AI and science frontiers · 2026-11-12

Alibaba DAMO Academy's HoloCT is a 3D CT foundation model that puts scan interpretation and voxel-level segmentation inside one vision–language model, so a textual finding can come with a checkable 3D mask. Previously, segmentation networks delineated without interpreting, and CT vision–language models wrote text without voxel-level localization. On 11,455 clinical CT scans from five hospitals, HoloCT-8B raised mean Dice over VoxTell from 0.107 to 0.488 on laryngeal lesions and from 0.543 to 0.715 on lung cancer, and in retrospective reader studies of thyroid tumours and colon lesions it exceeded all four participating radiologists in sensitivity and overall accuracy.

Alibaba DAMO Academy's HoloCT is a 3D CT foundation model that puts scan interpretation and voxel-level segmentation inside one vision–language model, so a textual finding can come with a checkable 3D mask. Previously, segmentation networks delineated without interpreting, and CT vision–language models wrote text without voxel-level localization. On 11,455 clinical CT scans from five hospitals, HoloCT-8B raised mean Dice over VoxTell from 0.107 to 0.488 on laryngeal lesions and from 0.543 to 0.715 on lung cancer, and in retrospective reader studies of thyroid tumours and colon lesions it exceeded all four participating radiologists in sensitivity and overall accuracy.

SO3LR Tauri extends the earlier SO3LR force field from eight elements and 4 million structures to 85 elements and roughly 148 million structures, and rewrites its short-range repulsion, electrostatics and dispersion modules. Keeping the fast SO3krates message-passing backbone, it reproduces nucleobase-pair and protein–ligand interactions and carbohydrate conformer ordering against coupled-cluster references, and yields water and organic-solvent densities, ion hydration and structures of several biomolecular classes in dynamics without system-specific fitting. The models ship in three sizes and scale to the million-atom range on a single GPU.

SO3LR Tauri extends the earlier SO3LR force field from eight elements and 4 million structures to 85 elements and roughly 148 million structures, and rewrites its short-range repulsion, electrostatics and dispersion modules. Keeping the fast SO3krates message-passing backbone, it reproduces nucleobase-pair and protein–ligand interactions and carbohydrate conformer ordering against coupled-cluster references, and yields water and organic-solvent densities, ion hydration and structures of several biomolecular classes in dynamics without system-specific fitting. The models ship in three sizes and scale to the million-atom range on a single GPU.

A team used RFdiffusion3 to design ubiquitin ligases from scratch around a cysteine catalytic triad, conjugating an activated ubiquitin donor directly to a chosen lysine on a target protein. Against a 19-residue model peptide with a single lysine, 35 of 48 designs showed measurable ubiquitylation activity, and four reached near-quantitative transfer without experimental optimization. Against the homotrimeric sliding clamp PCNA, only four of 48 designs were active, and PL4 modified K164 alone; a single N148A substitution then let PL4.1 convert 90% of PCNA to ubiquitylated product in 3 hours.

A team used RFdiffusion3 to design ubiquitin ligases from scratch around a cysteine catalytic triad, conjugating an activated ubiquitin donor directly to a chosen lysine on a target protein. Against a 19-residue model peptide with a single lysine, 35 of 48 designs showed measurable ubiquitylation activity, and four reached near-quantitative transfer without experimental optimization. Against the homotrimeric sliding clamp PCNA, only four of 48 designs were active, and PL4 modified K164 alone; a single N148A substitution then let PL4.1 convert 90% of PCNA to ubiquitylated product in 3 hours.

AlphaGenome 2 moves prediction of DNA variant effects from bulk-tissue averages down to specific cell types and cell states, recasting the model as a conditional one that takes a target biological condition as input. On the TenK10K single-cell map of 28 primary immune cell types, it identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and 43% versus 23% for directional gene effects. Trained on more than 107 million individual cell profiles across 68 tissues, it is available through the AlphaGenome API for non-commercial research.

AlphaGenome 2 moves prediction of DNA variant effects from bulk-tissue averages down to specific cell types and cell states, recasting the model as a conditional one that takes a target biological condition as input. On the TenK10K single-cell map of 28 primary immune cell types, it identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and 43% versus 23% for directional gene effects. Trained on more than 107 million individual cell profiles across 68 tissues, it is available through the AlphaGenome API for non-commercial research.

Hybrid quantum-classical Monte Carlo had been limited by the large classical cost of repeatedly processing quantum measurement data, with hardware demonstrations reaching only 16 qubits. Compressing those measurements into a differentiable machine-learning surrogate that supplies the wavefunction overlaps and derivatives Monte Carlo needs accelerates the end-to-end calculation by orders of magnitude on Google's Willow processor, bringing hydrogen chains and iron-sulfur clusters to chemical accuracy on up to 40 qubits, more than doubling the scale of earlier hardware demonstrations.

Hybrid quantum-classical Monte Carlo had been limited by the large classical cost of repeatedly processing quantum measurement data, with hardware demonstrations reaching only 16 qubits. Compressing those measurements into a differentiable machine-learning surrogate that supplies the wavefunction overlaps and derivatives Monte Carlo needs accelerates the end-to-end calculation by orders of magnitude on Google's Willow processor, bringing hydrogen chains and iron-sulfur clusters to chemical accuracy on up to 40 qubits, more than doubling the scale of earlier hardware demonstrations.

Glyph decomposes clinical foundation-model embeddings into sparse, nameable concepts so that features previously entangled in high-dimensional vectors can be tested like clinical variables. Across abdominal CT, radiology reports and laboratory data from 41,628 patients, the study tested 604,185 concept–outcome pairs and retained 397 high-risk concepts with 4,463 Bonferroni-significant associations. Most associations reproduced in three held-out cohorts without recalibration, and the top ten concepts reproduced more often than markers recovered from the literature.

Glyph decomposes clinical foundation-model embeddings into sparse, nameable concepts so that features previously entangled in high-dimensional vectors can be tested like clinical variables. Across abdominal CT, radiology reports and laboratory data from 41,628 patients, the study tested 604,185 concept–outcome pairs and retained 397 high-risk concepts with 4,463 Bonferroni-significant associations. Most associations reproduced in three held-out cohorts without recalibration, and the top ten concepts reproduced more often than markers recovered from the literature.

