AI and science frontiers · 2026-11-10

The Aaronson–Ambainis conjecture had been proved only for special classes of polynomials, while general bounded low-degree polynomials had influence bounds decaying exponentially in the degree. This work gives an elementary proof that a function of degree at most d with values in [-1,1] has a variable of influence at least a negative constant power of d; in the original normalization, a polynomial with variance at least 1 also has a variable of influence at least the same order. Consequently, superpolynomial quantum speedups in the query model require structure or a promise on the inputs.

The Aaronson–Ambainis conjecture had been proved only for special classes of polynomials, while general bounded low-degree polynomials had influence bounds decaying exponentially in the degree. This work gives an elementary proof that a function of degree at most d with values in [-1,1] has a variable of influence at least a negative constant power of d; in the original normalization, a polynomial with variance at least 1 also has a variable of influence at least the same order. Consequently, superpolynomial quantum speedups in the query model require structure or a promise on the inputs.

Cytosine methylation at plant CG sites is gained and lost spontaneously, far faster than DNA point mutation. Until now a single site's steady state could only be estimated from the epigenome itself, leaving open whether it is written in DNA or carried by chromatin. REED, built on the NTv3 backbone, predicts every cytosine's methylation from 10 kb of sequence, and that prediction is the site's set point; in Arabidopsis pedigrees, the rice germline and a 330-year-old poplar, the direction and the rate-implied equilibrium of measured epimutations followed it.

Cytosine methylation at plant CG sites is gained and lost spontaneously, far faster than DNA point mutation. Until now a single site's steady state could only be estimated from the epigenome itself, leaving open whether it is written in DNA or carried by chromatin. REED, built on the NTv3 backbone, predicts every cytosine's methylation from 10 kb of sequence, and that prediction is the site's set point; in Arabidopsis pedigrees, the rice germline and a 330-year-old poplar, the direction and the rate-implied equilibrium of measured epimutations followed it.

Most robot policies output action values at discrete control steps and leave the continuity of physical motion to be learned implicitly from data. ODE Policy instead learns a neural ordinary differential equation for the vector field of the action, and at deployment integrates each action from the previous one, so the executed trajectory is continuous by construction. After one epoch on LIBERO it reaches 79.2% success against at most 43.2% for the baselines, and it keeps that early lead as the action head of a VLA and of ImageWAM and as the actor of SuperMimic.

Most robot policies output action values at discrete control steps and leave the continuity of physical motion to be learned implicitly from data. ODE Policy instead learns a neural ordinary differential equation for the vector field of the action, and at deployment integrates each action from the previous one, so the executed trajectory is continuous by construction. After one epoch on LIBERO it reaches 79.2% success against at most 43.2% for the baselines, and it keeps that early lead as the action head of a VLA and of ImageWAM and as the actor of SuperMimic.

Treating childhood medulloblastoma and ependymoma means trading recurrence risk against the long-term side effects of radiotherapy, yet current clinical indicators and molecular subgroups only give group-level risk. MANTIS takes genome-wide DNA methylation array data plus gene-level copy-number variation and uses a neural network to output individualized 10-year survival probability curves, reaching a c-index as high as 0.80 ± 0.07 in independent clinical trial cohorts, above current clinical indicators and subgroup-only models. An extension of the same framework flags about 60% of average-risk medulloblastoma patients who showed no significant survival disadvantage (P = 0.683) under lower-dose craniospinal irradiation.

Treating childhood medulloblastoma and ependymoma means trading recurrence risk against the long-term side effects of radiotherapy, yet current clinical indicators and molecular subgroups only give group-level risk. MANTIS takes genome-wide DNA methylation array data plus gene-level copy-number variation and uses a neural network to output individualized 10-year survival probability curves, reaching a c-index as high as 0.80 ± 0.07 in independent clinical trial cohorts, above current clinical indicators and subgroup-only models. An extension of the same framework flags about 60% of average-risk medulloblastoma patients who showed no significant survival disadvantage (P = 0.683) under lower-dose craniospinal irradiation.

Sber Robotics Center released GreenWAM, a robot foundation model rebuilt from Cosmos 3 Edge to serve as both a world model and a policy. A shared 79-slot anatomical action space maps the native joints of 20 robot families onto common leg, torso, arm and finger coordinates, and one model is trained jointly in seven visual and action modes. On WorldArena 1.0 it has the highest trajectory, depth and 3D accuracy among the compared world-action models and video generators; on the real Green humanoid it completes 95.6% of in-distribution instructions, against 88.9% for a policy adapted from Cosmos 3 Edge and 84.4% for π0.5.

Sber Robotics Center released GreenWAM, a robot foundation model rebuilt from Cosmos 3 Edge to serve as both a world model and a policy. A shared 79-slot anatomical action space maps the native joints of 20 robot families onto common leg, torso, arm and finger coordinates, and one model is trained jointly in seven visual and action modes. On WorldArena 1.0 it has the highest trajectory, depth and 3D accuracy among the compared world-action models and video generators; on the real Green humanoid it completes 95.6% of in-distribution instructions, against 88.9% for a policy adapted from Cosmos 3 Edge and 84.4% for π0.5.

Detecting many-body entanglement no longer requires reconstructing the full density matrix or relying on a predefined witness. EntNet turns two- to four-body Pauli correlations on a one-dimensional spin chain into a representation invariant under local basis rotations, sparsely selects 16 feature channels, and learns their spatial arrangement with a Transformer using rotary positional embeddings. On a finite-temperature XY benchmark it reaches 100% AUROC from expectation values of these low-order observables alone and holds the false-positive rate to 0.003 on 2,000 constructively separable states; the same model transfers without retraining to XXZ, frustrated J1-J2 and random local Hamiltonians and to longer chains.

Detecting many-body entanglement no longer requires reconstructing the full density matrix or relying on a predefined witness. EntNet turns two- to four-body Pauli correlations on a one-dimensional spin chain into a representation invariant under local basis rotations, sparsely selects 16 feature channels, and learns their spatial arrangement with a Transformer using rotary positional embeddings. On a finite-temperature XY benchmark it reaches 100% AUROC from expectation values of these low-order observables alone and holds the false-positive rate to 0.003 on 2,000 constructively separable states; the same model transfers without retraining to XXZ, frustrated J1-J2 and random local Hamiltonians and to longer chains.

HybridAgent lets a single agent move back and forth between continuous latent reasoning and text interaction within one generation, so part of its thinking leaves token-by-token decoding while tool use stays intact. Text agents previously had to write every step as tokens, while latent-reasoning models lacked native tool access and were hard to inspect. Across 11 benchmarks in mathematics, science, medicine, search and agentic reasoning at 4B to 14B scale, it reports a 10.4% average gain over frontier text and latent agents, with 35.5%–54.7% fewer tokens and 1.8–2.6 times faster inference.

HybridAgent lets a single agent move back and forth between continuous latent reasoning and text interaction within one generation, so part of its thinking leaves token-by-token decoding while tool use stays intact. Text agents previously had to write every step as tokens, while latent-reasoning models lacked native tool access and were hard to inspect. Across 11 benchmarks in mathematics, science, medicine, search and agentic reasoning at 4B to 14B scale, it reports a 10.4% average gain over frontier text and latent agents, with 35.5%–54.7% fewer tokens and 1.8–2.6 times faster inference.

Long-context decoding is bottlenecked by repeatedly reading a growing KV cache, and sparse attention reads only part of it, yet token-level selection has not reliably delivered both accuracy and lower latency. JSSA scores candidate tokens in a shared low-dimensional query–key subspace and pairs it with a gather-free indexed-attention kernel that keeps a distinct selected token set for each KV head. Across GQA, hybrid sliding-window/global attention and MLA models it stays closest to full-KV accuracy, and on a B200 at 128K context its decoding throughput reaches up to that of full-KV attention.

Long-context decoding is bottlenecked by repeatedly reading a growing KV cache, and sparse attention reads only part of it, yet token-level selection has not reliably delivered both accuracy and lower latency. JSSA scores candidate tokens in a shared low-dimensional query–key subspace and pairs it with a gather-free indexed-attention kernel that keeps a distinct selected token set for each KV head. Across GQA, hybrid sliding-window/global attention and MLA models it stays closest to full-KV accuracy, and on a B200 at 128K context its decoding throughput reaches up to that of full-KV attention.

Subcellular-resolution spatial transcriptomics can locate transcripts inside cells, but earlier methods targeted single analysis tasks and lacked a unified framework for characterizing and predicting gene localization in a cell-specific, context-aware way. SVC represents each gene's within-cell spatial expression as an image patch and uses other genes in the same cell as context, learning cross-gene spatial dependencies through self-supervised masking. On seqFISH+ mouse embryonic fibroblast data, SVC beat four image-based baselines on all six metrics for predicting subcellular patterns of unmeasured genes, with median gains of 5.2%–21.0% over the more challenging baseline pair.

Subcellular-resolution spatial transcriptomics can locate transcripts inside cells, but earlier methods targeted single analysis tasks and lacked a unified framework for characterizing and predicting gene localization in a cell-specific, context-aware way. SVC represents each gene's within-cell spatial expression as an image patch and uses other genes in the same cell as context, learning cross-gene spatial dependencies through self-supervised masking. On seqFISH+ mouse embryonic fibroblast data, SVC beat four image-based baselines on all six metrics for predicting subcellular patterns of unmeasured genes, with median gains of 5.2%–21.0% over the more challenging baseline pair.

Instead of rewriting whole proofs, AI agents work inside human-built Lean blueprints and tighten intermediate estimates one at a time, so each local improvement propagates along the dependency graph to the final bound. On Grothendieck's constant this pipeline gives the first explicit numerical improvement over Krivine's 1977 upper bound; on Korenblum's constant it improves Wang's 2025 lower bound. With the blueprint removed, the agent wrote plenty of Lean but improved nothing.

Instead of rewriting whole proofs, AI agents work inside human-built Lean blueprints and tighten intermediate estimates one at a time, so each local improvement propagates along the dependency graph to the final bound. On Grothendieck's constant this pipeline gives the first explicit numerical improvement over Krivine's 1977 upper bound; on Korenblum's constant it improves Wang's 2025 lower bound. With the blueprint removed, the agent wrote plenty of Lean but improved nothing.

