distant but distributionally similar samples. Across 18 tasks spanning three TCGA cancer types, three omics modalities and two clinical endpoints, DA-ICL reaches the highest mean AUROC on 10 tasks for African-ancestry samples, with gains that vary by task.

Omics-based cancer outcome prediction usually assigns samples to discrete ancestry groups and fits a separate model within each group, yet accuracy falls continuously with genetic distance from the training set. DA-ICL treats ancestry as a continuous variable, projects genotypes into a low-dimensional space with principal component analysis, builds a context from nearby samples at each ancestry point, and lets a pretrained tabular foundation model predict at inference time, with distribution-aware context enhancement adding distant but distributionally similar samples. Across 18 tasks spanning three TCGA cancer types, three omics modalities and two clinical endpoints, DA-ICL reaches the highest mean AUROC on 10 tasks for African-ancestry samples, with gains that vary by task.

GRAC represents a surgical scene as a graph, retrieves graph-caption pairs of similar surgical activities from the training set, and compares instrument-tissue interactions between the current and retrieved scenes to condition caption generation. On the combined EndoVis-18 benchmark it reaches a CIDEr-D of 5.88, against 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline, and an exact instrument-action-target triplet-set accuracy of 52.6%, against 32.7% for SGT and 26.2% for the strongest conventional baseline. On TORS it reaches a CIDEr-D of 5.06, against 3.80 for the strongest baseline.

GRAC represents a surgical scene as a graph, retrieves graph-caption pairs of similar surgical activities from the training set, and compares instrument-tissue interactions between the current and retrieved scenes to condition caption generation. On the combined EndoVis-18 benchmark it reaches a CIDEr-D of 5.88, against 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline, and an exact instrument-action-target triplet-set accuracy of 52.6%, against 32.7% for SGT and 26.2% for the strongest conventional baseline. On TORS it reaches a CIDEr-D of 5.06, against 3.80 for the strongest baseline.

Anthropic added three collaborative artifact types to Claude, Sheets, Whiteboards and Data dashboards, and upgraded interactive connectors so partner apps can be embedded directly in Claude. Claude could already produce documents, slide decks and designs, but team-editable spreadsheets and whiteboards were not among them. An internal legal team's Slack-based review tool built with Claude cut marketing review from three days to 24 hours, and the finance team turned a long business update into an interactive page the whole team maintains.

Anthropic added three collaborative artifact types to Claude, Sheets, Whiteboards and Data dashboards, and upgraded interactive connectors so partner apps can be embedded directly in Claude. Claude could already produce documents, slide decks and designs, but team-editable spreadsheets and whiteboards were not among them. An internal legal team's Slack-based review tool built with Claude cut marketing review from three days to 24 hours, and the finance team turned a long business update into an interactive page the whole team maintains.

Longitudinal whole-body PET/CT registration has relied mainly on raw image intensities or sparse segmentation labels to establish correspondence; intensities are fragile under cross-modal appearance change, and labels assign only coarse identities to fine anatomy. This work instead uses per-voxel semantic representations from the 3D foundation model anatomix, combined with a hand-crafted MIND-SSC descriptor and segmentation labels, to drive the FireANTs solver through two optimization stages with lesion-volume preservation constraints. On the 2026 Learn2Reg PSMAReg Challenge test set the system ranked first without training on the challenge training set, with MTV error 1.78% and TLG error 1.39%.

Longitudinal whole-body PET/CT registration has relied mainly on raw image intensities or sparse segmentation labels to establish correspondence; intensities are fragile under cross-modal appearance change, and labels assign only coarse identities to fine anatomy. This work instead uses per-voxel semantic representations from the 3D foundation model anatomix, combined with a hand-crafted MIND-SSC descriptor and segmentation labels, to drive the FireANTs solver through two optimization stages with lesion-volume preservation constraints. On the 2026 Learn2Reg PSMAReg Challenge test set the system ranked first without training on the challenge training set, with MTV error 1.78% and TLG error 1.39%.

Neural surrogates for three-dimensional fluid simulation need large sets of costly high-fidelity CFD labels. SimpFlow instead generates supervision from simplified physics: a source-panel potential-flow solver that needs only a surface mesh labels about 9,000 unlabeled ShapeNet geometries with velocity and pressure fields, reusing precomputed velocity fields across inflow directions for an amortized 0.111 s per example. A Transolver pretrained on these labels and fine-tuned on independent CFD datasets improves over training from scratch in all 24 task–budget settings across five flow tasks; with fine-tuning fixed at 200 epochs, interpolation indicates 35–68% fewer CFD training cases are needed to match scratch error at each task's largest budget.

Neural surrogates for three-dimensional fluid simulation need large sets of costly high-fidelity CFD labels. SimpFlow instead generates supervision from simplified physics: a source-panel potential-flow solver that needs only a surface mesh labels about 9,000 unlabeled ShapeNet geometries with velocity and pressure fields, reusing precomputed velocity fields across inflow directions for an amortized 0.111 s per example. A Transolver pretrained on these labels and fine-tuned on independent CFD datasets improves over training from scratch in all 24 task–budget settings across five flow tasks; with fine-tuning fixed at 200 epochs, interpolation indicates 35–68% fewer CFD training cases are needed to match scratch error at each task's largest budget.

In reinforcement-learning post-training, the training engine and the rollout engine can compute different logprobs even with the same model, weights and inputs, because kernels, parallelism and reduction order differ, which affects the importance ratio, KL diagnostics and clipping. vime aligns the training timeline while RL-Kernel aligns numerical execution, and together they put attention, feed-forward, logprob and communication paths on one numerical contract on AMD Instinct MI300X. In a Qwen3-8B GRPO experiment on 8 MI300X GPUs, the strict path ran 200 consecutive steps with mismatch_count and max_abs_diff both zero.

In reinforcement-learning post-training, the training engine and the rollout engine can compute different logprobs even with the same model, weights and inputs, because kernels, parallelism and reduction order differ, which affects the importance ratio, KL diagnostics and clipping. vime aligns the training timeline while RL-Kernel aligns numerical execution, and together they put attention, feed-forward, logprob and communication paths on one numerical contract on AMD Instinct MI300X. In a Qwen3-8B GRPO experiment on 8 MI300X GPUs, the strict path ran 200 consecutive steps with mismatch_count and max_abs_diff both zero.

Anthropic released beta versions of Claude Whiteboards and Claude Sheets and made the former Claude Dashboards generally available as Claude Data. Working with Claude on a project previously meant moving output to another tool where Claude had no live access to follow changes; now boards, sheets and dashboards stay in Claude, teams can edit together, and outputs can still go to .xlsx, CSV or partner tools such as Lucid, Miro and Airtable. The three work on web, desktop and mobile, and Claude Data needs a data connector.

Anthropic released beta versions of Claude Whiteboards and Claude Sheets and made the former Claude Dashboards generally available as Claude Data. Working with Claude on a project previously meant moving output to another tool where Claude had no live access to follow changes; now boards, sheets and dashboards stay in Claude, teams can edit together, and outputs can still go to .xlsx, CSV or partner tools such as Lucid, Miro and Airtable. The three work on web, desktop and mobile, and Claude Data needs a data connector.

CryoEM can reconstruct a density map yet often cannot say which molecule the map contains, and matching hundreds or thousands of mass-spectrometry or genome-derived candidates by hand is a bottleneck. CryoID2 uses three 3D U-Nets to predict structure classes, key-atom positions and reduced residue or nucleotide types from the map, turns them into short sequence queries that rank candidates, then rigidly fits the candidates' PDB or AlphaFold models back into the density and ranks them by agreement. In a 166-protein mass-spectrometry pool it placed PF3D7_1446200/PfA-M17 first and selected PDB 4ZX8 as the best structural model; on the RNA side it ranked the Ca-raiA-containing interval fourth in a genome-derived non-coding pool, and structural evaluation then put PDB 9E73 first.