PharmAgents2 moves drug discovery from single-task models to an end-to-end autonomous workflow: it distills transferable decision knowledge from published literature and uses parallel agents to explore target, hypothesis and strategy spaces. On the antibacterial target LpxH it autonomously hypothesized a previously unreported allosteric site and screened against it, yielding AIPher-L1 with experimentally confirmed inhibitory activity and binding; JAK1 work and secondary-target exploration for the known molecule AIPher-X were supported by cell assays and SPR experiments respectively.

PharmAgents2 moves drug discovery from single-task models to an end-to-end autonomous workflow: it distills transferable decision knowledge from published literature and uses parallel agents to explore target, hypothesis and strategy spaces. On the antibacterial target LpxH it autonomously hypothesized a previously unreported allosteric site and screened against it, yielding AIPher-L1 with experimentally confirmed inhibitory activity and binding; JAK1 work and secondary-target exploration for the known molecule AIPher-X were supported by cell assays and SPR experiments respectively.

Latent-charge machine-learning potentials learn atomic charges as hidden variables from energies and forces, but their long-range electrostatic term had no exact periodic form, leaving net charge mixed with self-energy. Writing that term as a smeared Epstein–Hurwitz lattice sum separates the net-charge artefact from the local self-energy and yields a sphere-average theorem for cells of any shape, plus forces and virial. Projecting the latent charges onto fixed total charge gives projected latent Ewald summation (pLES); on charged NaCl clusters it recovers the total charge an unconstrained model misses by up to 1.8 and cuts a short-range model's test errors from 1.44 to 0.42 meV/atom and from 37 to 13 meV/Å.

Latent-charge machine-learning potentials learn atomic charges as hidden variables from energies and forces, but their long-range electrostatic term had no exact periodic form, leaving net charge mixed with self-energy. Writing that term as a smeared Epstein–Hurwitz lattice sum separates the net-charge artefact from the local self-energy and yields a sphere-average theorem for cells of any shape, plus forces and virial. Projecting the latent charges onto fixed total charge gives projected latent Ewald summation (pLES); on charged NaCl clusters it recovers the total charge an unconstrained model misses by up to 1.8 and cuts a short-range model's test errors from 1.44 to 0.42 meV/atom and from 37 to 13 meV/Å.

Open-HumanScale scales egocentric human manipulation video pretraining to 50,000 hours and releases the corpus, training pipelines and checkpoints. On a unified world-action model, pretraining hours and downstream robot task success follow an empirical log-linear relationship, and the curve steepens when task and scene diversity grow alongside duration. Pretrained on human data alone, HumanWAM-Vanilla reaches 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on the randomized Clean2Random setting.

Open-HumanScale scales egocentric human manipulation video pretraining to 50,000 hours and releases the corpus, training pipelines and checkpoints. On a unified world-action model, pretraining hours and downstream robot task success follow an empirical log-linear relationship, and the curve steepens when task and scene diversity grow alongside duration. Pretrained on human data alone, HumanWAM-Vanilla reaches 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on the randomized Clean2Random setting.

GenEA places an agent's coarse body biomechanics and its physical interactions with nearby objects into a multilevel state space, then embeds that space in a generative model to yield a joint probability distribution over goals and embodied plans. Earlier accounts of how people reason about plans mostly abstracted agents as point-like entities in two-dimensional grid worlds, which cannot carry the body and terrain constraints of natural settings. Across block-manipulation and obstacle-course navigation experiments, simple transformations of that joint distribution reproduce human goal inferences, action inferences, one-hand versus two-hand strategy preferences and difficulty judgments, outperforming a task-and-motion-planning baseline, heuristic models and a frontier vision-language model.

GenEA places an agent's coarse body biomechanics and its physical interactions with nearby objects into a multilevel state space, then embeds that space in a generative model to yield a joint probability distribution over goals and embodied plans. Earlier accounts of how people reason about plans mostly abstracted agents as point-like entities in two-dimensional grid worlds, which cannot carry the body and terrain constraints of natural settings. Across block-manipulation and obstacle-course navigation experiments, simple transformations of that joint distribution reproduce human goal inferences, action inferences, one-hand versus two-hand strategy preferences and difficulty judgments, outperforming a task-and-motion-planning baseline, heuristic models and a frontier vision-language model.

Long-horizon LLM agents interact with their environment through complete turns but generate each action token by token, and existing reinforcement learning either gives a whole trajectory one shared advantage or treats every token as an action. MPO splits credit assignment into two scales: turn-level generalized advantage estimation carries credit across interactions, a mean-preserving local correction refines it within each turn, and the policy update uses complete-turn importance ratios. Across search, tool-use and automation tasks, MPO reports the highest score on all nine metrics at both Qwen3-4B-Thinking and Qwen3-30B-A3B-Instruct scales, including a 4.33-point BrowseComp-Plus pass@1 gain over the strongest non-MPO baseline at 4B.

Long-horizon LLM agents interact with their environment through complete turns but generate each action token by token, and existing reinforcement learning either gives a whole trajectory one shared advantage or treats every token as an action. MPO splits credit assignment into two scales: turn-level generalized advantage estimation carries credit across interactions, a mean-preserving local correction refines it within each turn, and the policy update uses complete-turn importance ratios. Across search, tool-use and automation tasks, MPO reports the highest score on all nine metrics at both Qwen3-4B-Thinking and Qwen3-30B-A3B-Instruct scales, including a 4.33-point BrowseComp-Plus pass@1 gain over the strongest non-MPO baseline at 4B.

In multi-agent systems, one agent's faulty output can be read by peers and spread into external actions before anyone detects it. Airlock narrows permissions down a delegation tree and records cross-agent data consumption in an append-only provenance graph; after an alert it traces forward to isolate tainted executions and pending actions. Across 1,832 testbed runs, 161 physical smart-home runs and 24,266 simulation records, Airlock blocked all post-alert faulty actuations on its main testbed and smart-home grids with zero post-convergence leaks, retaining 68 to 96 percentage points more healthy work than native A2A and LangGraph controls.

In multi-agent systems, one agent's faulty output can be read by peers and spread into external actions before anyone detects it. Airlock narrows permissions down a delegation tree and records cross-agent data consumption in an append-only provenance graph; after an alert it traces forward to isolate tainted executions and pending actions. Across 1,832 testbed runs, 161 physical smart-home runs and 24,266 simulation records, Airlock blocked all post-alert faulty actuations on its main testbed and smart-home grids with zero post-convergence leaks, retaining 68 to 96 percentage points more healthy work than native A2A and LangGraph controls.