The Martian atmosphere has been described by reanalyses that assimilate spacecraft retrievals into general circulation models, so where observations are sparse their fields largely reflect the model. MarsNet trains a two-hour forecast and a per-window observation update jointly inside a fixed physical structure, using no reanalysis or model field as input or target. On MGS radio-occultation temperatures assimilated by no dataset, it matches or improves on the three reanalyses in the lowest three scale heights, with a near-surface RMSE of 4.2 K against 4.4–5.4 K, but is less accurate aloft.

The Martian atmosphere has been described by reanalyses that assimilate spacecraft retrievals into general circulation models, so where observations are sparse their fields largely reflect the model. MarsNet trains a two-hour forecast and a per-window observation update jointly inside a fixed physical structure, using no reanalysis or model field as input or target. On MGS radio-occultation temperatures assimilated by no dataset, it matches or improves on the three reanalyses in the lowest three scale heights, with a near-surface RMSE of 4.2 K against 4.4–5.4 K, but is less accurate aloft.

As Earth-system AI models approach kilometre scales, halving grid spacing quadruples the grid-point count and the end-to-end training cost. One2N splits resolution refinement into successive increments: at each level the complete parent predictor is frozen and only a cross-resolution residual correction module is trained, so the parent's activations and optimizer states are never retained. One refinement level on the NeuralGCM-0.7° baseline yields One2N-0.07°, trained for about four weeks on a single 80 GB A800 GPU; against HRES-0.07° analyses it keeps the parent's tropospheric skill while reducing upper-stratospheric errors and resolving finer structure.

As Earth-system AI models approach kilometre scales, halving grid spacing quadruples the grid-point count and the end-to-end training cost. One2N splits resolution refinement into successive increments: at each level the complete parent predictor is frozen and only a cross-resolution residual correction module is trained, so the parent's activations and optimizer states are never retained. One refinement level on the NeuralGCM-0.7° baseline yields One2N-0.07°, trained for about four weeks on a single 80 GB A800 GPU; against HRES-0.07° analyses it keeps the parent's tropospheric skill while reducing upper-stratospheric errors and resolving finer structure.

e-means replaces the k-means centroid update with a mini-batch exponential moving average and uses fitted scaling laws to set batch size and decay directly from the problem configuration, so no hyperparameter sweep is needed. Across embedding datasets it reaches the loss Lloyd k-means attains at 25 epochs in 1.4 to 2.6 epochs, translating to 9.5 to 45.5 times wall-clock speedups over Flash-KMeans while converging to a lower loss. On DINO10B, the exact-assignment version partitions 10 billion vectors into 16 million clusters in 2.5 epochs at a loss below Lloyd at 25 epochs.

e-means replaces the k-means centroid update with a mini-batch exponential moving average and uses fitted scaling laws to set batch size and decay directly from the problem configuration, so no hyperparameter sweep is needed. Across embedding datasets it reaches the loss Lloyd k-means attains at 25 epochs in 1.4 to 2.6 epochs, translating to 9.5 to 45.5 times wall-clock speedups over Flash-KMeans while converging to a lower loss. On DINO10B, the exact-assignment version partitions 10 billion vectors into 16 million clusters in 2.5 epochs at a loss below Lloyd at 25 epochs.

Reasoning models often keep re-checking after they have reached an answer, stretching the chain of thought. DoubtLess has a small model that shares the large model's tokenizer read the same chain twice, once told to go straight to the answer and once told to double-check every step, and adds the scaled difference of the two logit vectors to the large model's logits at every step, acting only at decoding time with no weight changes. On seven competition-mathematics, graduate-science and programming benchmarks, Qwen3-32B steered by Qwen3-0.6B uses 31% fewer thinking tokens at unchanged accuracy, a larger cut than five inference-time baselines.

Reasoning models often keep re-checking after they have reached an answer, stretching the chain of thought. DoubtLess has a small model that shares the large model's tokenizer read the same chain twice, once told to go straight to the answer and once told to double-check every step, and adds the scaled difference of the two logit vectors to the large model's logits at every step, acting only at decoding time with no weight changes. On seven competition-mathematics, graduate-science and programming benchmarks, Qwen3-32B steered by Qwen3-0.6B uses 31% fewer thinking tokens at unchanged accuracy, a larger cut than five inference-time baselines.

Agent evaluation has relied on fixed benchmarks that report task success or failure but cannot say whether reviewing or steering an agent mid-trajectory is worth it at deployment. Researchers bring micro-randomized trials into agent trajectories, assigning review or direct execution with equal probability at each eligible decision, then estimate causal effects with augmented inverse probability weighting and structural nested mean models. On τ-bench Retail, ToolSandbox and ImpossibleBench, review effects vary substantially across tool-call types and agent states, and the same data yield decision-group risk rankings.

Agent evaluation has relied on fixed benchmarks that report task success or failure but cannot say whether reviewing or steering an agent mid-trajectory is worth it at deployment. Researchers bring micro-randomized trials into agent trajectories, assigning review or direct execution with equal probability at each eligible decision, then estimate causal effects with augmented inverse probability weighting and structural nested mean models. On τ-bench Retail, ToolSandbox and ImpossibleBench, review effects vary substantially across tool-call types and agent states, and the same data yield decision-group risk rankings.

A theory treats spatial memory as a universal computer: a learner that reads only coincidences, whether two quantities are equal, rather than coordinates such as position or count that keep taking new values, has its rule fixed by finite experience at every length. The resulting machine pairs coprime-period rings, called clocks, with a controller that never reads them; trained on strings of up to 40 symbols, it holds its rule to length 500 on nine of fifteen benchmark tasks, five of which no LSTM or Transformer solves, and after training with new periods it reverses strings of 15,015 symbols without error before failing on the predicted curve.

A theory treats spatial memory as a universal computer: a learner that reads only coincidences, whether two quantities are equal, rather than coordinates such as position or count that keep taking new values, has its rule fixed by finite experience at every length. The resulting machine pairs coprime-period rings, called clocks, with a controller that never reads them; trained on strings of up to 40 symbols, it holds its rule to length 500 on nine of fifteen benchmark tasks, five of which no LSTM or Transformer solves, and after training with new periods it reverses strings of 15,015 symbols without error before failing on the predicted curve.

In large environments at naturalistic scale, rat and bat hippocampal CA1 place cells often carry multiple irregularly spaced firing fields, whereas the classical single-scale successor representation holds only in small enclosures. The factorised hierarchical successor representation (fHSR) folds one-step actions together with subgoal-directed subroutines into a predictive map, then decomposes the maps of several tasks into a non-negative overcomplete basis that gives each model cell its tuning. On a 48 m rat track and a 200 m bat track the model reproduced the gamma-Poisson overdispersion of field counts, and it predicted a spatial periodicity of population activity whose wavelength is independent of track length, a prediction tested in rat data.

In large environments at naturalistic scale, rat and bat hippocampal CA1 place cells often carry multiple irregularly spaced firing fields, whereas the classical single-scale successor representation holds only in small enclosures. The factorised hierarchical successor representation (fHSR) folds one-step actions together with subgoal-directed subroutines into a predictive map, then decomposes the maps of several tasks into a non-negative overcomplete basis that gives each model cell its tuning. On a 48 m rat track and a 200 m bat track the model reproduced the gamma-Poisson overdispersion of field counts, and it predicted a spatial periodicity of population activity whose wavelength is independent of track length, a prediction tested in rat data.

For a class of equivariant graph-neural-network interatomic potentials, a universal approximation theorem now covers atomic energies: provided every atom's neighbors are reached within a common number of layers and no more than a given number of neighbors sit at a common distance inside the layer cutoff, a multilayer network whose layer cutoff is shorter than the physical range still approximates the atomic energy over that range, with spherical harmonic filter degrees no larger than that neighbor bound. The production DPA4 architecture inherits the result by simulating the constructed layers and realizing the bilinear readout.

For a class of equivariant graph-neural-network interatomic potentials, a universal approximation theorem now covers atomic energies: provided every atom's neighbors are reached within a common number of layers and no more than a given number of neighbors sit at a common distance inside the layer cutoff, a multilayer network whose layer cutoff is shorter than the physical range still approximates the atomic energy over that range, with spherical harmonic filter degrees no larger than that neighbor bound. The production DPA4 architecture inherits the result by simulating the constructed layers and realizing the bilinear readout.

Methods that shape how an opponent learns in general-sum games have needed access to that opponent's policy, parameters or learning rule. A study carries the single-agent practice of meta-learning an update rule into multi-agent play, meta-training a rule that reads only the agent's own experience and payoff matrix, so deployment needs no privileged opponent information. Across round-robin tournaments in Prisoner's Dilemma, Matching Pennies, Chicken, Stag Hunt and Coin Game, the rule ranks first in normalised mean score over all five games and first in self-play.

Methods that shape how an opponent learns in general-sum games have needed access to that opponent's policy, parameters or learning rule. A study carries the single-agent practice of meta-learning an update rule into multi-agent play, meta-training a rule that reads only the agent's own experience and payoff matrix, so deployment needs no privileged opponent information. Across round-robin tournaments in Prisoner's Dilemma, Matching Pennies, Chicken, Stag Hunt and Coin Game, the rule ranks first in normalised mean score over all five games and first in self-play.

Low-bit KV caches cut memory traffic, yet existing kernels still convert every integer back to floating point and apply its scale and offset, so instruction issue becomes the new bottleneck. FOSCAR folds each quantization group's shared scale and offset into the other matrix operand or the partial product, and builds Tensor Core operands directly from packed integers with bit manipulation. Its INT2 kernels reach geometric-mean speedups over FlashInfer BF16 on H100 and B200, and the kernel library is open-sourced.

Low-bit KV caches cut memory traffic, yet existing kernels still convert every integer back to floating point and apply its scale and offset, so instruction issue becomes the new bottleneck. FOSCAR folds each quantization group's shared scale and offset into the other matrix operand or the partial product, and builds Tensor Core operands directly from packed integers with bit manipulation. Its INT2 kernels reach geometric-mean speedups over FlashInfer BF16 on H100 and B200, and the kernel library is open-sourced.

WebCoder-27B, trained during an internship at Zhipu AI, keeps reproduction, generation, modification and repair on one evolving project state and scores reinforcement learning by how well the executable application meets task requirements. On the new Webdev-Bench its requirement satisfaction rises from 64.8 for the Qwen3.8-27B base model to 72.8, above Claude-Opus-4.8 at 65.1; on Vision2Web its average visual and functional score rises from 41.1 to 62.4, on par with Claude-Opus-4.8 at 62.5.