CryoEM can reconstruct a density map yet often cannot say which molecule the map contains, and matching hundreds or thousands of mass-spectrometry or genome-derived candidates by hand is a bottleneck. CryoID2 uses three 3D U-Nets to predict structure classes, key-atom positions and reduced residue or nucleotide types from the map, turns them into short sequence queries that rank candidates, then rigidly fits the candidates' PDB or AlphaFold models back into the density and ranks them by agreement. In a 166-protein mass-spectrometry pool it placed PF3D7_1446200/PfA-M17 first and selected PDB 4ZX8 as the best structural model; on the RNA side it ranked the Ca-raiA-containing interval fourth in a genome-derived non-coding pool, and structural evaluation then put PDB 9E73 first.

The vLLM Project's Speculators training library now supports Kimi K3, a 2.8-trillion-parameter model, and produced a five-layer, five-billion-parameter DSpark draft model that combines parallel block drafting with sequential correction. On math-reasoning workloads, single-stream interactivity rises from about 110 to about 435 tok/s/user, output throughput reaches up to about 3.5x at matched interactivity under concurrent load, and the macro-average acceptance length across nine evaluation domains is 4.11 tokens per verification round.

The vLLM Project's Speculators training library now supports Kimi K3, a 2.8-trillion-parameter model, and produced a five-layer, five-billion-parameter DSpark draft model that combines parallel block drafting with sequential correction. On math-reasoning workloads, single-stream interactivity rises from about 110 to about 435 tok/s/user, output throughput reaches up to about 3.5x at matched interactivity under concurrent load, and the macro-average acceptance length across nine evaluation domains is 4.11 tokens per verification round.

Hugging Face has fixed a common interface for decision models: given a state and several typed questions, the model returns option probabilities, an ordered score or a yes/no probability without generating any text. Previously such structured judgments relied on an LLM emitting JSON, sometimes backed by constrained decoding, or on fixed-label and zero-shot classifiers. Transformers, llama.cpp and a Hub tag now follow the same contract, and compatible autoregressive language models can use the interface without retraining.

Hugging Face has fixed a common interface for decision models: given a state and several typed questions, the model returns option probabilities, an ordered score or a yes/no probability without generating any text. Previously such structured judgments relied on an LLM emitting JSON, sometimes backed by constrained decoding, or on fixed-label and zero-shot classifiers. Transformers, llama.cpp and a Hub tag now follow the same contract, and compatible autoregressive language models can use the interface without retraining.

Hugging Face has defined a common interface for decision models: given a state and a set of typed questions, a model returns option probabilities, an ordered score or a true/false probability without generating text token by token. Previously such structured judgments relied on prompting a large language model for JSON, on fixed-label classifiers, or on zero-shot classification. Transformers adds DecisionPipeline and related classes, llama.cpp serves the same /v1/systemone endpoint, and compatible autoregressive models can score answer options without producing text.

Hugging Face has defined a common interface for decision models: given a state and a set of typed questions, a model returns option probabilities, an ordered score or a true/false probability without generating text token by token. Previously such structured judgments relied on prompting a large language model for JSON, on fixed-label classifiers, or on zero-shot classification. Transformers adds DecisionPipeline and related classes, llama.cpp serves the same /v1/systemone endpoint, and compatible autoregressive models can score answer options without producing text.

Gravitational-wave population studies must repeatedly ask whether each of millions of simulated binary black holes would be detected, and each answer conventionally requires generating a waveform and computing a noise-weighted inner product. asmaNet maps 11 binary parameters directly to the phase-averaged optimal signal-to-noise ratio for a single detector, beating MLP, 1D-ResNet and Transformer-encoder baselines at every dataset scale from 100k to 1M samples under a matched training protocol with only about 732k parameters. On the 1M-sample held-out test set it reaches 99.29% accuracy after rank-preserving monotonic calibration, sustaining roughly 120,000 SNR evaluations per second on one GPU.

Gravitational-wave population studies must repeatedly ask whether each of millions of simulated binary black holes would be detected, and each answer conventionally requires generating a waveform and computing a noise-weighted inner product. asmaNet maps 11 binary parameters directly to the phase-averaged optimal signal-to-noise ratio for a single detector, beating MLP, 1D-ResNet and Transformer-encoder baselines at every dataset scale from 100k to 1M samples under a matched training protocol with only about 732k parameters. On the 1M-sample held-out test set it reaches 99.29% accuracy after rank-preserving monotonic calibration, sustaining roughly 120,000 SNR evaluations per second on one GPU.

Adobe and Google announced an expanded partnership that puts Adobe's creative and productivity tools directly inside Google products people use every day, while bringing Google's frontier AI models into Adobe's apps. The arrangements include a direct path from Google Drive and Gmail into Adobe Acrobat, expanded Acrobat and creative-tool connectors in Gemini and Gemini Enterprise, and Adobe creative tools in AI Mode in Google Search. Businesses will be able to use Adobe's CX Enterprise Coworker agent in Gemini Enterprise, and Adobe and Google DeepMind will explore image-model editing for Photoshop.

Adobe and Google announced an expanded partnership that puts Adobe's creative and productivity tools directly inside Google products people use every day, while bringing Google's frontier AI models into Adobe's apps. The arrangements include a direct path from Google Drive and Gmail into Adobe Acrobat, expanded Acrobat and creative-tool connectors in Gemini and Gemini Enterprise, and Adobe creative tools in AI Mode in Google Search. Businesses will be able to use Adobe's CX Enterprise Coworker agent in Gemini Enterprise, and Adobe and Google DeepMind will explore image-model editing for Photoshop.

A recurrent neural controller must keep behaving while neurons are born, move, and change their intrinsic time constants during development; controllers optimized only on the initial or the mature substrate generally failed at other ages. The study feeds developmental time into a compact connectivity generator and optimizes the same generator at three, five, or fifteen developmental snapshots, testing it on two geometrically distinct substrates and two visuomotor tasks. Optimization at only three snapshots was often enough to keep performance high and consistent at trained and untrained ages, while the generated connectomes changed substantially with age and did not converge on a single sparsity or weight trajectory.

A recurrent neural controller must keep behaving while neurons are born, move, and change their intrinsic time constants during development; controllers optimized only on the initial or the mature substrate generally failed at other ages. The study feeds developmental time into a compact connectivity generator and optimizes the same generator at three, five, or fifteen developmental snapshots, testing it on two geometrically distinct substrates and two visuomotor tasks. Optimization at only three snapshots was often enough to keep performance high and consistent at trained and untrained ages, while the generated connectomes changed substantially with age and did not converge on a single sparsity or weight trajectory.

MiniMax released H3-Image, an open-weight image generation model built on H3 that supports text-to-image, reference-based generation and image editing, and can produce a 4096×4096 4K image in under 10 seconds, which MiniMax says is up to 5× the speed of mainstream alternatives. Previously, 4K output carried more cost than the common 1K or 2K sizes, which kept it out of everyday creative workflows. The speed comes from changes across the whole post-training pipeline, including data, reinforcement learning, sampling-step reduction, PE and deployment optimization, rather than from a single change.

MiniMax released H3-Image, an open-weight image generation model built on H3 that supports text-to-image, reference-based generation and image editing, and can produce a 4096×4096 4K image in under 10 seconds, which MiniMax says is up to 5× the speed of mainstream alternatives. Previously, 4K output carried more cost than the common 1K or 2K sizes, which kept it out of everyday creative workflows. The speed comes from changes across the whole post-training pipeline, including data, reinforcement learning, sampling-step reduction, PE and deployment optimization, rather than from a single change.

At KubeCon + CloudNativeCon, NVIDIA said it is donating OpenShell, an open source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation (CNCF) as a Sandbox project, while also contributing to Google-led Agent Substrate and releasing three new components of the DSX AI Factory Platform. OpenShell applies deterministic controls to agent activity so teams can enforce policies as agents use tools, access inference services and execute workloads, and its extensible architecture connects their own compute, inference providers, credential storage and monitoring tools. NVIDIA says hundreds of organizations have put OpenShell to use and the project has received tens of thousands of contributions in less than a year since launch.