Safety alignment in open-weight models is highly sparse: pruning a small set of safety-critical neurons or ablating a single refusal direction can make a model produce harmful content. ABS-Align gradually prunes those identified safety-critical neurons during preference finetuning, forcing refusal behavior to be re-established through other parts of the network, and restores the pruned neurons afterward. Across Gemma-2-2B-IT, Llama-3-8B-Instruct, Llama-3.2-3B-Instruct and Qwen-3-4B, attack success generally falls against TwinBreak, NeuroStrike, refusal-direction ablation and ten SORRY-Bench black-box attacks, with reductions of about 90% for some attacks, while TruthfulQA improves and general capabilities change little.

Safety alignment in open-weight models is highly sparse: pruning a small set of safety-critical neurons or ablating a single refusal direction can make a model produce harmful content. ABS-Align gradually prunes those identified safety-critical neurons during preference finetuning, forcing refusal behavior to be re-established through other parts of the network, and restores the pruned neurons afterward. Across Gemma-2-2B-IT, Llama-3-8B-Instruct, Llama-3.2-3B-Instruct and Qwen-3-4B, attack success generally falls against TwinBreak, NeuroStrike, refusal-direction ablation and ten SORRY-Bench black-box attacks, with reductions of about 90% for some attacks, while TruthfulQA improves and general capabilities change little.

AI agents can now propose atom-level biological hypotheses faster than they can be tested, but trust in an agent-run molecular dynamics simulation rests on a process record, and existing agents leave that record to the language model itself. MDClaw pairs validating tools with a directed acyclic graph (DAG, a graph recording which calculation stages depend on which) that checks stages before execution and records each one; on MDDataBench, which grades against deposited simulations, four open-weight models with MDClaw passed all three runs on 81–100% of 31 test systems. An ablation without the DAG found 11 runs that passed every output check yet had skipped minimization or equilibration, or simulated a different structure.

AI agents can now propose atom-level biological hypotheses faster than they can be tested, but trust in an agent-run molecular dynamics simulation rests on a process record, and existing agents leave that record to the language model itself. MDClaw pairs validating tools with a directed acyclic graph (DAG, a graph recording which calculation stages depend on which) that checks stages before execution and records each one; on MDDataBench, which grades against deposited simulations, four open-weight models with MDClaw passed all three runs on 81–100% of 31 test systems. An ablation without the DAG found 11 runs that passed every output check yet had skipped minimization or equilibration, or simulated a different structure.

Dense shape correspondence on high-resolution 3D meshes has been held back by the quadratic cost of spatial-spectral coupling: for two shapes with k vertices each, a single coupling occupies roughly GB in single precision, so existing methods downsample heavily and transfer predictions back. Geo-FlashAttention rewrites that coupling as scaled dot-product attention and evaluates the exact softmax only over surface patches selected by geodesic distance, giving near-linear empirical scaling while preserving the correspondence accuracy of exact implementations. At vertices, peak training memory falls from 106.08 to 5.67 GiB, and each training step is faster than FlashAttention, making end-to-end correspondence learning feasible on meshes such as cortical surfaces.

Dense shape correspondence on high-resolution 3D meshes has been held back by the quadratic cost of spatial-spectral coupling: for two shapes with k vertices each, a single coupling occupies roughly GB in single precision, so existing methods downsample heavily and transfer predictions back. Geo-FlashAttention rewrites that coupling as scaled dot-product attention and evaluates the exact softmax only over surface patches selected by geodesic distance, giving near-linear empirical scaling while preserving the correspondence accuracy of exact implementations. At vertices, peak training memory falls from 106.08 to 5.67 GiB, and each training step is faster than FlashAttention, making end-to-end correspondence learning feasible on meshes such as cortical surfaces.

Physics-informed neural networks could previously show only small residuals at sampled points, not how large the error is over the continuous domain. This pipeline has a network search the Babenko equation in conformal coordinates for high-amplitude gravity–capillary waves, and staged neural corrections, Fourier projection and Newton polishing cut the maximum sampled residual by about seven orders of magnitude. The candidate is then frozen as an explicit trigonometric polynomial, interval arithmetic yields a rigorous uniform bound on the continuous residual, and Lean machine-checks the deductions from that numerical bound to the final residual theorem.

Physics-informed neural networks could previously show only small residuals at sampled points, not how large the error is over the continuous domain. This pipeline has a network search the Babenko equation in conformal coordinates for high-amplitude gravity–capillary waves, and staged neural corrections, Fourier projection and Newton polishing cut the maximum sampled residual by about seven orders of magnitude. The candidate is then frozen as an explicit trigonometric polynomial, interval arithmetic yields a rigorous uniform bound on the continuous residual, and Lean machine-checks the deductions from that numerical bound to the final residual theorem.

Separating a brain area's own dynamics from the inputs driving it is hard, because the same recorded activity can arise from many combinations of the two. DynISys models each recorded area as its own low-dimensional dynamical system, lets areas exchange signals only through a few explicit channels, and maps latent states to firing rates through an invertible readout. On simultaneous PMd and M1 recordings from two monkeys, eight latent dimensions per area matched the PSTH reconstruction of MR-LFADS at 128 per area, and the inferred PMd-to-M1 communication timing was consistent with target information traveling from PMd to M1.

Separating a brain area's own dynamics from the inputs driving it is hard, because the same recorded activity can arise from many combinations of the two. DynISys models each recorded area as its own low-dimensional dynamical system, lets areas exchange signals only through a few explicit channels, and maps latent states to firing rates through an invertible readout. On simultaneous PMd and M1 recordings from two monkeys, eight latent dimensions per area matched the PSTH reconstruction of MR-LFADS at 128 per area, and the inferred PMd-to-M1 communication timing was consistent with target information traveling from PMd to M1.

In multi-turn tool-use reinforcement learning, reasoning accuracy can rise early and then fall below the non-agentic baseline; TACA traces this to negative aggregate credit on tool-call tokens and instead supervises the step where reasoning turns into a tool call. Against MT-GRPO it raises mean accuracy over six reasoning benchmarks by 4.61 and 3.61 percentage points for Qwen2.5-1.5B and 3B, while tool execution success during training stays at 86.28%.

In multi-turn tool-use reinforcement learning, reasoning accuracy can rise early and then fall below the non-agentic baseline; TACA traces this to negative aggregate credit on tool-call tokens and instead supervises the step where reasoning turns into a tool call. Against MT-GRPO it raises mean accuracy over six reasoning benchmarks by 4.61 and 3.61 percentage points for Qwen2.5-1.5B and 3B, while tool execution success during training stays at 86.28%.