WebCoder-27B, trained during an internship at Zhipu AI, keeps reproduction, generation, modification and repair on one evolving project state and scores reinforcement learning by how well the executable application meets task requirements. On the new Webdev-Bench its requirement satisfaction rises from 64.8 for the Qwen3.8-27B base model to 72.8, above Claude-Opus-4.8 at 65.1; on Vision2Web its average visual and functional score rises from 41.1 to 62.4, on par with Claude-Opus-4.8 at 62.5.

In sparse-to-dense mesh motion completion, residual priors trained on one geometry family often fail to beat parameter-free harmonic extension, because residual amplitude varies by more than an order of magnitude across families and densities while the target motion is unknown at test time. Cepheid learns only a dimensionless residual structure and recovers physical amplitude from observables plus a lightweight correction fitted on training sources. The unified model stays below harmonic extension on all five evaluated families, cutting error by 15.6% on an unseen cloth sheet and 10.1% on procedural legged bodies, without skeletons, registration, or cross-mesh correspondence.

In sparse-to-dense mesh motion completion, residual priors trained on one geometry family often fail to beat parameter-free harmonic extension, because residual amplitude varies by more than an order of magnitude across families and densities while the target motion is unknown at test time. Cepheid learns only a dimensionless residual structure and recovers physical amplitude from observables plus a lightweight correction fitted on training sources. The unified model stays below harmonic extension on all five evaluated families, cutting error by 15.6% on an unseen cloth sheet and 10.1% on procedural legged bodies, without skeletons, registration, or cross-mesh correspondence.

Humanoid policies learned in simulation can transfer motions to hardware while failing to transfer the force interactions learned in training, which costs torque efficiency, force-tracking accuracy and impact absorption. Contact-UAN keeps the simulator's rigid-body and contact physics and learns only a residual correction on the actuator command channel, using onboard proprioception to infer a stance-foot pose so replay starts with hardware-consistent contact forces. On a Unitree G1, a single residual model trained on 6.5 minutes of real walking data yields 20% lower running torque and 43% lower force-tracking error, and supports stable jumping and landing.

Humanoid policies learned in simulation can transfer motions to hardware while failing to transfer the force interactions learned in training, which costs torque efficiency, force-tracking accuracy and impact absorption. Contact-UAN keeps the simulator's rigid-body and contact physics and learns only a residual correction on the actuator command channel, using onboard proprioception to infer a stance-foot pose so replay starts with hardware-consistent contact forces. On a Unitree G1, a single residual model trained on 6.5 minutes of real walking data yields 20% lower running torque and 43% lower force-tracking error, and supports stable jumping and landing.

StrataFlux, from Cornell and Carnegie Mellon researchers, keeps low-bit weights and KV cache resident on the GPU and fetches only a small set of refinement bits from host RAM or SSD for the computation at hand, recovering accuracy while retaining quantization's memory and speed advantages. On an RTX 5090, 2-bit Llama-3.1-8B-Instruct weights drop WikiText-2 perplexity from 9.46 to 8.40 while keeping 95.4% of the quantized base's throughput; with 2-bit weights and KV, Qwen3-8B's AIME 2024/2025 accuracy rises from 11.0% to 47.1% at unchanged GPU memory use.

StrataFlux, from Cornell and Carnegie Mellon researchers, keeps low-bit weights and KV cache resident on the GPU and fetches only a small set of refinement bits from host RAM or SSD for the computation at hand, recovering accuracy while retaining quantization's memory and speed advantages. On an RTX 5090, 2-bit Llama-3.1-8B-Instruct weights drop WikiText-2 perplexity from 9.46 to 8.40 while keeping 95.4% of the quantized base's throughput; with 2-bit weights and KV, Qwen3-8B's AIME 2024/2025 accuracy rises from 11.0% to 47.1% at unchanged GPU memory use.

Long-horizon video generation has traded between two kinds of memory: explicit geometric caches recall a place by location but grow with the scene, while recurrent memories stay constant-size yet let old observations fade with time. Smart turns a fixed-capacity slot memory into one addressed by camera pose, so writes and reads depend on where something was seen rather than how long ago, and a diffusion generator uses the memory's confidence to decide where to trust it. On long trajectories in Memory Maze and Minecraft it keeps high-fidelity recall far beyond its training horizon at constant memory and per-step cost.

Long-horizon video generation has traded between two kinds of memory: explicit geometric caches recall a place by location but grow with the scene, while recurrent memories stay constant-size yet let old observations fade with time. Smart turns a fixed-capacity slot memory into one addressed by camera pose, so writes and reads depend on where something was seen rather than how long ago, and a diffusion generator uses the memory's confidence to decide where to trust it. On long trajectories in Memory Maze and Minecraft it keeps high-fidelity recall far beyond its training horizon at constant memory and per-step cost.

Reinforcement learning from unit tests raises pass rates but also teaches code models to edit more and to invent bugs in programs that are already correct. PreciseCoder adds a signal for semantically verified edit precision and recall against a reference repair; on 4B, 9B and 27B models it cuts the average edit from 7–15 lines to 3 at a pass rate on par with unit-test-only training, and raises the no-edit rate on correct code to 47–52%.

Reinforcement learning from unit tests raises pass rates but also teaches code models to edit more and to invent bugs in programs that are already correct. PreciseCoder adds a signal for semantically verified edit precision and recall against a reference repair; on 4B, 9B and 27B models it cuts the average edit from 7–15 lines to 3 at a pass rate on par with unit-test-only training, and raises the no-edit rate on correct code to 47–52%.

Reinforcement-learning training of reasoning language models chains generation, assessment and training across several large models each step, and existing systems usually optimize only part of that design space. RPEG places pipeline structure, including bounded asynchrony and actor–critic sharing, and execution choices such as stage fusion, fine-grained scheduling and multi-dimensional parallelism into one graph searched jointly, backed by a hardware-calibrated analytical performance model. On up to 256 GH200 GPUs its runtime is up to 2x faster end to end than ReaL, cuts step latency by up to 2.48x versus verl, and raises hardware utilization by up to 30%.

Reinforcement-learning training of reasoning language models chains generation, assessment and training across several large models each step, and existing systems usually optimize only part of that design space. RPEG places pipeline structure, including bounded asynchrony and actor–critic sharing, and execution choices such as stage fusion, fine-grained scheduling and multi-dimensional parallelism into one graph searched jointly, backed by a hardware-calibrated analytical performance model. On up to 256 GH200 GPUs its runtime is up to 2x faster end to end than ReaL, cuts step latency by up to 2.48x versus verl, and raises hardware utilization by up to 30%.

Transformers have long passed information across depth by adding layer outputs in a residual stream, and recent alternatives mHC and AttnRes make that depth aggregation more expressive. Depth Memory recasts cross-layer flow as streams that store and retrieve layer representations, and from it comes Linear Depth Memory (LDM), which treats retrieval across depth as a regression solved exactly by recursive least squares while carrying a residual stream, a matrix memory stream and an inverse covariance stream. Pretrained on 400M Llama-like and 1.5B Qwen-like backbones with 50B and 100B FineWeb-Edu tokens, LDM reaches lower validation, WikiText-2 and LAMBADA perplexity than tuned PreLN, mHC and AttnRes, with competitive zero-shot accuracy.

Transformers have long passed information across depth by adding layer outputs in a residual stream, and recent alternatives mHC and AttnRes make that depth aggregation more expressive. Depth Memory recasts cross-layer flow as streams that store and retrieve layer representations, and from it comes Linear Depth Memory (LDM), which treats retrieval across depth as a regression solved exactly by recursive least squares while carrying a residual stream, a matrix memory stream and an inverse covariance stream. Pretrained on 400M Llama-like and 1.5B Qwen-like backbones with 50B and 100B FineWeb-Edu tokens, LDM reaches lower validation, WikiText-2 and LAMBADA perplexity than tuned PreLN, mHC and AttnRes, with competitive zero-shot accuracy.

Single-cell annotation has long treated cell identity (lineage) and state (active transcriptional programs) with separate tools, while fixed-output classifiers can only return predefined types. CellScribe feeds frozen single-cell foundation-model embeddings into a pretrained language model through AdaLN adapters, so one shared model generates either an ontology-grounded cell type or a meta-program profile with activity scores under task-specific prompts. On the 309-type, 4.44M-cell study-holdout split of CELLxGENE, the full pipeline reaches 0.223 macro-F1 and 0.484 weighted-F1, above the 0.194 and 0.366 of a supervised probe on frozen scFoundation embeddings; on malignant cells from held-out 3CA studies it reaches 0.344 hallmark accuracy against the probe's 0.227.

Single-cell annotation has long treated cell identity (lineage) and state (active transcriptional programs) with separate tools, while fixed-output classifiers can only return predefined types. CellScribe feeds frozen single-cell foundation-model embeddings into a pretrained language model through AdaLN adapters, so one shared model generates either an ontology-grounded cell type or a meta-program profile with activity scores under task-specific prompts. On the 309-type, 4.44M-cell study-holdout split of CELLxGENE, the full pipeline reaches 0.223 macro-F1 and 0.484 weighted-F1, above the 0.194 and 0.366 of a supervised probe on frozen scFoundation embeddings; on malignant cells from held-out 3CA studies it reaches 0.344 hallmark accuracy against the probe's 0.227.

Intravascular optical coherence tomography (OCT) images entangle tissue attenuation with the imaging system's own physics, so attenuation has been estimated with simplified closed-form inversions or with learned inverses tied to a specific scanner. DiffOCT writes the Beer-Lambert attenuation model as a differentiable forward model and runs gradient descent on a single B-scan to recover attenuation per pixel, with no training data or segmentation. Across 5,120 inversions on five lipid phantoms, recovered values rose monotonically with scatterer concentration and matched curve-fitting slopes; the same procedure ran unmodified on 29,269 clinical pullback frames from 128 patients, though the clinical data show geometric transfer only, not calibrated absolute values.

Intravascular optical coherence tomography (OCT) images entangle tissue attenuation with the imaging system's own physics, so attenuation has been estimated with simplified closed-form inversions or with learned inverses tied to a specific scanner. DiffOCT writes the Beer-Lambert attenuation model as a differentiable forward model and runs gradient descent on a single B-scan to recover attenuation per pixel, with no training data or segmentation. Across 5,120 inversions on five lipid phantoms, recovered values rose monotonically with scatterer concentration and matched curve-fitting slopes; the same procedure ran unmodified on 29,269 clinical pullback frames from 128 patients, though the clinical data show geometric transfer only, not calibrated absolute values.