At KubeCon + CloudNativeCon, NVIDIA said it is donating OpenShell, an open source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation (CNCF) as a Sandbox project, while also contributing to Google-led Agent Substrate and releasing three new components of the DSX AI Factory Platform. OpenShell applies deterministic controls to agent activity so teams can enforce policies as agents use tools, access inference services and execute workloads, and its extensible architecture connects their own compute, inference providers, credential storage and monitoring tools. NVIDIA says hundreds of organizations have put OpenShell to use and the project has received tens of thousands of contributions in less than a year since launch.

Cloth simulators usually cut motion into fixed timesteps and return states only on the solver grid, so querying arbitrary times or derivatives means re-running or finite differences. ContinuousCloth represents the cloth trajectory as a differentiable function of time, trained under the residual of a spring-mass ordinary differential equation, and answers position and derivative queries at any timestamp. Across six boundary-condition scenarios its trajectory accuracy matches a high-precision fourth-order Runge-Kutta reference while dense-time evaluation runs up to 17.4 times faster.

Cloth simulators usually cut motion into fixed timesteps and return states only on the solver grid, so querying arbitrary times or derivatives means re-running or finite differences. ContinuousCloth represents the cloth trajectory as a differentiable function of time, trained under the residual of a spring-mass ordinary differential equation, and answers position and derivative queries at any timestamp. Across six boundary-condition scenarios its trajectory accuracy matches a high-precision fourth-order Runge-Kutta reference while dense-time evaluation runs up to 17.4 times faster.

A self-driving-laboratory testbed built from a commercial lab robot, an absorbance plate reader and a swappable optimizer completed six independent campaigns of 48 closed-loop iterations each, over 24 cumulative hours of unattended operation. Five of the six campaigns matched their target absorbance spectra within measurement error, at an average RMSE of OD. In the same physical loop, an agentic optimizer driven by Claude Opus 5 matched all three targets in fewer iterations than Bayesian optimization, cutting iterations by 52% on target 1 and 32% on target 2, and converging in 11 iterations on target 3, which Bayesian optimization failed to match within 48.

A self-driving-laboratory testbed built from a commercial lab robot, an absorbance plate reader and a swappable optimizer completed six independent campaigns of 48 closed-loop iterations each, over 24 cumulative hours of unattended operation. Five of the six campaigns matched their target absorbance spectra within measurement error, at an average RMSE of OD. In the same physical loop, an agentic optimizer driven by Claude Opus 5 matched all three targets in fewer iterations than Bayesian optimization, cutting iterations by 52% on target 1 and 32% on target 2, and converging in 11 iterations on target 3, which Bayesian optimization failed to match within 48.

NVIDIA reframes AI-factory memory from how much capacity a GPU has to how many useful tokens each byte produces, combining NVFP4 weight and KV-cache compression, wide expert parallelism for MoE weights, Dynamo KV-cache reuse, Vera CPU staging of context during tool calls, and CMX pod-level context storage. On SemiAnalysis AgentX, which replays recorded coding-agent sessions, Vera Rubin NVL72 delivers up to 30x the throughput per megawatt of GB300 NVL72 at 160 tokens per second per user, with both at 288 GB per GPU.

NVIDIA reframes AI-factory memory from how much capacity a GPU has to how many useful tokens each byte produces, combining NVFP4 weight and KV-cache compression, wide expert parallelism for MoE weights, Dynamo KV-cache reuse, Vera CPU staging of context during tool calls, and CMX pod-level context storage. On SemiAnalysis AgentX, which replays recorded coding-agent sessions, Vera Rubin NVL72 delivers up to 30x the throughput per megawatt of GB300 NVL72 at 160 tokens per second per user, with both at 288 GB per GPU.

Periodic Neon is a trillion-parameter model post-trained with midtraining and reinforcement learning on data from the company's own labs and deployed to its high-throughput physical labs in Menlo Park to analyze X-ray diffraction results. It outperforms Claude Fable 5.1 and GPT-6 Astra on diffraction analysis, and inference is optimized so the model can be deployed at scale. Its training data comes from physical experiments rather than elastically scalable digital environments.

Periodic Neon is a trillion-parameter model post-trained with midtraining and reinforcement learning on data from the company's own labs and deployed to its high-throughput physical labs in Menlo Park to analyze X-ray diffraction results. It outperforms Claude Fable 5.1 and GPT-6 Astra on diffraction analysis, and inference is optimized so the model can be deployed at scale. Its training data comes from physical experiments rather than elastically scalable digital environments.

Soft demapping in mmWave MIMO receivers no longer treats each symbol on its own. Prior learned demappers used multilayer perceptrons on single symbols, while classical demappers rely on analytical decision rules and a Gaussian noise assumption, so neither exploits residual distortion structure across neighboring symbols. This work applies a one-dimensional convolutional network to whole blocks of equalized symbols; on over-the-air data from a 60 GHz, four-stream, 1.76 GHz-per-stream link, the single-stream variant reaches up to 2.6× BER gain over the conventional soft demapper for 16QAM and up to 17.6× for 64QAM, and a fixed-point RFSoC implementation runs about 34.3× faster.

Soft demapping in mmWave MIMO receivers no longer treats each symbol on its own. Prior learned demappers used multilayer perceptrons on single symbols, while classical demappers rely on analytical decision rules and a Gaussian noise assumption, so neither exploits residual distortion structure across neighboring symbols. This work applies a one-dimensional convolutional network to whole blocks of equalized symbols; on over-the-air data from a 60 GHz, four-stream, 1.76 GHz-per-stream link, the single-stream variant reaches up to 2.6× BER gain over the conventional soft demapper for 16QAM and up to 17.6× for 64QAM, and a fixed-point RFSoC implementation runs about 34.3× faster.

Google Cloud has released Gemini for healthcare, aimed at unifying fragmented clinical, operational and administrative systems so care teams can turn siloed data into coordinated action. Through specialized skills and data connectors to existing third-party systems, the platform pulls in information from electronic health records, billing platforms, clinical guidelines and Medicare datasets. It is initially available in preview with launch partners including HCA Healthcare, Highmark Health and Northwell Health.

Google Cloud has released Gemini for healthcare, aimed at unifying fragmented clinical, operational and administrative systems so care teams can turn siloed data into coordinated action. Through specialized skills and data connectors to existing third-party systems, the platform pulls in information from electronic health records, billing platforms, clinical guidelines and Medicare datasets. It is initially available in preview with launch partners including HCA Healthcare, Highmark Health and Northwell Health.

Mercor announced a $750 million Series D at a $20 billion valuation, a round that included an employee tender and participation from Benchmark, NVIDIA, Salesforce and others. For the past two years Mercor built infrastructure to train and evaluate frontier models on real-world work for frontier labs, drawing on a network of more than five million experts paid to train models from their own knowledge. It is now offering that expert network, evaluation methods and post-training infrastructure to big tech, open-model developers, application-layer companies and Fortune 500 firms, so enterprises can post-train a capable open model and build custom evals around their own workflows.

Mercor announced a $750 million Series D at a $20 billion valuation, a round that included an employee tender and participation from Benchmark, NVIDIA, Salesforce and others. For the past two years Mercor built infrastructure to train and evaluate frontier models on real-world work for frontier labs, drawing on a network of more than five million experts paid to train models from their own knowledge. It is now offering that expert network, evaluation methods and post-training infrastructure to big tech, open-model developers, application-layer companies and Fortune 500 firms, so enterprises can post-train a capable open model and build custom evals around their own workflows.

Conservation laws for Newtonian systems usually require writing the equations of motion and then solving symmetry partial differential equations that depend on the potential. Here the dynamics is embedded through the Eisenhart-Duval lift into an extended metric, so the symmetry conditions become Killing or conformal Killing equations of that metric, solved by a Hamiltonian network that learns the potential from phase-space data plus a random-feature network. For the one-dimensional Ermakov-Pinney equation, the three Killing vectors of the flat extended plane and their nonlocal conservation laws are recovered; for the one-dimensional harmonic oscillator, the ten-dimensional conformal algebra is reconstructed under noise.