Computer-use agents encode each screenshot into thousands of visual tokens, and those tokens pile up into long contexts on long-horizon tasks and in real-time games. VTR trains a vision-language model jointly with a lightweight spatio-temporal router that inspects the same image region in the current and previous screenshots and decides whether that token enters the language model, using an adaptive loss to keep as few tokens as possible without losing action prediction. After only 2,000 training steps, VTR matches the full-resolution model on OSWorld, CUA-World and video-game action prediction while using 2.5–5.8× fewer visual tokens, cutting inference cost by up to 3.3× and speeding prefill by up to 2.9×.

Computer-use agents encode each screenshot into thousands of visual tokens, and those tokens pile up into long contexts on long-horizon tasks and in real-time games. VTR trains a vision-language model jointly with a lightweight spatio-temporal router that inspects the same image region in the current and previous screenshots and decides whether that token enters the language model, using an adaptive loss to keep as few tokens as possible without losing action prediction. After only 2,000 training steps, VTR matches the full-resolution model on OSWorld, CUA-World and video-game action prediction while using 2.5–5.8× fewer visual tokens, cutting inference cost by up to 3.3× and speeding prefill by up to 2.9×.

Tool-using LLM agents turn model output into real actions, making indirect prompt injection an execution-security problem. SecureGARD inserts an executable ontology between planning and tool execution, grounding heterogeneous tool calls into shared semantic actions and judging each as ready, waiting, or forbidden against the current task state. Under benign and indirect-injection settings on AgentDojo, ASB, and AgentDyn, the observed attack success rate is 0% on both backbones, with higher task utility under attack than the compared baselines.

Tool-using LLM agents turn model output into real actions, making indirect prompt injection an execution-security problem. SecureGARD inserts an executable ontology between planning and tool execution, grounding heterogeneous tool calls into shared semantic actions and judging each as ready, waiting, or forbidden against the current task state. Under benign and indirect-injection settings on AgentDojo, ASB, and AgentDyn, the observed attack success rate is 0% on both backbones, with higher task utility under attack than the compared baselines.

Most members of an Alzheimer's cohort contribute only one partially observed cross-sectional snapshot, leaving their stage, subtype and stage-varying multimodal coupling unresolved. ELPAC anchors subtype trajectories at clinically reliable endpoint groups and integrates backward along biologically constrained velocity fields, then factorizes residual multimodal variation around that stage axis into sparse coordination axes that strengthen or fade with stage. On synthetic data and a harmonized ADNI+NACC/SCAN cohort it recovers the latent structure, returns calibrated prediction intervals and yields multimodal coordination programs that differ by subtype.

Most members of an Alzheimer's cohort contribute only one partially observed cross-sectional snapshot, leaving their stage, subtype and stage-varying multimodal coupling unresolved. ELPAC anchors subtype trajectories at clinically reliable endpoint groups and integrates backward along biologically constrained velocity fields, then factorizes residual multimodal variation around that stage axis into sparse coordination axes that strengthen or fade with stage. On synthetic data and a harmonized ADNI+NACC/SCAN cohort it recovers the latent structure, returns calibrated prediction intervals and yields multimodal coordination programs that differ by subtype.

Vision models still fail on seemingly simple scenes, and a fixed dataset cannot reliably estimate worst-case error. Procedural adversarial learning treats an interpretable procedural generator as a controllable data source and uses zeroth-order optimization to search its low-dimensional control parameters for hard instances. On four tasks spanning surface normals, image segmentation and human pose estimation, Simple-PAT lowers worst-case error below uniform training and two robust reweighting baselines against a cross-model black-box adversary with 50M aggregate queries.

Vision models still fail on seemingly simple scenes, and a fixed dataset cannot reliably estimate worst-case error. Procedural adversarial learning treats an interpretable procedural generator as a controllable data source and uses zeroth-order optimization to search its low-dimensional control parameters for hard instances. On four tasks spanning surface normals, image segmentation and human pose estimation, Simple-PAT lowers worst-case error below uniform training and two robust reweighting baselines against a cross-model black-box adversary with 50M aggregate queries.

In multi-turn image editing, each output becomes the next input, so small unintended changes to untouched content accumulate into visible drift. TailSteer concentrates preservation corrections near the tail of the diffusion denoising trajectory while keeping early dynamics close to the base model, and learns them from the editor's own on-policy rollouts using unchanged source content as reference, without curated editing pairs. Across four backbones in the FLUX, SenseNova and Qwen families, cumulative drift over 15 to 20 editing turns drops substantially while text-to-image and single-turn editing benchmarks stay largely intact.

In multi-turn image editing, each output becomes the next input, so small unintended changes to untouched content accumulate into visible drift. TailSteer concentrates preservation corrections near the tail of the diffusion denoising trajectory while keeping early dynamics close to the base model, and learns them from the editor's own on-policy rollouts using unchanged source content as reference, without curated editing pairs. Across four backbones in the FLUX, SenseNova and Qwen families, cumulative drift over 15 to 20 editing turns drops substantially while text-to-image and single-turn editing benchmarks stay largely intact.

CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.

CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.

Quantizing SAM 3's image-encoder weights to 4 bits while keeping activations at 8 bits produces prompt-alignment drift: masks look plausible but land on the wrong object or lose the target across video frames. SPQ gives more calibration weight to parameters that affect text-concept scores, constrains feature error along low-variance residual directions, and folds a mean-error correction estimated from paired video rollouts into existing biases. Under W4A8, RefCOCO Pr@0.5 rises from 42.1 for PTQ4SAM to 72.5 while retaining 54.8 COCO mask AP, with no added inference operators.

Quantizing SAM 3's image-encoder weights to 4 bits while keeping activations at 8 bits produces prompt-alignment drift: masks look plausible but land on the wrong object or lose the target across video frames. SPQ gives more calibration weight to parameters that affect text-concept scores, constrains feature error along low-variance residual directions, and folds a mean-error correction estimated from paired video rollouts into existing biases. Under W4A8, RefCOCO Pr@0.5 rises from 42.1 for PTQ4SAM to 72.5 while retaining 54.8 COCO mask AP, with no added inference operators.

Answer probes on chain-of-thought hidden states are often read as evidence that a model has settled its final answer, yet a hidden state also carries the candidate whose name is about to be written. Emission-controlled dual-target probing (ECDP) fits the same low-rank reader separately to the next candidate mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed in advance. On DecisionDepth, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696; Llama-3.1-8B gives .575. The order is not universal: in a Qwen3-8B MMLU setting the final answer is read more strongly.