Whether research agents can keep improving from their own research experience depends on turning a hypothesis into an investigation that can actually run and on knowing which contribution to reinforce. DiscoveryRSI builds research trees of hypotheses through multi-agent evolutionary search, keeps the evidence and decisions behind every proposal and revision, and uses those records to train the model and refine the workflow. After two rounds of self-improvement, mean hypothesis quality rises from 4.86 to 5.39, with gains concentrated in testability: falsifiability from 4.00 to 6.20 and data resolvability from 3.73 to 4.99.

Whether research agents can keep improving from their own research experience depends on turning a hypothesis into an investigation that can actually run and on knowing which contribution to reinforce. DiscoveryRSI builds research trees of hypotheses through multi-agent evolutionary search, keeps the evidence and decisions behind every proposal and revision, and uses those records to train the model and refine the workflow. After two rounds of self-improvement, mean hypothesis quality rises from 4.86 to 5.39, with gains concentrated in testability: falsifiability from 4.00 to 6.20 and data resolvability from 3.73 to 4.99.

Mixture-of-experts models must keep every expert in memory at inference even though each token uses only a few, so deleting experts by independent importance scores is a common fix. HOPE writes the pruning error as pairwise interaction terms between experts and selects a whole prune-set with a quadratic program, reaching the best average rank across three models, two calibration datasets and several benchmarks. On Qwen3.5-35B-A3B at 50% pruning it scores 42.3 on SWE-Bench Pro, matching the unpruned model's 41.8 while the first-order method REAP falls to 36.2.

Mixture-of-experts models must keep every expert in memory at inference even though each token uses only a few, so deleting experts by independent importance scores is a common fix. HOPE writes the pruning error as pairwise interaction terms between experts and selects a whole prune-set with a quadratic program, reaching the best average rank across three models, two calibration datasets and several benchmarks. On Qwen3.5-35B-A3B at 50% pruning it scores 42.3 on SWE-Bench Pro, matching the unpruned model's 41.8 while the first-order method REAP falls to 36.2.

Diffusion models for 3D molecules learn atomic distributions implicitly from data, so samples can be statistically plausible yet chemically invalid. SensDiff turns four physicochemical rules, steric non-overlap, bond length, bond angle and valence, into differentiable energies that are activated in a coarse-to-fine order with adaptive weighting during training, so the rules enter the generative dynamics rather than only correcting sampling. On the QM9 and GEOM-Drugs benchmarks it reaches 96.9% and 97.2% validity, which the authors report as better than the compared baselines.

Diffusion models for 3D molecules learn atomic distributions implicitly from data, so samples can be statistically plausible yet chemically invalid. SensDiff turns four physicochemical rules, steric non-overlap, bond length, bond angle and valence, into differentiable energies that are activated in a coarse-to-fine order with adaptive weighting during training, so the rules enter the generative dynamics rather than only correcting sampling. On the QM9 and GEOM-Drugs benchmarks it reaches 96.9% and 97.2% validity, which the authors report as better than the compared baselines.

State-inferer recovers a protocol implementation's state machine by program slicing: from each state-variable assignment it runs depth-bounded forward and backward searches on the system dependence graph, hands the LLM one compact fragment per transition, then specializes RFC security properties into linear temporal logic formulas over that model and checks them with nuXmv. On seven open-source TLS, IKEv2 and SSH implementations it reaches a macro-average strict transition-level F-score of 95.2 against 78.9 for the strongest baseline, Codex, and uncovers four previously unknown vulnerabilities, all assigned CVE or CNVD identifiers.

State-inferer recovers a protocol implementation's state machine by program slicing: from each state-variable assignment it runs depth-bounded forward and backward searches on the system dependence graph, hands the LLM one compact fragment per transition, then specializes RFC security properties into linear temporal logic formulas over that model and checks them with nuXmv. On seven open-source TLS, IKEv2 and SSH implementations it reaches a macro-average strict transition-level F-score of 95.2 against 78.9 for the strongest baseline, Codex, and uncovers four previously unknown vulnerabilities, all assigned CVE or CNVD identifiers.

PepPr uses a peptide language model trained on a reference proteome as a prior, fused with spectral evidence at every step of autoregressive decoding rather than applied as post hoc correction. On the nine-species dataset with Casanovo as the base model, median peptide recall rose by 15.8 percentage points, improving all nine species; on single-amino-acid variants and antibodies, variant identification was not suppressed, and recall of peptides spanning the variable-constant boundary rose from 0.23 to 0.68.

PepPr uses a peptide language model trained on a reference proteome as a prior, fused with spectral evidence at every step of autoregressive decoding rather than applied as post hoc correction. On the nine-species dataset with Casanovo as the base model, median peptide recall rose by 15.8 percentage points, improving all nine species; on single-amino-acid variants and antibodies, variant identification was not suppressed, and recall of peptides spanning the variable-constant boundary rose from 0.23 to 0.68.

Reinforcement-learning control policies for a pneumatically actuated soft arm can be trained entirely on an offline reduced model instead of through repeated interaction with fragile hardware. Soft robots have lacked a reliable, fast simulator, and trial and error on the hardware itself is costly. The arm's dynamics are reduced to a six-dimensional adiabatic spectral submanifold (aSSM), and a polynomial ordinary differential equation with only 6,924 parameters serves as the training environment; a tracking policy trains in three hours and deploys without fine-tuning, with 40% lower tracking error than otherwise identical policies trained on a 4.8-million-parameter RNN or a 1.7-million-parameter transformer.

Reinforcement-learning control policies for a pneumatically actuated soft arm can be trained entirely on an offline reduced model instead of through repeated interaction with fragile hardware. Soft robots have lacked a reliable, fast simulator, and trial and error on the hardware itself is costly. The arm's dynamics are reduced to a six-dimensional adiabatic spectral submanifold (aSSM), and a polynomial ordinary differential equation with only 6,924 parameters serves as the training environment; a tracking policy trains in three hours and deploys without fine-tuning, with 40% lower tracking error than otherwise identical policies trained on a 4.8-million-parameter RNN or a 1.7-million-parameter transformer.

Causal discovery no longer has to rest on conditional-independence tests or a single best-scoring model. causalBMS scores each candidate parent by how often it appears in expressions sampled from a Bayesian symbolic-regression posterior, so it still gives a direction in bivariate, strongly autocorrelated and high-noise settings. On the CauseMe benchmark it ranked first in 13 of 21 dataset-metric pairs and second in four more, while also returning closed-form functional forms.

Causal discovery no longer has to rest on conditional-independence tests or a single best-scoring model. causalBMS scores each candidate parent by how often it appears in expressions sampled from a Bayesian symbolic-regression posterior, so it still gives a direction in bivariate, strongly autocorrelated and high-noise settings. On the CauseMe benchmark it ranked first in 13 of 21 dataset-metric pairs and second in four more, while also returning closed-form functional forms.

A pathology case often spans dozens of slides from different specimens, yet existing pathology agents analyze one slide at a time and describe morphology rather than measure it. PlanPath turns a diagnostic question into an exploration plan, routes to the slides likely to carry the evidence, then zooms from tissue architecture down to cells for counts and proportions, aggregating cross-slide evidence into a traceable answer. On patient-level renal and gastric cancer datasets, PlanPath reaches total accuracy of 72.71% and 60.24%, 12.14 and 9.43 percentage points above the strongest baseline.

A pathology case often spans dozens of slides from different specimens, yet existing pathology agents analyze one slide at a time and describe morphology rather than measure it. PlanPath turns a diagnostic question into an exploration plan, routes to the slides likely to carry the evidence, then zooms from tissue architecture down to cells for counts and proportions, aggregating cross-slide evidence into a traceable answer. On patient-level renal and gastric cancer datasets, PlanPath reaches total accuracy of 72.71% and 60.24%, 12.14 and 9.43 percentage points above the strongest baseline.

Ground-state calculations for the two-dimensional Fermi–Hubbard model have been stuck on a basis mismatch: momentum-space matrix product states are compact, while the diagonal correlators that capture short-range correlation act naturally in real space. The new scheme has a quantum processor prepare and rotate a bond-dimension-4 momentum-space matrix product state, sample in bases fixed by the Hamiltonian, and optimize the correlator entirely in classical post-processing with no quantum-classical feedback loop. On 32 qubits of Quantinuum's H2-2, postselection gives a 4.40% relative ground-state energy error for the half-filled lattice, reduced further by calibration using low-order moments from the matrix product state.

Ground-state calculations for the two-dimensional Fermi–Hubbard model have been stuck on a basis mismatch: momentum-space matrix product states are compact, while the diagonal correlators that capture short-range correlation act naturally in real space. The new scheme has a quantum processor prepare and rotate a bond-dimension-4 momentum-space matrix product state, sample in bases fixed by the Hamiltonian, and optimize the correlator entirely in classical post-processing with no quantum-classical feedback loop. On 32 qubits of Quantinuum's H2-2, postselection gives a 4.40% relative ground-state energy error for the half-filled lattice, reduced further by calibration using low-order moments from the matrix product state.

Causal structure learning built on differentiable acyclicity constraints has lacked finite-sample theory for nonlinear models. This work sets each node's conditional distribution to a generalized linear model, searches all topological orders through one continuous objective, and proves that global minimizers select the sparsest representation at the population level and attain an estimation error rate up to log factors at the sample level. Experiments examine structure recovery accuracy and runtime on logistic, Poisson and Gaussian DAGs.

Causal structure learning built on differentiable acyclicity constraints has lacked finite-sample theory for nonlinear models. This work sets each node's conditional distribution to a generalized linear model, searches all topological orders through one continuous objective, and proves that global minimizers select the sparsest representation at the population level and attain an estimation error rate up to log factors at the sample level. Experiments examine structure recovery accuracy and runtime on logistic, Poisson and Gaussian DAGs.

Mouse behavior analysis has relied on self-supervised pose models that encode each animal independently, yet aggression, courtship, mounting and allogrooming are inherently multi-animal. MICE aligns six mouse-pose datasets to a universal 12-keypoint vocabulary and pretrains jointly, pairing a hierarchical individual encoder for per-mouse kinematics with a Perceiver-style social encoder whose learned latent codebook handles any group size, so one checkpoint serves different animal counts without retraining. Across six datasets and four evaluation protocols, MICE matches or exceeds prior keypoint-based baselines, and layer-wise probing shows the social stage improves discriminability beyond the individual encoder.