Conservation laws for Newtonian systems usually require writing the equations of motion and then solving symmetry partial differential equations that depend on the potential. Here the dynamics is embedded through the Eisenhart-Duval lift into an extended metric, so the symmetry conditions become Killing or conformal Killing equations of that metric, solved by a Hamiltonian network that learns the potential from phase-space data plus a random-feature network. For the one-dimensional Ermakov-Pinney equation, the three Killing vectors of the flat extended plane and their nonlocal conservation laws are recovered; for the one-dimensional harmonic oscillator, the ten-dimensional conformal algebra is reconstructed under noise.

The hidden Markov Ising model makes an Ising model, a binary graphical model from statistical physics, the emission distribution of a hidden Markov chain, so that a binary network's dependence structure is modeled directly as its state changes over time. Because the Ising normalizing constant grows exponentially with the number of nodes, fitting uses pseudolikelihood inside Baum-Welch reestimation, with results on the algorithm's validity, model identifiability and large-network consistency. Exogenous covariates extend it to a hidden Markov autologistic model, applied to a five-year sequence of weekly gun-death events across 3109 continental U.S. counties from 2021 to 2025, where it identifies state changes in gun violence activity.

The hidden Markov Ising model makes an Ising model, a binary graphical model from statistical physics, the emission distribution of a hidden Markov chain, so that a binary network's dependence structure is modeled directly as its state changes over time. Because the Ising normalizing constant grows exponentially with the number of nodes, fitting uses pseudolikelihood inside Baum-Welch reestimation, with results on the algorithm's validity, model identifiability and large-network consistency. Exogenous covariates extend it to a hidden Markov autologistic model, applied to a five-year sequence of weekly gun-death events across 3109 continental U.S. counties from 2021 to 2025, where it identifies state changes in gun violence activity.

Antifreeze peptide discovery has been limited by narrow candidate diversity and reliance on known protein scaffolds. AFPFusionLM fuses sequence and structure-aware representations to reach an AUROC of 0.937 and an AUPRC of 0.957 on an independent reference set; paired with a LoRA-adapted generative model and multi-model consensus screening, it retained 87,305 candidate protein records from UniProt, glacier metagenomes and generated sequences and revealed periodic tandem-repeat architectures across several candidate clusters. Short peptides from the resulting repeat-associated families reached or exceeded the ice recrystallization inhibition of a literature reference peptide measured in the same batch.

Antifreeze peptide discovery has been limited by narrow candidate diversity and reliance on known protein scaffolds. AFPFusionLM fuses sequence and structure-aware representations to reach an AUROC of 0.937 and an AUPRC of 0.957 on an independent reference set; paired with a LoRA-adapted generative model and multi-model consensus screening, it retained 87,305 candidate protein records from UniProt, glacier metagenomes and generated sequences and revealed periodic tandem-repeat architectures across several candidate clusters. Short peptides from the resulting repeat-associated families reached or exceeded the ice recrystallization inhibition of a literature reference peptide measured in the same batch.

Ampersand, a language model from Humans&, is now available free on AI Gateway through November 13, 2026. It is built for multi-person conversations such as team chats, deciding on each turn whether to speak or stay quiet, and is designed to understand individual and shared goals and help people work through disagreements. An application exposes speaking and silence as two caller-defined tools, send_message and pass, which the application executes and records; the model accepts auto and none tool choices but rejects required and named ones.

Ampersand, a language model from Humans&, is now available free on AI Gateway through November 13, 2026. It is built for multi-person conversations such as team chats, deciding on each turn whether to speak or stay quiet, and is designed to understand individual and shared goals and help people work through disagreements. An application exposes speaking and silence as two caller-defined tools, send_message and pass, which the application executes and records; the model accepts auto and none tool choices but rejects required and named ones.

Cloudflare puts the identity problem that personal AI agents create for websites on the table: a remotely hosted agent acts for a user, yet the site sees only the agent and struggles to recognize, serve and hold accountable the user behind it. It proposes identity, reputation and accountability as the three mechanisms that preserve the site-user relationship, and has shipped three capabilities: listing personal assistants as a separate traffic category, setting access per agent, and showing at login and sign-up which agent made the request.

Cloudflare puts the identity problem that personal AI agents create for websites on the table: a remotely hosted agent acts for a user, yet the site sees only the agent and struggles to recognize, serve and hold accountable the user behind it. It proposes identity, reputation and accountability as the three mechanisms that preserve the site-user relationship, and has shipped three capabilities: listing personal assistants as a separate traffic category, setting access per agent, and showing at login and sign-up which agent made the request.

In marine seismic exploration, SRME (surface-related multiple elimination) predicts multiples and then subtracts them from the recorded data with a matching filter; the classical route uses either a fixed-length stationary filter or a non-stationary regression with a local smoothness constraint, which limits it where signal energy changes rapidly. PDL replaces that matching filter with a self-supervised neural network trained under the adaptive matching filtering equation, cutting the seismic profile into many patches so training yields a globally optimal non-stationary match. On synthetic data and the Viking Graben field dataset, PDL leaves fewer multiple residuals than the classical non-stationary least-squares matching filter and better preserves weak primaries and waveform continuity.

In marine seismic exploration, SRME (surface-related multiple elimination) predicts multiples and then subtracts them from the recorded data with a matching filter; the classical route uses either a fixed-length stationary filter or a non-stationary regression with a local smoothness constraint, which limits it where signal energy changes rapidly. PDL replaces that matching filter with a self-supervised neural network trained under the adaptive matching filtering equation, cutting the seismic profile into many patches so training yields a globally optimal non-stationary match. On synthetic data and the Viking Graben field dataset, PDL leaves fewer multiple residuals than the classical non-stationary least-squares matching filter and better preserves weak primaries and waveform continuity.

Anthropic introduced Claude Opus 5.5, the first model in its new Claude 5.5 family, which the company says performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Pricing is $4/$20 per million input/output tokens, 20% lower than Opus 5, with cache pricing at $0.20, 60% lower; Anthropic estimates typical overall costs will drop 40% and speed rise 30%. The model is available with zero data retention, its guardrails are similar to Fable 5.1, and the company highlights agentic coding, security and communication.

Anthropic introduced Claude Opus 5.5, the first model in its new Claude 5.5 family, which the company says performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Pricing is $4/$20 per million input/output tokens, 20% lower than Opus 5, with cache pricing at $0.20, 60% lower; Anthropic estimates typical overall costs will drop 40% and speed rise 30%. The model is available with zero data retention, its guardrails are similar to Fable 5.1, and the company highlights agentic coding, security and communication.

Tumor subtypes form a hierarchy running from primary cancer type down to molecular subtype, yet most classifiers flatten the labels and ignore the biological constraints between levels. HHRM combines cancer hallmark gene networks with recursive modules to predict level by level across 11,476 pan-cancer samples covering 26 primary cancers, 405 subtypes and up to eight levels, reaching 0.7924 accuracy in local output mode, above the hierarchical models it was compared with. Retrained on an independent set of 1,608 breast cancer samples, it exceeded 0.86 accuracy on both test and validation data.

Tumor subtypes form a hierarchy running from primary cancer type down to molecular subtype, yet most classifiers flatten the labels and ignore the biological constraints between levels. HHRM combines cancer hallmark gene networks with recursive modules to predict level by level across 11,476 pan-cancer samples covering 26 primary cancers, 405 subtypes and up to eight levels, reaching 0.7924 accuracy in local output mode, above the hierarchical models it was compared with. Retrained on an independent set of 1,608 breast cancer samples, it exceeded 0.86 accuracy on both test and validation data.

Anthropic's Claude Fable 5.2 is now callable through AI Gateway, built for open-ended technical work and improving on earlier Fable models in coding, reasoning and vision, and able to run long, ambiguous tasks unattended. It can act as an orchestrator that hands simpler tasks to smaller models, or as an advisor other Claude models consult on harder problems, with clearer communication and faster output. The model ships with cybersecurity and biology safety classifiers, and access requires data retention: prompts and completions are kept for 30 days and not used to train Claude, while teams with Zero Data Retention enabled get an error instead of a forwarded request.