Answer probes on chain-of-thought hidden states are often read as evidence that a model has settled its final answer, yet a hidden state also carries the candidate whose name is about to be written. Emission-controlled dual-target probing (ECDP) fits the same low-rank reader separately to the next candidate mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed in advance. On DecisionDepth, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696; Llama-3.1-8B gives .575. The order is not universal: in a Qwen3-8B MMLU setting the final answer is read more strongly.

Large-vision-language-model logical anomaly detection splits each constraint into a textual checklist, yet every check still reads the same raw image, so the model must pull different kinds of visual evidence from one input. CASCADE extends the decomposition to the image side, overlaying per-instance color for binding constraints or an object-anchored grid for calibration constraints, and CASCADE++ derives each scaffold's configuration from the constraint structure and the reference distribution. On MVTec-LOCO, CASCADE++ with three references and no training exceeds the strongest full-shot method and transfers across several vision-language backbones.

Large-vision-language-model logical anomaly detection splits each constraint into a textual checklist, yet every check still reads the same raw image, so the model must pull different kinds of visual evidence from one input. CASCADE extends the decomposition to the image side, overlaying per-instance color for binding constraints or an object-anchored grid for calibration constraints, and CASCADE++ derives each scaffold's configuration from the constraint structure and the reference distribution. On MVTec-LOCO, CASCADE++ with three references and no training exceeds the strongest full-shot method and transfers across several vision-language backbones.

The information a fine-tune writes into its adapter is now measured not by parameter count or weight reconstruction error but as the smallest serialized size of a compressed adapter that, reloaded into the frozen model, keeps 90% of the held-out fine-tuning gain. The same dataset demands very different rates across frozen models: keeping 90% of the math gain takes 1.03 bits per value on Qwen2.5-Math-7B but 0.41 on Qwen2.5-7B, and on an adapter trained with deliberately mismatched reasoning supervision, compression lifts GSM8K accuracy from 23.7% to 82.9%, above the frozen Qwen2.5-7B's 77.6%.

The information a fine-tune writes into its adapter is now measured not by parameter count or weight reconstruction error but as the smallest serialized size of a compressed adapter that, reloaded into the frozen model, keeps 90% of the held-out fine-tuning gain. The same dataset demands very different rates across frozen models: keeping 90% of the math gain takes 1.03 bits per value on Qwen2.5-Math-7B but 0.41 on Qwen2.5-7B, and on an adapter trained with deliberately mismatched reasoning supervision, compression lifts GSM8K accuracy from 23.7% to 82.9%, above the frozen Qwen2.5-7B's 77.6%.

For a fixed feature distribution and an arbitrary model class, agnostic learning (training a discriminator that matches the best member of the class) and multiaccuracy (training a generator whose labels no observer in the class can distinguish from real ones) are proved equivalent in statistical and computational complexity. Only the direction from agnostic learning to multiaccuracy was known; the converse was blocked because a single multiaccurate predictor can fail to beat random guessing even after calibration. The new reduction repeatedly queries a multiaccuracy oracle, splits by prediction level sets and averages the iterates, yielding a predictor that is both multiaccurate and calibrated and thus performs agnostic learning.

For a fixed feature distribution and an arbitrary model class, agnostic learning (training a discriminator that matches the best member of the class) and multiaccuracy (training a generator whose labels no observer in the class can distinguish from real ones) are proved equivalent in statistical and computational complexity. Only the direction from agnostic learning to multiaccuracy was known; the converse was blocked because a single multiaccurate predictor can fail to beat random guessing even after calibration. The new reduction repeatedly queries a multiaccuracy oracle, splits by prediction level sets and averages the iterates, yielding a predictor that is both multiaccurate and calibrated and thus performs agnostic learning.

For retrieval-augmented generation and agent workloads, Nebula promotes position-independent caching (PIC) segments, which let the same text segment reuse already-computed key-value cache under a different prefix, into cluster-level cache objects, enabling remote reuse and parallel prefill. Distributed caching systems had reused key-value cache only under strict prefix matching, while single-node PIC could not draw on other nodes' cache or compute. Across four models and four real RAG and agentic workloads, Nebula cuts median time-to-first-token by an average of 61.4% against the strongest baseline and recovers an average of 76.1% of the tokens missed by prefix caching.

For retrieval-augmented generation and agent workloads, Nebula promotes position-independent caching (PIC) segments, which let the same text segment reuse already-computed key-value cache under a different prefix, into cluster-level cache objects, enabling remote reuse and parallel prefill. Distributed caching systems had reused key-value cache only under strict prefix matching, while single-node PIC could not draw on other nodes' cache or compute. Across four models and four real RAG and agentic workloads, Nebula cuts median time-to-first-token by an average of 61.4% against the strongest baseline and recovers an average of 76.1% of the tokens missed by prefix caching.

ZoomGain is a data-construction pipeline that compares a full-image description of a region with a zoomed-in description of the same region and rewrites details visible only after zooming into questions about the whole image, verifying answers on the zoomed view with several models and discarding items answerable from language priors alone. On the resulting 1,000-question bilingual ZoomGainBench, the best of 28 vision-language models, GPT-5.6-Sol, scores only 61.08; after GRPO reinforcement learning on just 9K ZoomGain samples, Qwen3.5-9B averages 66.10 across 13 held-out benchmarks, above Qwen3.5-397B-A17B at 65.55.

ZoomGain is a data-construction pipeline that compares a full-image description of a region with a zoomed-in description of the same region and rewrites details visible only after zooming into questions about the whole image, verifying answers on the zoomed view with several models and discarding items answerable from language priors alone. On the resulting 1,000-question bilingual ZoomGainBench, the best of 28 vision-language models, GPT-5.6-Sol, scores only 61.08; after GRPO reinforcement learning on just 9K ZoomGain samples, Qwen3.5-9B averages 66.10 across 13 held-out benchmarks, above Qwen3.5-397B-A17B at 65.55.

Continually updated models keep a sequence of queryable snapshots, so an attacker can infer not only whether a sample was used in training but when it entered and left. TMIA treats a sample's confidence across snapshots as a time series, labels entry and exit points with a sliding-window bidirectional LSTM, and pairs them chronologically into valid membership intervals. Across three language models and three image classifiers, TMIA reaches timestamp-level F1 above 83% in every language-model setting and improves interval-level F1 over both baselines by at least 21.82%.