Mouse behavior analysis has relied on self-supervised pose models that encode each animal independently, yet aggression, courtship, mounting and allogrooming are inherently multi-animal. MICE aligns six mouse-pose datasets to a universal 12-keypoint vocabulary and pretrains jointly, pairing a hierarchical individual encoder for per-mouse kinematics with a Perceiver-style social encoder whose learned latent codebook handles any group size, so one checkpoint serves different animal counts without retraining. Across six datasets and four evaluation protocols, MICE matches or exceeds prior keypoint-based baselines, and layer-wise probing shows the social stage improves discriminability beyond the individual encoder.

The interactive segmentation model SEEM performs poorly on OCTA retinal vessels out of the box, reaching only 75.83% Dice on large vessels with a frozen encoder, and every candidate mask it can produce is read from the same quarter-resolution feature maps, too coarse for vessels one to three pixels wide. The Read-out Basis Adapter (RBA) adds 32 half-resolution basis maps at the read-out, which each candidate accesses through a zero-initialised projection, for 0.47M trainable parameters. On OCTA-500 3 mm large vessels, averaged over three seeds, it reaches 91.93% Dice and 85.16% IoU, 1.72 and 2.90 points above Conv-LoRA, with gains concentrated on thin vessels and edges and no significant difference on the FAZ.

The interactive segmentation model SEEM performs poorly on OCTA retinal vessels out of the box, reaching only 75.83% Dice on large vessels with a frozen encoder, and every candidate mask it can produce is read from the same quarter-resolution feature maps, too coarse for vessels one to three pixels wide. The Read-out Basis Adapter (RBA) adds 32 half-resolution basis maps at the read-out, which each candidate accesses through a zero-initialised projection, for 0.47M trainable parameters. On OCTA-500 3 mm large vessels, averaged over three seeds, it reaches 91.93% Dice and 85.16% IoU, 1.72 and 2.90 points above Conv-LoRA, with gains concentrated on thin vessels and edges and no significant difference on the FAZ.

When the original RTL and documentation are gone and only a gate-level netlist or a complete truth table remains, RTLREX has a large language model choose analyses among domain-specific tools, build candidate RTL, and confirm each result with independent equivalence checking. On the 30 benchmarks of ICCAD 2022 CAD Contest Problem A it recovered equivalent high-level RTL for all of them, above WolFEx's 26, and on the 26 recovered by both it cut mean syntax cost by 20.5% relative to WolFEx.

When the original RTL and documentation are gone and only a gate-level netlist or a complete truth table remains, RTLREX has a large language model choose analyses among domain-specific tools, build candidate RTL, and confirm each result with independent equivalence checking. On the 30 benchmarks of ICCAD 2022 CAD Contest Problem A it recovered equivalent high-level RTL for all of them, above WolFEx's 26, and on the 26 recovered by both it cut mean syntax cost by 20.5% relative to WolFEx.

SCOUT compresses high-dimensional, noisy baseline variables into at most two adjustment scores and learns a treatment-responsive composite outcome inside that low-dimensional space, yielding average and heterogeneous treatment-effect estimates from observational data. Among 481 STEP-BD participants with bipolar disorder, the learned composite put its largest weights on nocturia, tremor, polyuria and nausea/vomiting, with a mean effect of 0.250 and Holm-adjusted values of 0.0026 for the average effect and 0.033 for heterogeneity. On synthetic and semi-synthetic data its average-effect error was close to CV-TMLE while CATE error for its score-indexed target was substantially lower.

SCOUT compresses high-dimensional, noisy baseline variables into at most two adjustment scores and learns a treatment-responsive composite outcome inside that low-dimensional space, yielding average and heterogeneous treatment-effect estimates from observational data. Among 481 STEP-BD participants with bipolar disorder, the learned composite put its largest weights on nocturia, tremor, polyuria and nausea/vomiting, with a mean effect of 0.250 and Holm-adjusted values of 0.0026 for the average effect and 0.033 for heterogeneity. On synthetic and semi-synthetic data its average-effect error was close to CV-TMLE while CATE error for its score-indexed target was substantially lower.

MAAS replaces the text-embedding conditioning of a pretrained audio diffusion transformer with region-of-interest fMRI responses, removing the decoder that previously predicted an external stimulus representation such as CLAP or YamNet. On a 7T fMRI dataset it improves acoustic and semantic metrics, including CLAP similarity, pairwise identification accuracy and Mel-SSIM, over baselines such as two-stage fMRI-to-latent conditioning. The framework also uses a fidelity metric grounded in a computational model of auditory cortex to quantify how individual auditory regions contribute to the reconstructed audio.

MAAS replaces the text-embedding conditioning of a pretrained audio diffusion transformer with region-of-interest fMRI responses, removing the decoder that previously predicted an external stimulus representation such as CLAP or YamNet. On a 7T fMRI dataset it improves acoustic and semantic metrics, including CLAP similarity, pairwise identification accuracy and Mel-SSIM, over baselines such as two-stage fMRI-to-latent conditioning. The framework also uses a fidelity metric grounded in a computational model of auditory cortex to quantify how individual auditory regions contribute to the reconstructed audio.

Intracranial EEG electrode count and placement differ from patient to patient, so detectors with a fixed channel topology generalize poorly under subject-independent evaluation. MOSAIC encodes each channel with a shared convolutional encoder, then a single channel-attention layer infers a weighted inter-channel adjacency from the encoded features without any preset montage graph; an exponential moving average self-distillation term pulls each window's adjacency toward a slowly updated per-patient teacher. Across four public iEEG corpora, MOSAIC's F2 on both the SWEC-ETHZ short-term and long-term corpora exceeds the listed baselines, by 1.8 points over the strongest baseline on the short-term corpus and 3.3 points on the long-term one.

Intracranial EEG electrode count and placement differ from patient to patient, so detectors with a fixed channel topology generalize poorly under subject-independent evaluation. MOSAIC encodes each channel with a shared convolutional encoder, then a single channel-attention layer infers a weighted inter-channel adjacency from the encoded features without any preset montage graph; an exponential moving average self-distillation term pulls each window's adjacency toward a slowly updated per-patient teacher. Across four public iEEG corpora, MOSAIC's F2 on both the SWEC-ETHZ short-term and long-term corpora exceeds the listed baselines, by 1.8 points over the strongest baseline on the short-term corpus and 3.3 points on the long-term one.

In ultra-central heavy-ion collisions the measured ratio of elliptic to triangular flow approaches unity, whereas linear scaling from initial eccentricity predicts a clear hierarchy. UCC-DiffHIC, an AI-accelerated event-by-event hydrodynamic model, generates these extremely rare collisions and decomposes the flow ratio exactly into a linear-response term and the ratio of the Pearson response correlations for elliptic and triangular flow. For the standard TRENTo parameter set, the suppression from residual fluctuations offsets the linear-scaling enhancement, bringing the ratio near unity in the most central collisions in agreement with CMS; two Bayesian parameter sets compensate less completely.

In ultra-central heavy-ion collisions the measured ratio of elliptic to triangular flow approaches unity, whereas linear scaling from initial eccentricity predicts a clear hierarchy. UCC-DiffHIC, an AI-accelerated event-by-event hydrodynamic model, generates these extremely rare collisions and decomposes the flow ratio exactly into a linear-response term and the ratio of the Pearson response correlations for elliptic and triangular flow. For the standard TRENTo parameter set, the suppression from residual fluctuations offsets the linear-scaling enhancement, bringing the ratio near unity in the most central collisions in agreement with CMS; two Bayesian parameter sets compensate less completely.

Multi-step reaction mechanism prediction has largely stayed at the level of single elementary steps and generalizes poorly to unseen reaction spaces. GRMP couples a chemical mechanism knowledge graph, completed with expert templates, to a language model: a graph convolutional network generates candidate pathways that become preference data for direct preference optimization, and at inference a relation retrieval module pulls graph evidence to refine the pathway. GRMP reaches pathway accuracy of 48.1% on ChemMechQA and 57.9% on oMe-Gold, the highest among the compared methods.

Multi-step reaction mechanism prediction has largely stayed at the level of single elementary steps and generalizes poorly to unseen reaction spaces. GRMP couples a chemical mechanism knowledge graph, completed with expert templates, to a language model: a graph convolutional network generates candidate pathways that become preference data for direct preference optimization, and at inference a relation retrieval module pulls graph evidence to refine the pathway. GRMP reaches pathway accuracy of 48.1% on ChemMechQA and 57.9% on oMe-Gold, the highest among the compared methods.

Deciding whether a mixed two-qutrit state is steerable has relied on semidefinite programming over a finite measurement set, and that does not necessarily carry over to a different state family. Correlation-enhanced quantum reservoir processing (CEQRP) adds two-body spin-correlation readouts from a spin-1 reservoir to the feature vector, letting logistic regression reach high accuracy on unseen Werner and isotropic states. With training augmented by samples certified by two steering criteria, classifiers flagged partially entangled states whose steerability semidefinite programming then independently confirmed using seven projective measurement settings.

Deciding whether a mixed two-qutrit state is steerable has relied on semidefinite programming over a finite measurement set, and that does not necessarily carry over to a different state family. Correlation-enhanced quantum reservoir processing (CEQRP) adds two-body spin-correlation readouts from a spin-1 reservoir to the feature vector, letting logistic regression reach high accuracy on unseen Werner and isotropic states. With training augmented by samples certified by two steering criteria, classifiers flagged partially entangled states whose steerability semidefinite programming then independently confirmed using seven projective measurement settings.

DNA-LLM2Vec converts pretrained causal DNA decoders into frozen bidirectional encoders in stages: it opens bidirectional context, repairs the resulting mismatch with masked next-token prediction, then adds contrastive views encoding strand and positional invariance. Across two backbones, DNAGPT and HyenaDNA, and seven biological task families, the benefit is not uniform but depends on task structure and backbone; crop contrastive learning helps most for extended-context families such as enhancer–promoter interaction and Virus-COVID, reaching frozen Avg-EPI MCC of 27.5% and 26.7%.

DNA-LLM2Vec converts pretrained causal DNA decoders into frozen bidirectional encoders in stages: it opens bidirectional context, repairs the resulting mismatch with masked next-token prediction, then adds contrastive views encoding strand and positional invariance. Across two backbones, DNAGPT and HyenaDNA, and seven biological task families, the benefit is not uniform but depends on task structure and backbone; crop contrastive learning helps most for extended-context families such as enhancer–promoter interaction and Virus-COVID, reaching frozen Avg-EPI MCC of 27.5% and 26.7%.