Anthropic's Claude Fable 5.2 is now callable through AI Gateway, built for open-ended technical work and improving on earlier Fable models in coding, reasoning and vision, and able to run long, ambiguous tasks unattended. It can act as an orchestrator that hands simpler tasks to smaller models, or as an advisor other Claude models consult on harder problems, with clearer communication and faster output. The model ships with cybersecurity and biology safety classifiers, and access requires data retention: prompts and completions are kept for 30 days and not used to train Claude, while teams with Zero Data Retention enabled get an error instead of a forwarded request.

Micius-Flash is a 27-billion-parameter engineering model with native multimodal input and a 256K-token context window, specialized and evaluated primarily on semiconductor engineering, and released as an open-source model. A knowledge-tree-driven data flywheel organizes semiconductor knowledge, synthesizes data and validates quality, and multi-teacher on-policy distillation merges four expert capabilities, knowledge and reasoning, hardware coding, multimodal understanding and agentic tool use, into one student model. It averages 64.2% on 12 hardware coding benchmarks, 15.0 points above a same-size general-purpose baseline, and scores 62.4% on a semiconductor-specific agent benchmark, 17.2 points above that baseline.

Micius-Flash is a 27-billion-parameter engineering model with native multimodal input and a 256K-token context window, specialized and evaluated primarily on semiconductor engineering, and released as an open-source model. A knowledge-tree-driven data flywheel organizes semiconductor knowledge, synthesizes data and validates quality, and multi-teacher on-policy distillation merges four expert capabilities, knowledge and reasoning, hardware coding, multimodal understanding and agentic tool use, into one student model. It averages 64.2% on 12 hardware coding benchmarks, 15.0 points above a same-size general-purpose baseline, and scores 62.4% on a semiconductor-specific agent benchmark, 17.2 points above that baseline.

NVIDIA built the AI Factory Operations Agent for its own DGX SuperPOD operations so site reliability engineers no longer have to check compute, network, hardware and scheduling dashboards one by one. The agent is a set of domain-specific sub-agents that communicate over MCP, detect failures in production, correlate root causes across layers and surface recommendations through a chat interface. The framework runs as a Kubernetes control-plane application, its tool calls are constrained by read-only credentials and RBAC, and the orchestration framework is released as experimental open source software.

NVIDIA built the AI Factory Operations Agent for its own DGX SuperPOD operations so site reliability engineers no longer have to check compute, network, hardware and scheduling dashboards one by one. The agent is a set of domain-specific sub-agents that communicate over MCP, detect failures in production, correlate root causes across layers and surface recommendations through a chat interface. The framework runs as a Kubernetes control-plane application, its tool calls are constrained by read-only credentials and RBAC, and the orchestration framework is released as experimental open source software.

Clinical prediction models often arrive as files without the data they were trained on, so an auditor must work out what the model actually computes. Researchers trained tanh networks of 105 to 38,401 weights on simulated subarachnoid haemorrhage cohorts, planted known relationships or a label bias, and asked GPT-6 Astra to answer with no tools, with a calculation-only white-box tool, or with a black-box tool exposing only predictions. Without tools Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights; with the calculator it recovered the relationship in all 29 runs that had one, raised no false alarm in 15 null runs, and black-box matched white-box.

Clinical prediction models often arrive as files without the data they were trained on, so an auditor must work out what the model actually computes. Researchers trained tanh networks of 105 to 38,401 weights on simulated subarachnoid haemorrhage cohorts, planted known relationships or a label bias, and asked GPT-6 Astra to answer with no tools, with a calculation-only white-box tool, or with a black-box tool exposing only predictions. Without tools Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights; with the calculator it recovered the relationship in all 29 runs that had one, raised no false alarm in 15 null runs, and black-box matched white-box.

A MODIS pixel often covers one or more Prairie fields, so the coarse signal is a mixture of crop types and direct coarse-to-fine classifiers struggle. UnMixMamba uses the abundance field from unmixing as a shared intermediate, so reconstruction, classification and sparse refinement all read the same fractions, while a learnable per-crop variant bank absorbs intra-class spectral spread. On 183 spatially blocked Alberta 2015 test patches it reaches OA 0.8397, mIoU 0.6640, AA 0.7837 and Kappa 0.7715, ahead of ST-DRes and other baselines, with the best barley recall in the comparison at 0.5496; the task is closed-set four-crop mapping inside a known agricultural mask.

A MODIS pixel often covers one or more Prairie fields, so the coarse signal is a mixture of crop types and direct coarse-to-fine classifiers struggle. UnMixMamba uses the abundance field from unmixing as a shared intermediate, so reconstruction, classification and sparse refinement all read the same fractions, while a learnable per-crop variant bank absorbs intra-class spectral spread. On 183 spatially blocked Alberta 2015 test patches it reaches OA 0.8397, mIoU 0.6640, AA 0.7837 and Kappa 0.7715, ahead of ST-DRes and other baselines, with the best barley recall in the comparison at 0.5496; the task is closed-set four-crop mapping inside a known agricultural mask.

Google's Android team released a repository called Android skills that turns Android development tasks into modular technical specifications as SKILL.md files following the open agent-skills standard, giving AI agents information on specialized domains and workflows. The skills prioritize use cases where evaluations show large language models underperform, rather than areas where models are already proficient, such as basic Jetpack Compose best practices. Skills install through the Android CLI or the Claude Code and Codex plugin marketplaces, and agents activate relevant ones automatically.

Google's Android team released a repository called Android skills that turns Android development tasks into modular technical specifications as SKILL.md files following the open agent-skills standard, giving AI agents information on specialized domains and workflows. The skills prioritize use cases where evaluations show large language models underperform, rather than areas where models are already proficient, such as basic Jetpack Compose best practices. Skills install through the Android CLI or the Claude Code and Codex plugin marketplaces, and agents activate relevant ones automatically.

A radiology agentic framework separates a large-language-model orchestrator from a modular tool interface and uses 1,989 finding classes consolidated from FDA-cleared radiology AI devices as its tool catalog, simulating structured report generation on the PARROT fictional-report corpus. Planning from the clinical indication alone, actionable-finding recall is 0.290 with the cleared catalog and 0.437 with an unlimited catalog; when the planner reads the target report, these rise to 0.426 and 0.740. Tool outputs are simulated from the reference report, so the comparison isolates tool coverage and orchestration rather than tool error.

A radiology agentic framework separates a large-language-model orchestrator from a modular tool interface and uses 1,989 finding classes consolidated from FDA-cleared radiology AI devices as its tool catalog, simulating structured report generation on the PARROT fictional-report corpus. Planning from the clinical indication alone, actionable-finding recall is 0.290 with the cleared catalog and 0.437 with an unlimited catalog; when the planner reads the target report, these rise to 0.426 and 0.740. Tool outputs are simulated from the reference report, so the comparison isolates tool coverage and orchestration rather than tool error.

In preoperative multimodal cataract data, SLO images are often missing, and prior methods mostly suppressed the effect of the missing modality during fusion or filled it with zeros rather than explicitly recovering a task-relevant representation for it. This work treats missing SLO as proxy representation recovery in latent feature space: OCT image features and clinical text features condition a diffusion model whose reverse denoising estimates a sample-specific SLO proxy representation, which then enters the fusion prediction network. On 2,141 preoperative samples with a 58.80% SLO missing rate, the method reached 0.9714 accuracy under an absolute error within 0.1 LogMAR and a mean absolute error of 0.0423, versus 0.9625 and 0.0457 for the UML backbone baseline.