Continually updated models keep a sequence of queryable snapshots, so an attacker can infer not only whether a sample was used in training but when it entered and left. TMIA treats a sample's confidence across snapshots as a time series, labels entry and exit points with a sliding-window bidirectional LSTM, and pairs them chronologically into valid membership intervals. Across three language models and three image classifiers, TMIA reaches timestamp-level F1 above 83% in every language-model setting and improves interval-level F1 over both baselines by at least 21.82%.

Long-horizon language-model agents accumulate ever-longer interaction histories, so context compaction is necessary, yet existing learned compaction methods often depend on downstream task outcomes or solver behavior. SumOne is a standalone 2-billion-parameter summarizer trained with a task-agnostic reward built from masked reconstruction and optimized with GRPO, using only raw trajectories and no reference summaries or downstream task execution. On DeepSearchQA it cuts cumulative agent input from 2.83M to 591K tokens with F1 0.48 versus 0.52 for full context; on SWE-bench Verified it reaches a 0.59 resolved rate versus 0.63, with input falling from 4.10M to 1.09M tokens.

Long-horizon language-model agents accumulate ever-longer interaction histories, so context compaction is necessary, yet existing learned compaction methods often depend on downstream task outcomes or solver behavior. SumOne is a standalone 2-billion-parameter summarizer trained with a task-agnostic reward built from masked reconstruction and optimized with GRPO, using only raw trajectories and no reference summaries or downstream task execution. On DeepSearchQA it cuts cumulative agent input from 2.83M to 591K tokens with F1 0.48 versus 0.52 for full context; on SWE-bench Verified it reaches a 0.59 resolved rate versus 0.63, with input falling from 4.10M to 1.09M tokens.

For GUI agents, the safety signal may lie not in the screen but in the transition that produced it: the same credential screen can be a legitimate result of tapping Sign in or a hijacked result of tapping View order. LGWM skips rendering the next screen and predicts its representation directly in the observation encoder's space, reducing verification to a vector comparison; on RSWT-Bench, where each credential screen appears pixel-identically under both labels, the training-free consistency score reaches 0.987 AUC at 17 ms per decision, on par with the strongest closed-source vision-language models and about ten AUC points above generative GUI world models.

For GUI agents, the safety signal may lie not in the screen but in the transition that produced it: the same credential screen can be a legitimate result of tapping Sign in or a hijacked result of tapping View order. LGWM skips rendering the next screen and predicts its representation directly in the observation encoder's space, reducing verification to a vector comparison; on RSWT-Bench, where each credential screen appears pixel-identically under both labels, the training-free consistency score reaches 0.987 AUC at 17 ms per decision, on par with the strongest closed-source vision-language models and about ten AUC points above generative GUI world models.

A deep eutectic solvent is a multicomponent hydrogen-bond network of acceptors, donors and solutes, so its solubility cannot be read off the properties of individual constituents. TiNN represents the process in two stages, first forming the acceptor-donor environment and then accommodating the solute, and constrains the temperature and pressure response through a learned free-energy pathway, reaching a test-set $R^2$ of 0.917 on 938 records. In new measurements of glucose, xylan and Kraft lignin in acidic, basic and neutral deep eutectic solvents, TiNN reduces the geometric mean fold error of solubility from more than 23 for COSMO-RS to about 1.5.

A deep eutectic solvent is a multicomponent hydrogen-bond network of acceptors, donors and solutes, so its solubility cannot be read off the properties of individual constituents. TiNN represents the process in two stages, first forming the acceptor-donor environment and then accommodating the solute, and constrains the temperature and pressure response through a learned free-energy pathway, reaching a test-set $R^2$ of 0.917 on 938 records. In new measurements of glucose, xylan and Kraft lignin in acidic, basic and neutral deep eutectic solvents, TiNN reduces the geometric mean fold error of solubility from more than 23 for COSMO-RS to about 1.5.

Optimizing register-transfer-level (RTL) code has long relied on manual rewriting, and existing evaluations mostly measure gains on manually degraded designs under weak synthesis flows, so they cannot show whether a rewrite beats a commercial synthesis tool. Dr. RTL closes a loop of timing analysis, parallel rewriting, commercial synthesis and sequential equivalence checking, and distills successful and failed rewrites into reusable pattern-strategy skills. On 20 original human-written RTL designs it improves worst negative slack (WNS) by 21.3% and total negative slack (TNS) by 16.9% on average over commercial synthesis, with a 5.8% area reduction and an average sequential equivalence check pass rate of 86%.

Optimizing register-transfer-level (RTL) code has long relied on manual rewriting, and existing evaluations mostly measure gains on manually degraded designs under weak synthesis flows, so they cannot show whether a rewrite beats a commercial synthesis tool. Dr. RTL closes a loop of timing analysis, parallel rewriting, commercial synthesis and sequential equivalence checking, and distills successful and failed rewrites into reusable pattern-strategy skills. On 20 original human-written RTL designs it improves worst negative slack (WNS) by 21.3% and total negative slack (TNS) by 16.9% on average over commercial synthesis, with a 5.8% area reduction and an average sequential equivalence check pass rate of 86%.

Most ICU sepsis warning models answer only whether sepsis will occur, leaving clinicians unsure of the onset context and early course once an alarm sounds. SepsisFlow places hourly pre-onset risk, the patient state at onset and the 72-hour organ-support trajectory on one clock advancing hourly from ICU admission: before onset it uses only bedside-visible records without antimicrobial or culture orders, reaching a 24-hour AUROC of 0.822 on the MIMIC-IV internal test set, and its onset states form three contexts with 28-day in-hospital mortality of 17.1%, 15.1% and 2.2%.

Most ICU sepsis warning models answer only whether sepsis will occur, leaving clinicians unsure of the onset context and early course once an alarm sounds. SepsisFlow places hourly pre-onset risk, the patient state at onset and the 72-hour organ-support trajectory on one clock advancing hourly from ICU admission: before onset it uses only bedside-visible records without antimicrobial or culture orders, reaching a 24-hour AUROC of 0.822 on the MIMIC-IV internal test set, and its onset states form three contexts with 28-day in-hospital mortality of 17.1%, 15.1% and 2.2%.

MEOW converts non-interactive 3D segmentation backbones into open-world interactive segmentation systems without updating any learnable parameters. Interactive segmentation had been confined to the open-set setting, able to predict or reject unknown structures but not to accumulate those predictions. MEOW pairs a frozen feature extractor with parameter-free transductive reasoning and memory consolidation, storing user-confirmed pseudo-labels in memory for later queries to retrieve; on classes never seen in training it keeps improving its own predictions as memory grows, and it transfers zero-shot to micro-CT and other imaging modalities.