Particle image velocimetry (PIV) usually fixes the imaging position before the measurement, so if the structure of interest departs from expectation, no later algorithm can recover information that was never captured. This work treats camera placement as a closed-loop decision variable: a DreamerV3 agent trained in Isaac Sim drives a 6-DoF arm using only the velocity field and uncertainty estimated by LIMA to locate, frame and resolve a hydrofoil wake. After zero-shot transfer to a real water tunnel, the agent locates wakes of randomly placed hydrofoils and measures them at higher resolution and lower flow-statistics error than a fixed view, in roughly half the time of a full scan and with far fewer image pairs.

Particle image velocimetry (PIV) usually fixes the imaging position before the measurement, so if the structure of interest departs from expectation, no later algorithm can recover information that was never captured. This work treats camera placement as a closed-loop decision variable: a DreamerV3 agent trained in Isaac Sim drives a 6-DoF arm using only the velocity field and uncertainty estimated by LIMA to locate, frame and resolve a hydrofoil wake. After zero-shot transfer to a real water tunnel, the agent locates wakes of randomly placed hydrofoils and measures them at higher resolution and lower flow-statistics error than a fixed view, in roughly half the time of a full scan and with far fewer image pairs.

Remote-sensing image fusion and semantic interpretation have usually been optimized separately, so good fusion metrics need not preserve downstream retrieval or detection utility. Lance maps each input to a closed-vocabulary text record of scene, objects and layout, conditions fusion on that language, re-describes the fused image as text, and back-propagates through text consistency and an anti-collapse text–mask contrastive loss with negative queries. Across seven test sets on GF2, QuickBird, WorldView-III and WorldView-II, reduced-scale reference metrics improve by up to 11.7% (median 2.1%) over the strongest learning baseline, and a held-out RemoteCLIP verifier plus a supervised DOTA detector give +6.4 mAP@0.5 over the closest competitor.

Remote-sensing image fusion and semantic interpretation have usually been optimized separately, so good fusion metrics need not preserve downstream retrieval or detection utility. Lance maps each input to a closed-vocabulary text record of scene, objects and layout, conditions fusion on that language, re-describes the fused image as text, and back-propagates through text consistency and an anti-collapse text–mask contrastive loss with negative queries. Across seven test sets on GF2, QuickBird, WorldView-III and WorldView-II, reduced-scale reference metrics improve by up to 11.7% (median 2.1%) over the strongest learning baseline, and a held-out RemoteCLIP verifier plus a supervised DOTA detector give +6.4 mAP@0.5 over the closest competitor.

Bulk mucosal RNA sequencing in inflammatory bowel disease has been analyzed one cohort at a time, so historical data accumulate without becoming a reusable reference. IBDex trained a tissue-aware conditional variational autoencoder on 3,168 biopsies from 1,680 patients across four cohorts and suppressed cohort identity adversarially, then froze the gene interface, encoder, decoder, pathway basis and phenotype directions. After freezing, dataset identity fell to balanced accuracy 0.307 against chance 0.250 while inflammation stayed recoverable (AUC 0.82); in 93 patients with concordant endoscopic and histological remission, inflammation approached controls (AUC 0.58) but disease state stayed displaced (AUC 0.725).

Bulk mucosal RNA sequencing in inflammatory bowel disease has been analyzed one cohort at a time, so historical data accumulate without becoming a reusable reference. IBDex trained a tissue-aware conditional variational autoencoder on 3,168 biopsies from 1,680 patients across four cohorts and suppressed cohort identity adversarially, then froze the gene interface, encoder, decoder, pathway basis and phenotype directions. After freezing, dataset identity fell to balanced accuracy 0.307 against chance 0.250 while inflammation stayed recoverable (AUC 0.82); in 93 patients with concordant endoscopic and histological remission, inflammation approached controls (AUC 0.58) but disease state stayed displaced (AUC 0.725).

Real-time digital twins of permanent magnet synchronous machines have relied on offline-identified parameters or finite-element flux look-up tables, which must be regenerated when the machine changes. The Inferencer-in-the-Loop architecture has a field-programmable gate array execute a physics-guided network while a GPU retrains it from reference data, switching weights only after validation. Across damper circuits, shifted parameters and an FEA-calibrated saturated machine, one kernel form reaches average normalized mean absolute errors of 0.19%–0.67% at measured kernel latencies of 0.30–0.34 microseconds; on a 200-W drive, a residual form gives errors of 5.9%–9.4% of measured RMS over six held-out runs.

Real-time digital twins of permanent magnet synchronous machines have relied on offline-identified parameters or finite-element flux look-up tables, which must be regenerated when the machine changes. The Inferencer-in-the-Loop architecture has a field-programmable gate array execute a physics-guided network while a GPU retrains it from reference data, switching weights only after validation. Across damper circuits, shifted parameters and an FEA-calibrated saturated machine, one kernel form reaches average normalized mean absolute errors of 0.19%–0.67% at measured kernel latencies of 0.30–0.34 microseconds; on a 200-W drive, a residual form gives errors of 5.9%–9.4% of measured RMS over six held-out runs.

Physics-informed neural networks (PINNs) often lose accuracy on PDEs with sharp gradients, partly because the network struggles to represent high-frequency variation. CTPINN adds a learnable coordinate transformation alongside the solution network, mapping the physical domain to a computational domain where the solution is approximately affine and gradients are more evenly distributed. The explicit construction restores accuracy on boundary-layer problems, and the implicit construction, plugged into PirateNet, SPINN and QRPINN, improves accuracy across several benchmark equations.

Physics-informed neural networks (PINNs) often lose accuracy on PDEs with sharp gradients, partly because the network struggles to represent high-frequency variation. CTPINN adds a learnable coordinate transformation alongside the solution network, mapping the physical domain to a computational domain where the solution is approximately affine and gradients are more evenly distributed. The explicit construction restores accuracy on boundary-layer problems, and the implicit construction, plugged into PirateNet, SPINN and QRPINN, improves accuracy across several benchmark equations.

Similarity between multimodal large language model representations and human visual cortex shows that the two are alike but not which semantics they share. SAIL recasts voxel interpretation as a correspondence between each voxel and multiple sparse semantic units: it trains a unified cross-layer sparse autoencoder on several model layers, matches each voxel to several units, and compares the semantic similarity of their preferred image sets with optimal transport. Across eight subjects in the Natural Scenes Dataset, joint cross-layer training reduced unit response redundancy and raised semantic similarity over layer-wise training, in both whole cortex and selective regions of interest.

Similarity between multimodal large language model representations and human visual cortex shows that the two are alike but not which semantics they share. SAIL recasts voxel interpretation as a correspondence between each voxel and multiple sparse semantic units: it trains a unified cross-layer sparse autoencoder on several model layers, matches each voxel to several units, and compares the semantic similarity of their preferred image sets with optimal transport. Across eight subjects in the Natural Scenes Dataset, joint cross-layer training reduced unit response redundancy and raised semantic similarity over layer-wise training, in both whole cortex and selective regions of interest.

Ottoman Divan poetry is written in aruz, a quantitative meter that counts each syllable as long or short, and modern Turkish spelling drops the marks that record long vowels, so length has to be inferred from context. A lattice-based hidden Markov model lays out candidate scholarly spellings of each word together with linking and lengthening choices as paths, is trained by EM from each poem's printed meter, and scans 77.6% of held-out lines fully correctly, against 26.8% for a rule-based scanner, 55.0% for a neural tagger trained with the same supervision and 63.6% for Claude Opus 5.5 few-shot on the same 500 lines. The code is open-sourced and the tool is released on the web.

Ottoman Divan poetry is written in aruz, a quantitative meter that counts each syllable as long or short, and modern Turkish spelling drops the marks that record long vowels, so length has to be inferred from context. A lattice-based hidden Markov model lays out candidate scholarly spellings of each word together with linking and lengthening choices as paths, is trained by EM from each poem's printed meter, and scans 77.6% of held-out lines fully correctly, against 26.8% for a rule-based scanner, 55.0% for a neural tagger trained with the same supervision and 63.6% for Claude Opus 5.5 few-shot on the same 500 lines. The code is open-sourced and the tool is released on the web.

On 509 unseen perturbations from the Replogle K562 screen, GEARS reaches a global Pearson correlation of 0.977 between predicted and observed transcriptomes, but only 0.076 once the control baseline is subtracted and the perturbation-induced delta is scored; a trivial baseline that predicts the same mean training delta for every perturbation reaches 0.267. The gap arises because global correlation is dominated by thousands of non-responsive genes, so delta-space evaluation is proposed as the primary benchmark, together with two post-hoc corrections that need no retraining and a retrieval upper bound that localizes the residual.

On 509 unseen perturbations from the Replogle K562 screen, GEARS reaches a global Pearson correlation of 0.977 between predicted and observed transcriptomes, but only 0.076 once the control baseline is subtracted and the perturbation-induced delta is scored; a trivial baseline that predicts the same mean training delta for every perturbation reaches 0.267. The gap arises because global correlation is dominated by thousands of non-responsive genes, so delta-space evaluation is proposed as the primary benchmark, together with two post-hoc corrections that need no retraining and a retrieval upper bound that localizes the residual.

MongoDB Research has released LEAF, a knowledge distillation framework, together with two 23M-parameter text embedding models, mdbr-leaf-ir and mdbr-leaf-mt. State-of-the-art embedding models have grown to hundreds of millions or billions of parameters, typically need GPU servers, and are not compatible with one another, so swapping models means recomputing every document embedding. LEAF aligns only the teacher and student output embeddings; the distilled models retain up to 97% of their teacher's performance while being 5–15× smaller and 7–24× faster, and can run offline on CPU-only servers, IoT and mobile devices.

MongoDB Research has released LEAF, a knowledge distillation framework, together with two 23M-parameter text embedding models, mdbr-leaf-ir and mdbr-leaf-mt. State-of-the-art embedding models have grown to hundreds of millions or billions of parameters, typically need GPU servers, and are not compatible with one another, so swapping models means recomputing every document embedding. LEAF aligns only the teacher and student output embeddings; the distilled models retain up to 97% of their teacher's performance while being 5–15× smaller and 7–24× faster, and can run offline on CPU-only servers, IoT and mobile devices.