In preoperative multimodal cataract data, SLO images are often missing, and prior methods mostly suppressed the effect of the missing modality during fusion or filled it with zeros rather than explicitly recovering a task-relevant representation for it. This work treats missing SLO as proxy representation recovery in latent feature space: OCT image features and clinical text features condition a diffusion model whose reverse denoising estimates a sample-specific SLO proxy representation, which then enters the fusion prediction network. On 2,141 preoperative samples with a 58.80% SLO missing rate, the method reached 0.9714 accuracy under an absolute error within 0.1 LogMAR and a mean absolute error of 0.0423, versus 0.9625 and 0.0457 for the UML backbone baseline.

Soliton solutions of the KdV equation normally require retraining a physics-informed neural network whenever the dispersion parameter changes, and long-time extrapolation tends to drift. The scaled TW-PINN first normalizes the equation, trains once in normalized coordinates, and inverse-scales back to the original variables; the one-soliton solver uses a traveling-wave coordinate, and the two-soliton case trains in two steps, learning individual speeds first and the interaction second. In numerical experiments the two-soliton model preserved profiles ten times beyond the training domain, and a three-soliton experiment reproduced collision and separation.

Soliton solutions of the KdV equation normally require retraining a physics-informed neural network whenever the dispersion parameter changes, and long-time extrapolation tends to drift. The scaled TW-PINN first normalizes the equation, trains once in normalized coordinates, and inverse-scales back to the original variables; the one-soliton solver uses a traveling-wave coordinate, and the two-soliton case trains in two steps, learning individual speeds first and the interaction second. In numerical experiments the two-soliton model preserved profiles ten times beyond the training domain, and a three-soliton experiment reproduced collision and separation.

Tether AI Research released BotPsy, a 0.8B-parameter on-device personal assistant built on Qwen3.5-0.8B and post-trained in four stages for tool calling. Against its base model, BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and τ-bench from 23.2 to 51.5; it leads five of six tool-use and agentic benchmarks, trailing only on Claw-Eval behind the 2.5-times-larger Qwen3.5-2B-Think.

Tether AI Research released BotPsy, a 0.8B-parameter on-device personal assistant built on Qwen3.5-0.8B and post-trained in four stages for tool calling. Against its base model, BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and τ-bench from 23.2 to 51.5; it leads five of six tool-use and agentic benchmarks, trailing only on Claw-Eval behind the 2.5-times-larger Qwen3.5-2B-Think.

On two-dimensional decaying Navier–Stokes flow, three pre-registered studies separate three properties often conflated as evidence that a latent has learned physics: decodability, prominence and steerability. A joint-embedding predictive architecture (JEPA) beats an untrained network of the same architecture on three probe targets yet fails the pre-registered criteria for organizing its latent around fluid quantities; autoencoders that render the field make enstrophy prominent in their slowest unsupervised directions, and adding VICReg lowers that prominence in every comparison while raising linear decodability; a routed design yields a label-free loudness dial in all 8 trained models and a spectral-shape dial in only 3 of 5 fresh seeds, below the pre-registered bar of 4.

On two-dimensional decaying Navier–Stokes flow, three pre-registered studies separate three properties often conflated as evidence that a latent has learned physics: decodability, prominence and steerability. A joint-embedding predictive architecture (JEPA) beats an untrained network of the same architecture on three probe targets yet fails the pre-registered criteria for organizing its latent around fluid quantities; autoencoders that render the field make enstrophy prominent in their slowest unsupervised directions, and adding VICReg lowers that prominence in every comparison while raising linear decodability; a routed design yields a label-free loudness dial in all 8 trained models and a spectral-shape dial in only 3 of 5 fresh seeds, below the pre-registered bar of 4.

Sakana AI and Modal put Sakana Namazu, a roughly 1-trillion-parameter mixture-of-experts model, into production API service at NVFP4 precision. Compression was applied only to the routed experts that hold most of the parameters, while attention, the router and the shared experts stayed at BF16, and calibration used Japanese-heavy conversations so the model matched its production checkpoint on Japanese instruction following. A custom draft model trained on more than 30 billion raw tokens raised speculative token acceptance by 25% over a generic drafter; a second drafter retrained on about 1 million real requests lifted average acceptance length from 3.51 to 5.01 and cut p99 latency from 68s to 52s across 1,000 held-out requests.

Sakana AI and Modal put Sakana Namazu, a roughly 1-trillion-parameter mixture-of-experts model, into production API service at NVFP4 precision. Compression was applied only to the routed experts that hold most of the parameters, while attention, the router and the shared experts stayed at BF16, and calibration used Japanese-heavy conversations so the model matched its production checkpoint on Japanese instruction following. A custom draft model trained on more than 30 billion raw tokens raised speculative token acceptance by 25% over a generic drafter; a second drafter retrained on about 1 million real requests lifted average acceptance length from 3.51 to 5.01 and cut p99 latency from 68s to 52s across 1,000 held-out requests.

OpenAI's Daybreak Red brings purpose-trained cybersecurity models into AWS GovCloud (US-East), letting approved government security teams call GPT‑5.6 Cyber inside the AWS environments they use for critical missions. Daybreak models were already available on Amazon Bedrock; this expansion extends availability to a government cloud region. The model supports authorized vulnerability research, exploit validation and security testing, helping defenders investigate complex software weaknesses, reproduce vulnerabilities in controlled settings and develop mitigations, while qualified practitioners must still review findings and proposed fixes.

OpenAI's Daybreak Red brings purpose-trained cybersecurity models into AWS GovCloud (US-East), letting approved government security teams call GPT‑5.6 Cyber inside the AWS environments they use for critical missions. Daybreak models were already available on Amazon Bedrock; this expansion extends availability to a government cloud region. The model supports authorized vulnerability research, exploit validation and security testing, helping defenders investigate complex software weaknesses, reproduce vulnerabilities in controlled settings and develop mitigations, while qualified practitioners must still review findings and proposed fixes.

Individual-level trajectory generators can produce plausible single traces, yet once many individuals are aggregated their displacement distribution often misses the heavy tail of real data. ScaleTraj splits next-location prediction into two steps: a Scale Selection Head decides which spatial scale a move belongs to, and a Scale-Aware Cell Selection Head predicts the destination conditioned on that scale. On GeoLife and Tencent Beijing it lowers both the Scale Fidelity Profile error and the deviations of the fitted truncated power-law parameters relative to the compared model-driven and data-driven baselines.

Individual-level trajectory generators can produce plausible single traces, yet once many individuals are aggregated their displacement distribution often misses the heavy tail of real data. ScaleTraj splits next-location prediction into two steps: a Scale Selection Head decides which spatial scale a move belongs to, and a Scale-Aware Cell Selection Head predicts the destination conditioned on that scale. On GeoLife and Tencent Beijing it lowers both the Scale Fidelity Profile error and the deviations of the fitted truncated power-law parameters relative to the compared model-driven and data-driven baselines.

OpenAI announced a partnership with the security firm Hacktron in which engineers from both organizations and Hacktron's security researchers jointly pick high-priority areas to test, focusing on how OpenAI's systems interact with third-party services. OpenAI came to know Hacktron through its Security Bug Bounty program; the team builds AI agents for code review, penetration testing and red teaming, and its research has uncovered vulnerabilities in how applications, cloud infrastructure and identity systems interact. As part of the partnership, OpenAI is expanding Hacktron's access to its frontier models through Daybreak Red.

OpenAI announced a partnership with the security firm Hacktron in which engineers from both organizations and Hacktron's security researchers jointly pick high-priority areas to test, focusing on how OpenAI's systems interact with third-party services. OpenAI came to know Hacktron through its Security Bug Bounty program; the team builds AI agents for code review, penetration testing and red teaming, and its research has uncovered vulnerabilities in how applications, cloud infrastructure and identity systems interact. As part of the partnership, OpenAI is expanding Hacktron's access to its frontier models through Daybreak Red.

Tether AI Research released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows that runs on-device. Trained through supervised fine-tuning, multi-teacher on-policy distillation and two reinforcement-learning rounds, it lifts BFCL v4 from its base's 23.4 to 44.0 and τ-bench from 23.2 to 51.5, and beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tool-use and agentic benchmarks, trailing only on Claw-Eval (46.5 vs 48.9). It still makes mistakes, targets low-to-medium difficulty tasks and is not intended for high-stakes use.