MEOW converts non-interactive 3D segmentation backbones into open-world interactive segmentation systems without updating any learnable parameters. Interactive segmentation had been confined to the open-set setting, able to predict or reject unknown structures but not to accumulate those predictions. MEOW pairs a frozen feature extractor with parameter-free transductive reasoning and memory consolidation, storing user-confirmed pseudo-labels in memory for later queries to retrieve; on classes never seen in training it keeps improving its own predictions as memory grows, and it transfers zero-shot to micro-CT and other imaging modalities.

Training neural operators needs expensive numerical solution labels, while PDE inputs and governing equations are cheap physical information. PI-Init spends that information on initialization: it pretrains on unlabeled inputs with physics losses, then keeps the parameters for supervised fine-tuning. At matched label budgets, FNO prediction error falls 53.3% on Burgers and 62.7% on 1D Advection, and on Burgers 200 labeled trajectories beat supervised-only training with 2,000.

Training neural operators needs expensive numerical solution labels, while PDE inputs and governing equations are cheap physical information. PI-Init spends that information on initialization: it pretrains on unlabeled inputs with physics losses, then keeps the parameters for supervised fine-tuning. At matched label budgets, FNO prediction error falls 53.3% on Burgers and 62.7% on 1D Advection, and on Burgers 200 labeled trajectories beat supervised-only training with 2,000.

RIVER is a training-free multi-agent retrieval framework that organizes large language models, biochemical computational tools and external knowledge resources into traceable evidence chains for bidirectional enzyme-to-reaction and reaction-to-enzyme retrieval. In four experimentally supported enzyme–reaction cases drawn from an E. coli metabolomics study, it ranked among 667 candidate proteins with a mean reciprocal rank of 0.833 using Qwen3.6-35B-A3B and 0.750 using GPT-5.6 Luna, above the general-purpose coding-agent combinations and domain methods listed in the comparison table. On public benchmarks it reached Top-1 query success rates of 82.3% on Enzyme-405 and 72.8% on Orphan-335.

RIVER is a training-free multi-agent retrieval framework that organizes large language models, biochemical computational tools and external knowledge resources into traceable evidence chains for bidirectional enzyme-to-reaction and reaction-to-enzyme retrieval. In four experimentally supported enzyme–reaction cases drawn from an E. coli metabolomics study, it ranked among 667 candidate proteins with a mean reciprocal rank of 0.833 using Qwen3.6-35B-A3B and 0.750 using GPT-5.6 Luna, above the general-purpose coding-agent combinations and domain methods listed in the comparison table. On public benchmarks it reached Top-1 query success rates of 82.3% on Enzyme-405 and 72.8% on Orphan-335.

Post-disaster building damage mapping is limited by labels: parameter-efficient fine-tuning adapts geospatial foundation models but still assumes large reliable annotation sets. PUMA recasts the task as positive-unlabeled learning, fine-tuning with a few confirmed damaged buildings and many unlabeled ones, while non-contrastive representation refinement and latent-state label disambiguation reinforce each other. On the xBD tsunami, hurricane, tornado and wildfire events, PUMA reaches the highest overall accuracy and F1 among the compared methods on tsunami, hurricane and wildfire, with F1 gains of 5.98 and 4.47 percentage points over the strongest competitor on tsunami and hurricane; on tornado it reaches 95.72% overall accuracy and 92.88% F1, competitive with the best NcPU.

Post-disaster building damage mapping is limited by labels: parameter-efficient fine-tuning adapts geospatial foundation models but still assumes large reliable annotation sets. PUMA recasts the task as positive-unlabeled learning, fine-tuning with a few confirmed damaged buildings and many unlabeled ones, while non-contrastive representation refinement and latent-state label disambiguation reinforce each other. On the xBD tsunami, hurricane, tornado and wildfire events, PUMA reaches the highest overall accuracy and F1 among the compared methods on tsunami, hurricane and wildfire, with F1 gains of 5.98 and 4.47 percentage points over the strongest competitor on tsunami and hurricane; on tornado it reaches 95.72% overall accuracy and 92.88% F1, competitive with the best NcPU.

EvoDiff recasts active-learning sample selection as generate-then-retrieve: diffusion sampling in feature space produces synthetic probes, and only the real unlabeled ultrasound images those probes retrieve are sent to experts for annotation. Across 15 datasets and 16,696 images spanning six organ tasks, it beat random sampling and several active-learning baselines in most settings on both ResNet-50 and the USFM foundation model, and a model-based annotation-effort metric estimated a 24% to 84% workload reduction relative to random sampling.

EvoDiff recasts active-learning sample selection as generate-then-retrieve: diffusion sampling in feature space produces synthetic probes, and only the real unlabeled ultrasound images those probes retrieve are sent to experts for annotation. Across 15 datasets and 16,696 images spanning six organ tasks, it beat random sampling and several active-learning baselines in most settings on both ResNet-50 and the USFM foundation model, and a model-based annotation-effort metric estimated a 24% to 84% workload reduction relative to random sampling.

Real-time video-to-audio has been held back by two bottlenecks: visual frontends shrink their receptive fields to cut look-ahead, and causal diffusion backends denoise each patch in isolation, so both alignment and audio quality trail offline models. LiveFoley uses a trailing sliding window to turn off-the-shelf non-causal segment-level encoders into streaming modules without retraining, cutting frontend look-ahead to 0.08 s, and widens the backend denoising window toward history. On the VGGSound test set, LiveFoley-48k reaches a DeSync of 0.48, on par with offline MMAudio-L at 0.47, with 156 ms end-to-end latency on a single A800 GPU.

Real-time video-to-audio has been held back by two bottlenecks: visual frontends shrink their receptive fields to cut look-ahead, and causal diffusion backends denoise each patch in isolation, so both alignment and audio quality trail offline models. LiveFoley uses a trailing sliding window to turn off-the-shelf non-causal segment-level encoders into streaming modules without retraining, cutting frontend look-ahead to 0.08 s, and widens the backend denoising window toward history. On the VGGSound test set, LiveFoley-48k reaches a DeSync of 0.48, on par with offline MMAudio-L at 0.47, with 156 ms end-to-end latency on a single A800 GPU.

Multimodal large language models can read text in an image accurately yet still fail to reason over the same problem as well as they do over plain text, a shortfall the authors call a modality reasoning-access gap. TIGER renders text-only math problems as images, mines cases the model solves from text but fails from the image, and post-trains on those image inputs with GRPO and related objectives. Trained only on synthetic rendered-text images, Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse, while Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA.