GenISys splits early-stage bulk-carrier concept design into 20 subsystems and uses a simultaneous conditional denoising diffusion model so each subsystem reads the others' intermediate states at every reverse step, instead of generating them in a fixed order. In a five-fit confirmation at a fixed 5,000-nautical-mile voyage range, the model returned designs that both satisfied the 187 formulated engineering constraints and met a common low-cost target in 55.70% of generation requests, versus 51.10% for the monolithic model and 47.54% for the sequential model, corresponding to 9.0% and 17.2% more qualifying designs with similar estimated transport costs.

GenISys splits early-stage bulk-carrier concept design into 20 subsystems and uses a simultaneous conditional denoising diffusion model so each subsystem reads the others' intermediate states at every reverse step, instead of generating them in a fixed order. In a five-fit confirmation at a fixed 5,000-nautical-mile voyage range, the model returned designs that both satisfied the 187 formulated engineering constraints and met a common low-cost target in 55.70% of generation requests, versus 51.10% for the monolithic model and 47.54% for the sequential model, corresponding to 9.0% and 17.2% more qualifying designs with similar estimated transport costs.

After progressive Gaussian noise was added to LaMem images, mainstream memorability models such as ResMem, ViTMem and AMNet estimated lower memorability for initially memorable images and higher memorability for initially forgettable ones, a reversal resembling stochastic resonance. In a continuous recognition task with 88 participants, memorable images proved more resistant to noise while forgettable images declined sharply. The authors built GraphMem, which estimates memorability from an image's neighborhood in a similarity graph rather than from single-image features; it aligned most closely with human data under noise and blur, and after noise training its slope effect reached ρ=0.57 (95% bootstrap CI [0.29, 0.75]), the only model to turn positive.

After progressive Gaussian noise was added to LaMem images, mainstream memorability models such as ResMem, ViTMem and AMNet estimated lower memorability for initially memorable images and higher memorability for initially forgettable ones, a reversal resembling stochastic resonance. In a continuous recognition task with 88 participants, memorable images proved more resistant to noise while forgettable images declined sharply. The authors built GraphMem, which estimates memorability from an image's neighborhood in a similarity graph rather than from single-image features; it aligned most closely with human data under noise and blur, and after noise training its slope effect reached ρ=0.57 (95% bootstrap CI [0.29, 0.75]), the only model to turn positive.

RNA inverse folding has long relied on iterative optimization scored by an external folding algorithm, which is sensitive to initialization and prone to local optima. This work defines multinomial diffusion directly over nucleotide sequences, guides denoising with a target secondary structure, and compares a 2D contact-map projection with a purely 1D pairing-aware architecture. On the in-distribution test set, the 2D and 1D models reach median normalized RNAforester similarities of 0.82 and 0.79, while generated sequences stay close to random-sequence diversity.

RNA inverse folding has long relied on iterative optimization scored by an external folding algorithm, which is sensitive to initialization and prone to local optima. This work defines multinomial diffusion directly over nucleotide sequences, guides denoising with a target secondary structure, and compares a 2D contact-map projection with a purely 1D pairing-aware architecture. On the in-distribution test set, the 2D and 1D models reach median normalized RNAforester similarities of 0.82 and 0.79, while generated sequences stay close to random-sequence diversity.

HPE Labs proposed and validated SEPhIA, an optoelectronic spiking neural network architecture built on microring modulators and multi-wavelength lasers that lets many spiking neurons share one multi-wavelength light source, pushing the laser-per-neuron ratio below one. Prior photonic spiking work focused mostly on individual spiking devices or all-optical laser networks and rarely addressed optical power, crosstalk and footprint together at the architecture level. Using experimentally derived device parameters, a trainable optoelectronic spiking model reached 91.35% test accuracy on four-class spike-encoded image classification, about one percentage point behind a same-size fully connected spiking baseline at 93.65%.

HPE Labs proposed and validated SEPhIA, an optoelectronic spiking neural network architecture built on microring modulators and multi-wavelength lasers that lets many spiking neurons share one multi-wavelength light source, pushing the laser-per-neuron ratio below one. Prior photonic spiking work focused mostly on individual spiking devices or all-optical laser networks and rarely addressed optical power, crosstalk and footprint together at the architecture level. Using experimentally derived device parameters, a trainable optoelectronic spiking model reached 91.35% test accuracy on four-class spike-encoded image classification, about one percentage point behind a same-size fully connected spiking baseline at 93.65%.

Attention aggregation on heterogeneous graphs has long treated node types as static labels, and metapath aggregation can overlook the semantic contribution of intermediate nodes. TAHAN instead collects neighbors along metapath prefixes, merges them across metapaths and levels by node type, then fuses them with within-type attention followed by across-type attention. On DBLP and IMDB node classification its mean Macro-F1 and Micro-F1 exceed the listed baselines, with both IMDB gains significant after Holm correction; on ACM it falls below SeHGNN.

Attention aggregation on heterogeneous graphs has long treated node types as static labels, and metapath aggregation can overlook the semantic contribution of intermediate nodes. TAHAN instead collects neighbors along metapath prefixes, merges them across metapaths and levels by node type, then fuses them with within-type attention followed by across-type attention. On DBLP and IMDB node classification its mean Macro-F1 and Micro-F1 exceed the listed baselines, with both IMDB gains significant after Holm correction; on ACM it falls below SeHGNN.

Atlas is an autoregressive world model that previously took tens of seconds to minutes to produce a generated video, so users planned camera paths cautiously. The newly released Frameboy research preview serves a highly optimized Atlas variant from cloud GPUs and streams frames to the browser at up to 15 fps, updating the view continuously as the camera moves. The demo can generate scenes from a single image or text caption and retains and improves novel view synthesis from multiple images, but it must run on a powerful cloud GPU server and costs more than local rendering.

Atlas is an autoregressive world model that previously took tens of seconds to minutes to produce a generated video, so users planned camera paths cautiously. The newly released Frameboy research preview serves a highly optimized Atlas variant from cloud GPUs and streams frames to the browser at up to 15 fps, updating the view continuously as the camera moves. The demo can generate scenes from a single image or text caption and retains and improves novel view synthesis from multiple images, but it must run on a powerful cloud GPU server and costs more than local rendering.

The study introduces "trap molecules," an adversarial evaluation idea in which an AI agent uses reaction SMARTS templates and USPTO reaction records to deliberately design molecules carrying sensitive groups, near-duplicate reactive sites or bulky groups that lure retrosynthesis algorithms into characteristic errors. On 100 chemoselectivity, 100 regioselectivity and 27 steric trap molecules, three algorithms built on different principles, UAlign, SimpRetro and RetroChimera, all showed clear weaknesses in their top-20 predictions, with overall problem rates of 66%–77%, 85%–93% and 92.6%–100% for the three problem types.

The study introduces "trap molecules," an adversarial evaluation idea in which an AI agent uses reaction SMARTS templates and USPTO reaction records to deliberately design molecules carrying sensitive groups, near-duplicate reactive sites or bulky groups that lure retrosynthesis algorithms into characteristic errors. On 100 chemoselectivity, 100 regioselectivity and 27 steric trap molecules, three algorithms built on different principles, UAlign, SimpRetro and RetroChimera, all showed clear weaknesses in their top-20 predictions, with overall problem rates of 66%–77%, 85%–93% and 92.6%–100% for the three problem types.

Multi-disease retinal screening needs a per-image decision on whether an automated call can stand without review. RetinaGuard combines evidence from quality, diagnostic, uncertainty, verification and explanation agents into a learned probability that a decision is unsafe, invokes tools only when their expected value exceeds their cost, and certifies the acceptance threshold with conformal risk control. On the RFMiD test split with the weaker primary model (macro-AUROC 0.896), it captured 119 of 280 unsafe decisions at a 20% referral budget against 88 for lowest-confidence referral; with the stronger model (0.925) the verifier's gain shrank to a few percentage points.

Multi-disease retinal screening needs a per-image decision on whether an automated call can stand without review. RetinaGuard combines evidence from quality, diagnostic, uncertainty, verification and explanation agents into a learned probability that a decision is unsafe, invokes tools only when their expected value exceeds their cost, and certifies the acceptance threshold with conformal risk control. On the RFMiD test split with the weaker primary model (macro-AUROC 0.896), it captured 119 of 280 unsafe decisions at a 20% referral budget against 88 for lowest-confidence referral; with the stronger model (0.925) the verifier's gain shrank to a few percentage points.

In the dilepton channel, two neutrinos escape undetected, so reconstructing top-quark-pair kinematics has long relied on analytic approximations. Using a 7 TeV simulated sample from CMS Open Data, researchers first refined the loose kinematic reconstruction into LKRv3, then trained a residual multilayer perceptron and a conditional normalizing flow to regress the top-pair four-momenta directly from visible particles; relative to LKRv3, the two networks improve the resolution of the invariant mass, transverse momentum and rapidity by about 30%, 20% and 15%, reaching 30 GeV invariant-mass resolution near the production threshold.

In the dilepton channel, two neutrinos escape undetected, so reconstructing top-quark-pair kinematics has long relied on analytic approximations. Using a 7 TeV simulated sample from CMS Open Data, researchers first refined the loose kinematic reconstruction into LKRv3, then trained a residual multilayer perceptron and a conditional normalizing flow to regress the top-pair four-momenta directly from visible particles; relative to LKRv3, the two networks improve the resolution of the invariant mass, transverse momentum and rapidity by about 30%, 20% and 15%, reaching 30 GeV invariant-mass resolution near the production threshold.

Discount curves are usually estimated day by day with kernel ridge regression, but a good daily fit does not guarantee that the estimated curves can serve as states of a continuous-time arbitrage-free model. This work asks whether the finite-dimensional hypothesis space of kernel ridge regression is invariant under the risk-neutral drift of the discount curve for every grid of cash flow dates. Apart from kernels whose Hilbert space has dimension at most two, only the family built from the Segal–Bargmann kernel qualifies; for that kernel the estimate is an affine combination of flat discount curves, and the drift-invariant subspaces that allow dimension reduction correspond to sub-grids of the cash flow dates.

Discount curves are usually estimated day by day with kernel ridge regression, but a good daily fit does not guarantee that the estimated curves can serve as states of a continuous-time arbitrage-free model. This work asks whether the finite-dimensional hypothesis space of kernel ridge regression is invariant under the risk-neutral drift of the discount curve for every grid of cash flow dates. Apart from kernels whose Hilbert space has dimension at most two, only the family built from the Segal–Bargmann kernel qualifies; for that kernel the estimate is an affine combination of flat discount curves, and the drift-invariant subspaces that allow dimension reduction correspond to sub-grids of the cash flow dates.