Tether AI Research released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows that runs on-device. Trained through supervised fine-tuning, multi-teacher on-policy distillation and two reinforcement-learning rounds, it lifts BFCL v4 from its base's 23.4 to 44.0 and τ-bench from 23.2 to 51.5, and beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tool-use and agentic benchmarks, trailing only on Claw-Eval (46.5 vs 48.9). It still makes mistakes, targets low-to-medium difficulty tasks and is not intended for high-stakes use.

When a retrieval-augmented generation corpus must be trimmed under a storage or annotation budget, how much query coverage the chosen fragments lose relative to the best feasible set has lacked a finite-sample guarantee that survives data-dependent selection. This work separates selection from auditing: a pilot picks the set, screens competitors and builds a comparison network, then an independent audit acquires relevance labels with probabilities constant within observable strata and yields a simultaneous bound on all retained coverage contrasts, with simulations confirming the nominal level.

When a retrieval-augmented generation corpus must be trimmed under a storage or annotation budget, how much query coverage the chosen fragments lose relative to the best feasible set has lacked a finite-sample guarantee that survives data-dependent selection. This work separates selection from auditing: a pilot picks the set, screens competitors and builds a comparison network, then an independent audit acquires relevance labels with probabilities constant within observable strata and yields a simultaneous bound on all retained coverage contrasts, with simulations confirming the nominal level.

Google is extending Gemini Intelligence from the Pixel 11 and Samsung Galaxy Z Fold8 Ultra, Z Fold8, Z Flip8 and S26 series that began shipping this summer to Motorola's first such device, the motorola signature 27, and to the Pixel 10 series. The new Create My Widget builds a personalized widget from a plain-language description with no coding, pulling from Gmail, Calendar and Search, while Device Assistance lets users ask the phone to free up space, check app permissions or mute distractions. Rambler turns spoken thoughts into organized text in real time, task automation expands to the UK, India and Japan, and proactive suggestions are coming to Android Auto.

Google is extending Gemini Intelligence from the Pixel 11 and Samsung Galaxy Z Fold8 Ultra, Z Fold8, Z Flip8 and S26 series that began shipping this summer to Motorola's first such device, the motorola signature 27, and to the Pixel 10 series. The new Create My Widget builds a personalized widget from a plain-language description with no coding, pulling from Gmail, Calendar and Search, while Device Assistance lets users ask the phone to free up space, check app permissions or mute distractions. Rambler turns spoken thoughts into organized text in real time, task automation expands to the UK, India and Japan, and proactive suggestions are coming to Android Auto.

Meta introduced Your algorithm, which lets Facebook users state in natural language what they want and do not want to see, with those preferences applied to both Feed and Reels. Previously the recommendation systems learned mainly from implicit engagement signals, which scale well but adapt slowly and capture little nuance. The system uses a collection of large language models to infer interests, validate preference changes and perform a final rerank, while a supplementary fetch and cache invalidation make changes take effect faster; Meta notes that interpreting ambiguous input and responding more precisely remain open problems.

Meta introduced Your algorithm, which lets Facebook users state in natural language what they want and do not want to see, with those preferences applied to both Feed and Reels. Previously the recommendation systems learned mainly from implicit engagement signals, which scale well but adapt slowly and capture little nuance. The system uses a collection of large language models to infer interests, validate preference changes and perform a final rerank, while a supplementary fetch and cache invalidation make changes take effect faster; Meta notes that interpreting ambiguous input and responding more precisely remain open problems.

Preferred Networks built PLaMo 3.0 Fin on PLaMo 3 NICT, a pretrained model from joint research with NICT, by further training on Japanese finance data, and released it as an open-weight finance-focused small language model. Aimed at financial institutions' operations, it reasons before answering with the reasoning trace in Japanese and can call external tools such as web search, calculation and internal databases. Its small size lets it run from a single GPU in closed cloud or on-premises environments, positioned as the core model of AI agents handling tasks such as inquiry classification, document extraction and alert analysis.

Preferred Networks built PLaMo 3.0 Fin on PLaMo 3 NICT, a pretrained model from joint research with NICT, by further training on Japanese finance data, and released it as an open-weight finance-focused small language model. Aimed at financial institutions' operations, it reasons before answering with the reasoning trace in Japanese and can call external tools such as web search, calculation and internal databases. Its small size lets it run from a single GPU in closed cloud or on-premises environments, positioned as the core model of AI agents handling tasks such as inquiry classification, document extraction and alert analysis.

At the MSIT-NVIDIA AI Startup Partnership Forum at COEX in Seoul, NVIDIA announced an investment in Korean startup Point2 and plans to invest in Motif Technologies, RLWRLD and Upstage. The four companies span AI interconnect, foundation models for agents, robotics foundation models and enterprise document AI, covering the AI stack from infrastructure to models to robotics applications. The announcement states investment intent; NVIDIA disclosed no amounts, terms or closing dates.

At the MSIT-NVIDIA AI Startup Partnership Forum at COEX in Seoul, NVIDIA announced an investment in Korean startup Point2 and plans to invest in Motif Technologies, RLWRLD and Upstage. The four companies span AI interconnect, foundation models for agents, robotics foundation models and enterprise document AI, covering the AI stack from infrastructure to models to robotics applications. The announcement states investment intent; NVIDIA disclosed no amounts, terms or closing dates.

Mistral has moved Mistral Large 4 on AI Gateway to updated weights and lowered the price of input, cached input and output tokens. According to Mistral's release announcement, the new weights improve agentic coding and defensive cybersecurity, use tokens more efficiently on coding tasks, and also improve spreadsheet creation and editing, web research, and resistance to prompt injection. Users already calling mistral/mistral-large-4 through AI Gateway need no model change to get the new weights and pricing.

Mistral has moved Mistral Large 4 on AI Gateway to updated weights and lowered the price of input, cached input and output tokens. According to Mistral's release announcement, the new weights improve agentic coding and defensive cybersecurity, use tokens more efficiently on coding tasks, and also improve spreadsheet creation and editing, web research, and resistance to prompt injection. Users already calling mistral/mistral-large-4 through AI Gateway need no model change to get the new weights and pricing.

IBM researchers released Palette, which splits deck building into a designer that writes a structured brief and a coder that writes pptxgenjs code for each slide, both served by one LoRA adapter, and trains a 20B open-weight model on roughly 1,000 synthetic decks covering about 10,000 slides. The training data follows the IBM Design Language, and every generated slide is rendered and measured by a geometry checker, with repairs kept only when the checker confirms improvement. On PresentBench, Palette ranks first in Economics and Education, third in Talk, and second on the three-domain mean at 59.7 versus NotebookLM's 60.8, 1.1 points behind.

IBM researchers released Palette, which splits deck building into a designer that writes a structured brief and a coder that writes pptxgenjs code for each slide, both served by one LoRA adapter, and trains a 20B open-weight model on roughly 1,000 synthetic decks covering about 10,000 slides. The training data follows the IBM Design Language, and every generated slide is rendered and measured by a geometry checker, with repairs kept only when the checker confirms improvement. On PresentBench, Palette ranks first in Economics and Education, third in Talk, and second on the three-domain mean at 59.7 versus NotebookLM's 60.8, 1.1 points behind.

Younger breast cancer patients have too few samples to train reliable prognostic models, so LSAT-EN treats older patients as a source cohort and sets the strength of borrowing toward the source coefficients separately for each of six omics layers, while elastic-net regularization selects variables. In TCGA-BRCA, with patients aged 60 and older as source and those under 50 as target, four-year binary survival prediction reached an AUC of 0.819, above 0.774 for pooled fitting, 0.757 for target-only fitting, 0.728 for glmtrans and 0.780 for GAS-EN. In simulation the advantage grew as source-target differences and target-sample scarcity increased.