AstraZeneca and Oracle Health and Life Sciences announced a global AI collaboration combining AstraZeneca's clinical development and regulatory expertise with Oracle's real-world data, cloud infrastructure and AI-enabled clinical capabilities, aiming to speed clinical research and bring evidence into care workflows. The stated early areas include clinical trial recruitment, earlier diagnosis, drug-safety monitoring and real-world evidence across a medicine's lifecycle, spanning oncology, cardiovascular, renal and metabolic, respiratory and immunology, and rare disease. The first programme focuses on hypertension care and clinical research in Abu Dhabi with healthcare company M42; the companies signed a non-binding term sheet and disclosed no financials.

AstraZeneca and Oracle Health and Life Sciences announced a global AI collaboration combining AstraZeneca's clinical development and regulatory expertise with Oracle's real-world data, cloud infrastructure and AI-enabled clinical capabilities, aiming to speed clinical research and bring evidence into care workflows. The stated early areas include clinical trial recruitment, earlier diagnosis, drug-safety monitoring and real-world evidence across a medicine's lifecycle, spanning oncology, cardiovascular, renal and metabolic, respiratory and immunology, and rare disease. The first programme focuses on hypertension care and clinical research in Abu Dhabi with healthcare company M42; the companies signed a non-binding term sheet and disclosed no financials.

A small transformer with a single attention layer was trained to predict the left and right descent sets of elements of the infinite Coxeter group Ã₂, meaning the possible first and last letters of minimal-length words. Test accuracy improved as training data moved from random words to reduced words to ShortLex normal-form words: the random-word and reduced-word datasets each began deteriorating at some length, while the set of all 1,998 ShortLex words of length at most 36 never deteriorated. From the trained model's attention patterns, the researchers formulated and independently proved two word-theory propositions about right and left descent sets.

A small transformer with a single attention layer was trained to predict the left and right descent sets of elements of the infinite Coxeter group Ã₂, meaning the possible first and last letters of minimal-length words. Test accuracy improved as training data moved from random words to reduced words to ShortLex normal-form words: the random-word and reduced-word datasets each began deteriorating at some length, while the set of all 1,998 ShortLex words of length at most 36 never deteriorated. From the trained model's attention patterns, the researchers formulated and independently proved two word-theory propositions about right and left descent sets.

Autonomous-driving policies react in closed loop, so a scenario scripted with fixed routes often drifts away from the interaction it was meant to stage, a problem called closed-loop drift. A scenario orchestrator continuously monitors the joint traffic state and adapts non-ego actors through re-timing, re-routing and role reassignment, solving a constrained optimization at each step so the test intent survives while interventions stay limited and traffic stays feasible. In the CARLA simulator across six scenario families and six ego policies, the method fully realizes the intended interaction in every scenario family, while baselines such as ProSim, LASER and OSC2Runner mostly satisfy only the two easiest scenes.

Autonomous-driving policies react in closed loop, so a scenario scripted with fixed routes often drifts away from the interaction it was meant to stage, a problem called closed-loop drift. A scenario orchestrator continuously monitors the joint traffic state and adapts non-ego actors through re-timing, re-routing and role reassignment, solving a constrained optimization at each step so the test intent survives while interventions stay limited and traffic stays feasible. In the CARLA simulator across six scenario families and six ego policies, the method fully realizes the intended interaction in every scenario family, while baselines such as ProSim, LASER and OSC2Runner mostly satisfy only the two easiest scenes.

The open-source SRAM design tool OpenYieldV2 turns bitcell topology from a fixed input into an optimizable variable searched jointly with transistor sizing and threshold-voltage assignment. Previously OpenRAM generated layouts without optimizing, and OpenYield's high-fidelity optimization covered only a fixed 6T cell. OpenYieldV2 compiles 6T and 10T macros under identical high-fidelity conditions, predicts performance with a few-shot TabPFN surrogate, and searches with a coarse-to-refined multi-objective optimizer; in mixed-topology scenarios its hypervolume improves by up to 17.3% over the strongest single-topology baseline.

The open-source SRAM design tool OpenYieldV2 turns bitcell topology from a fixed input into an optimizable variable searched jointly with transistor sizing and threshold-voltage assignment. Previously OpenRAM generated layouts without optimizing, and OpenYield's high-fidelity optimization covered only a fixed 6T cell. OpenYieldV2 compiles 6T and 10T macros under identical high-fidelity conditions, predicts performance with a few-shot TabPFN surrogate, and searches with a coarse-to-refined multi-objective optimizer; in mixed-topology scenarios its hypervolume improves by up to 17.3% over the strongest single-topology baseline.

MongoDB released and open-sourced an integration that turns MongoDB Atlas into the persistent memory store for AI applications and agents built with the Vercel AI SDK. The SDK itself ships no database, so an application previously started every conversation with no knowledge of the last one, and developers had to design schemas, write cleanup jobs and set up vector search themselves. The integration sorts memories into session, semantic, procedural, episodic and scratchpad types with separate retention policies, and creates vector indexes automatically while detecting embedding dimensions at startup.

MongoDB released and open-sourced an integration that turns MongoDB Atlas into the persistent memory store for AI applications and agents built with the Vercel AI SDK. The SDK itself ships no database, so an application previously started every conversation with no knowledge of the last one, and developers had to design schemas, write cleanup jobs and set up vector search themselves. The integration sorts memories into session, semantic, procedural, episodic and scratchpad types with separate retention policies, and creates vector indexes automatically while detecting embedding dimensions at startup.

MongoDB released Atlas App Connections, which lets AI coding tools connect to an Atlas account with a single click of authorization. Until now the only route was sharing long-lived service account credentials, which slowed builders and left administrators managing a security risk they were reluctant to accept, so some organizations avoided third-party integrations. Built on OAuth 2.1, it lets AI clients act on the user's behalf through the Atlas MCP Server, scoped to the user's existing permissions, revocable at any time, and off by default for existing organizations until an administrator opts in.

MongoDB released Atlas App Connections, which lets AI coding tools connect to an Atlas account with a single click of authorization. Until now the only route was sharing long-lived service account credentials, which slowed builders and left administrators managing a security risk they were reluctant to accept, so some organizations avoided third-party integrations. Built on OAuth 2.1, it lets AI clients act on the user's behalf through the Atlas MCP Server, scoped to the user's existing permissions, revocable at any time, and off by default for existing organizations until an administrator opts in.

EEG fatigue detection has treated multi-band modeling and temporal weighting as separate concerns, without defining how the spectral structure inside a window should shape that window's contribution. LBSNet encodes each window's relative cross-band activity as a normalized descriptor, softly matches it to learnable spectral prototypes, and combines the assignments with prototype importance to weight power aggregation. In subject-independent experiments on SEED-VIG and SADT it beats seven baselines on accuracy and F1 while using the fewest parameters and multiply-accumulate operations among the compared models.

EEG fatigue detection has treated multi-band modeling and temporal weighting as separate concerns, without defining how the spectral structure inside a window should shape that window's contribution. LBSNet encodes each window's relative cross-band activity as a normalized descriptor, softly matches it to learnable spectral prototypes, and combines the assignments with prototype importance to weight power aggregation. In subject-independent experiments on SEED-VIG and SADT it beats seven baselines on accuracy and F1 while using the fewest parameters and multiply-accumulate operations among the compared models.

MongoDB has released a federated Copilot connector for MongoDB Atlas, now generally available, letting users ask Microsoft 365 Copilot questions about live Atlas data. Previously such answers lived in operational applications built on Atlas while the questions were asked in everyday productivity tools, or relied on a static copy of the data. The connector queries Atlas on demand through MongoDB's managed remote MCP server, with no migration or indexed copy, and each query runs as the signed-in user under existing Atlas permissions.

MongoDB has released a federated Copilot connector for MongoDB Atlas, now generally available, letting users ask Microsoft 365 Copilot questions about live Atlas data. Previously such answers lived in operational applications built on Atlas while the questions were asked in everyday productivity tools, or relied on a static copy of the data. The connector queries Atlas on demand through MongoDB's managed remote MCP server, with no migration or indexed copy, and each query runs as the signed-in user under existing Atlas permissions.

MongoDB released an integration with Google ADK that gives agent state durable memory. Previously developers still had to build their own tooling for a data store that could retain agent context over time. The integration uses MongoDbSessionService to persist sessions, events, and application-, user-, and session-scoped state to MongoDB, and MongoDBMemoryService to ingest session events into collections and recall memories through keyword matching; retrieval reuses MongoDB vector search combined with full-text search. It supports MongoDB Atlas and self-managed MongoDB 8.0+ deployments.

MongoDB released an integration with Google ADK that gives agent state durable memory. Previously developers still had to build their own tooling for a data store that could retain agent context over time. The integration uses MongoDbSessionService to persist sessions, events, and application-, user-, and session-scoped state to MongoDB, and MongoDBMemoryService to ingest session events into collections and recall memories through keyword matching; retrieval reuses MongoDB vector search combined with full-text search. It supports MongoDB Atlas and self-managed MongoDB 8.0+ deployments.

MongoDB has released Automated Embedding in MongoDB Vector Search as a public preview in Community Edition: a developer chooses an embedding model while creating a vector search index, and text vectors are generated and kept in sync during initial data sync, document inserts and updates, and query processing. Previously developers called external embedding providers from application code before inserting data or running a vector query, and built custom pipelines to keep source data and vectors aligned on every insert, update or delete, plus an extra API call per query. The feature ships Voyage AI's voyage-4-large, voyage-4, voyage-4-lite and voyage-code-3 models and uses asynchronous dynamic batching that adjusts batch size to document shape.

MongoDB has released Automated Embedding in MongoDB Vector Search as a public preview in Community Edition: a developer chooses an embedding model while creating a vector search index, and text vectors are generated and kept in sync during initial data sync, document inserts and updates, and query processing. Previously developers called external embedding providers from application code before inserting data or running a vector query, and built custom pipelines to keep source data and vectors aligned on every insert, update or delete, plus an extra API call per query. The feature ships Voyage AI's voyage-4-large, voyage-4, voyage-4-lite and voyage-code-3 models and uses asynchronous dynamic batching that adjusts batch size to document shape.