Younger breast cancer patients have too few samples to train reliable prognostic models, so LSAT-EN treats older patients as a source cohort and sets the strength of borrowing toward the source coefficients separately for each of six omics layers, while elastic-net regularization selects variables. In TCGA-BRCA, with patients aged 60 and older as source and those under 50 as target, four-year binary survival prediction reached an AUC of 0.819, above 0.774 for pooled fitting, 0.757 for target-only fitting, 0.728 for glmtrans and 0.780 for GAS-EN. In simulation the advantage grew as source-target differences and target-sample scarcity increased.

Several U.S. Department of Energy Genesis Mission teams are integrating Undermind directly into the autonomous research systems they are building, so agents can investigate prior work, evaluate scientific evidence and carry findings into later stages while keeping citations scientists can inspect. The mission brings together national laboratories, universities and industry with the goal of doubling the productivity and impact of American research within a decade; Undermind is an official member of the Genesis Mission Consortium, whose directory describes its role as core infrastructure for the autonomous research loop.

Several U.S. Department of Energy Genesis Mission teams are integrating Undermind directly into the autonomous research systems they are building, so agents can investigate prior work, evaluate scientific evidence and carry findings into later stages while keeping citations scientists can inspect. The mission brings together national laboratories, universities and industry with the goal of doubling the productivity and impact of American research within a decade; Undermind is an official member of the Genesis Mission Consortium, whose directory describes its role as core infrastructure for the autonomous research loop.

OpenAI has created a new role, Senior Counselor for National Security Modernization, and named Dan Driscoll, the 26th U.S. Secretary of the Army, to fill it. He will advise OpenAI leadership and partners across government, industry, academia and allied institutions on how advanced AI can strengthen national and cyber security, with a focus on responsible use of democratic AI with U.S. allies, and will help internal OpenAI teams understand the operational needs of global national security institutions.

OpenAI has created a new role, Senior Counselor for National Security Modernization, and named Dan Driscoll, the 26th U.S. Secretary of the Army, to fill it. He will advise OpenAI leadership and partners across government, industry, academia and allied institutions on how advanced AI can strengthen national and cyber security, with a focus on responsible use of democratic AI with U.S. allies, and will help internal OpenAI teams understand the operational needs of global national security institutions.

OpenAI has updated its Model Spec, the public description of how it wants its models to behave, reorganizing the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The previous version spread these commitments across separate principles alongside extensive detailed style guidance; the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and retires most of the style detail, while calling the model "GPT" rather than "the assistant." OpenAI says its models do not yet fully reflect the Spec and that results from related training work will come in future updates.

OpenAI has updated its Model Spec, the public description of how it wants its models to behave, reorganizing the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The previous version spread these commitments across separate principles alongside extensive detailed style guidance; the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and retires most of the style detail, while calling the model "GPT" rather than "the assistant." OpenAI says its models do not yet fully reflect the Spec and that results from related training work will come in future updates.

AMD announced general availability of its 6th Gen EPYC server SP7 processors on the new Zen 6 architecture, with systems shipping from major server vendors and Azure and OCI cloud instances to follow soon. AMD claims the general-purpose EPYC 9996, in dual-socket systems on SPECrate 2017_int_base, delivers up to 2.2x the performance and 2.0x the performance per dollar of an Intel Xeon 6980P system, and up to 3.7x an Nvidia Vera system. The Vera comparison rests on AMD internal estimates, and the company positions the line as a workload-optimized portfolio rather than one model leading everywhere.

AMD announced general availability of its 6th Gen EPYC server SP7 processors on the new Zen 6 architecture, with systems shipping from major server vendors and Azure and OCI cloud instances to follow soon. AMD claims the general-purpose EPYC 9996, in dual-socket systems on SPECrate 2017_int_base, delivers up to 2.2x the performance and 2.0x the performance per dollar of an Intel Xeon 6980P system, and up to 3.7x an Nvidia Vera system. The Vera comparison rests on AMD internal estimates, and the company positions the line as a workload-optimized portfolio rather than one model leading everywhere.

ElevenLabs added synthetic voice detection to ElevenAgents, which picks up traces in a live audio stream and can tell whether a caller is human or AI, so agencies can set queueing, routing and limits on sensitive actions. Governments facing a surge of requests filed by AI agents had previously relied on fees, IP blocks and stricter forms to add friction. On Czech employment and benefits hotlines, agents handle roughly 5,000 calls a day, independently resolve around 85% of calls, and raise answer rates from about 60% to near-universal coverage.

ElevenLabs added synthetic voice detection to ElevenAgents, which picks up traces in a live audio stream and can tell whether a caller is human or AI, so agencies can set queueing, routing and limits on sensitive actions. Governments facing a surge of requests filed by AI agents had previously relied on fees, IP blocks and stricter forms to add friction. On Czech employment and benefits hotlines, agents handle roughly 5,000 calls a day, independently resolve around 85% of calls, and raise answer rates from about 60% to near-universal coverage.

Asymptote Labs has open-sourced Beacon, a memory layer that works across agent harnesses. Previously, switching among coding agents such as Claude Code, Codex and Cursor meant re-explaining tasks, workflows and common mistake patterns, and the team found no good free alternative. Beacon hooks into more than 20 agent harnesses, normalizes session telemetry such as prompts, tool calls and command executions into one format, distills it into reusable patterns with a human-approval step, and exposes the result to later agents through skills or MCP.

Asymptote Labs has open-sourced Beacon, a memory layer that works across agent harnesses. Previously, switching among coding agents such as Claude Code, Codex and Cursor meant re-explaining tasks, workflows and common mistake patterns, and the team found no good free alternative. Beacon hooks into more than 20 agent harnesses, normalizes session telemetry such as prompts, tool calls and command executions into one format, distills it into reusable patterns with a human-approval step, and exposes the result to later agents through skills or MCP.

Macro merges email, messages, docs, tasks, calls, agents, CRM and databases into one workspace with shared team-level memory, where everything is @-linked and searchable. It targets the situation where a growing team runs separate tools held together by MCP and Zapier. Mail, messages, docs and tasks share one backend, cross-references are stored natively as a bidirectional graph, and agents can search PDF attachments parsed from email through a unified search tool.

Macro merges email, messages, docs, tasks, calls, agents, CRM and databases into one workspace with shared team-level memory, where everything is @-linked and searchable. It targets the situation where a growing team runs separate tools held together by MCP and Zapier. Mail, messages, docs and tasks share one backend, cross-references are stored natively as a bidirectional graph, and agents can search PDF attachments parsed from email through a unified search tool.

Lantern Pharma reported Phase 1a data for LP-184 (zirdafulven) in recurrent glioblastoma and set an expansion dose of 0.39 mg/kg. Because only about 20% of systemic drug is expected to reach brain tumors, single-agent brain concentrations are not projected to stay at therapeutic levels, so the company plans a Phase 1b/2a trial combining LP-184 with spironolactone, which degrades the DNA repair protein ERCC3. In preclinical models, spironolactone lowered the LP-184 concentration needed to kill tumor cells 3- to 6-fold.

Lantern Pharma reported Phase 1a data for LP-184 (zirdafulven) in recurrent glioblastoma and set an expansion dose of 0.39 mg/kg. Because only about 20% of systemic drug is expected to reach brain tumors, single-agent brain concentrations are not projected to stay at therapeutic levels, so the company plans a Phase 1b/2a trial combining LP-184 with spironolactone, which degrades the DNA repair protein ERCC3. In preclinical models, spironolactone lowered the LP-184 concentration needed to kill tumor cells 3- to 6-fold.

Cloudflare has added Zero Wire to Radar, a hub of briefings for investigating critical Internet vulnerabilities and threats. Each briefing separates confirmed facts from open questions, lists remediation steps, and shows what Cloudflare observes about the threat; briefings are grouped by urgency, searchable by CVE, name or tag, and followable via RSS.

Cloudflare has added Zero Wire to Radar, a hub of briefings for investigating critical Internet vulnerabilities and threats. Each briefing separates confirmed facts from open questions, lists remediation steps, and shows what Cloudflare observes about the threat; briefings are grouped by urgency, searchable by CVE, name or tag, and followable via RSS.