Multimodal large language models can read text in an image accurately yet still fail to reason over the same problem as well as they do over plain text, a shortfall the authors call a modality reasoning-access gap. TIGER renders text-only math problems as images, mines cases the model solves from text but fails from the image, and post-trains on those image inputs with GRPO and related objectives. Trained only on synthetic rendered-text images, Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse, while Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA.

For runtime attack detection in cyber-physical systems, PIA has a large language model agent write physical residuals such as pump characteristics, tank mass balances and controller logic from a plant description and normal operating data, then test each candidate on an evaluation workbench before deployment. On the BATADAL water-distribution attack benchmark's test attacks, PIA with Claude Opus 5.5 reaches an F1 of 0.878 against 0.504 for the strongest of six statistical and deep baselines; ablations show that denying the agent data testing costs up to 0.28 F1.

For runtime attack detection in cyber-physical systems, PIA has a large language model agent write physical residuals such as pump characteristics, tank mass balances and controller logic from a plant description and normal operating data, then test each candidate on an evaluation workbench before deployment. On the BATADAL water-distribution attack benchmark's test attacks, PIA with Claude Opus 5.5 reaches an F1 of 0.878 against 0.504 for the strongest of six statistical and deep baselines; ablations show that denying the agent data testing costs up to 0.28 F1.

Automated chest X-ray reports often read fluently while leaving unclear which image evidence supports which finding. ARIVE interleaves anatomical inspection with diagnostic reasoning before composing the report, trained on 159,416 multimodal chain-of-thought trajectories built from MIMIC-CXR. On MIMIC-CXR and the abnormality-only MIMIC-ABN it scored highest overall on clinical-efficacy measures among the compared models, and in blinded grading of 100 cases its report quality was close to Qwen3.5 (mean 3.47 versus 3.51, against 3.88 for reference reports).

Automated chest X-ray reports often read fluently while leaving unclear which image evidence supports which finding. ARIVE interleaves anatomical inspection with diagnostic reasoning before composing the report, trained on 159,416 multimodal chain-of-thought trajectories built from MIMIC-CXR. On MIMIC-CXR and the abnormality-only MIMIC-ABN it scored highest overall on clinical-efficacy measures among the compared models, and in blinded grading of 100 cases its report quality was close to Qwen3.5 (mean 3.47 versus 3.51, against 3.88 for reference reports).

Robot-arm inverse kinematics is usually handled by a numerical solver plus a hand-tuned controller; PolicyIK replaces that stage with a goal-conditioned reinforcement-learning policy that updates joint references from pose feedback through a joint-position interface. On a simulated 7-DoF OpenArm, the trained PPO policy keeps the mean error over the final second within 1 mm on 93.5% of held-out tasks in a single attempt, and a deployment protocol raises success to 98.9%; trained directly on a physical arm, it reaches and holds more tasks than four classical IK controllers.

Robot-arm inverse kinematics is usually handled by a numerical solver plus a hand-tuned controller; PolicyIK replaces that stage with a goal-conditioned reinforcement-learning policy that updates joint references from pose feedback through a joint-position interface. On a simulated 7-DoF OpenArm, the trained PPO policy keeps the mean error over the final second within 1 mm on 93.5% of held-out tasks in a single attempt, and a deployment protocol raises success to 98.9%; trained directly on a physical arm, it reaches and holds more tasks than four classical IK controllers.

Radar 3D perception has long compressed raw spectra into sparse point clouds or compact Doppler statistics before supervised training. L-SPMAE instead uses masked self-supervision: whatever remains reconstructable under localized corruption counts as signal, and spectral leakage supplies that corruption, so pretraining runs directly on unreduced radar spectra. On K-Radar, L-SPMAE with the RAD representation reaches 55.9/50.1 mAP after fine-tuning, above the best point-cloud self-supervised baseline in the table at 50.8/45.0.

Radar 3D perception has long compressed raw spectra into sparse point clouds or compact Doppler statistics before supervised training. L-SPMAE instead uses masked self-supervision: whatever remains reconstructable under localized corruption counts as signal, and spectral leakage supplies that corruption, so pretraining runs directly on unreduced radar spectra. On K-Radar, L-SPMAE with the RAD representation reaches 55.9/50.1 mAP after fine-tuning, above the best point-cloud self-supervised baseline in the table at 50.8/45.0.

GraphRAG links information across documents through entity relations, but black-box queries can also be used to reconstruct the knowledge graph, leaking whether entities are connected and what those relations mean. RelGuard rewrites source documents before indexing: it generates candidate relations from entity types, then uses an exponential mechanism to sample a replacement predicate or delete each relation, with a relation-level differential privacy guarantee. Across LightRAG, PathRAG and HippoRAG 2, two datasets and two reconstruction attacks, its relation-recovery F1 is 26.98 to 44.35 percentage points below the strongest baseline, while the extra question-answering F1 loss is 0.59 to 4.78 points in most settings.

GraphRAG links information across documents through entity relations, but black-box queries can also be used to reconstruct the knowledge graph, leaking whether entities are connected and what those relations mean. RelGuard rewrites source documents before indexing: it generates candidate relations from entity types, then uses an exponential mechanism to sample a replacement predicate or delete each relation, with a relation-level differential privacy guarantee. Across LightRAG, PathRAG and HippoRAG 2, two datasets and two reconstruction attacks, its relation-recovery F1 is 26.98 to 44.35 percentage points below the strongest baseline, while the extra question-answering F1 loss is 0.59 to 4.78 points in most settings.

Audio language model benchmark scores can mistake language ability for acoustic understanding. Using synthetic readings and transcripts as references, the study measures the recording gain of the original audio over words across 92 tasks from 6 benchmarks, finding that 19 open models trail Gemini 3.1 Pro by 28.6 and 41.9 percentage points on global-attribute and within-recording structure tasks, with about half or more of each gap lying in the recording gain. Training LoRA adapters with a content anchor in the readout-decline layers of Gemma-4 12B and Qwen2.5-Omni-7B raises attribute and structure accuracy by 12.8 to 44.4 points while mean lexical accuracy stays within 1 point of the base models.

Audio language model benchmark scores can mistake language ability for acoustic understanding. Using synthetic readings and transcripts as references, the study measures the recording gain of the original audio over words across 92 tasks from 6 benchmarks, finding that 19 open models trail Gemini 3.1 Pro by 28.6 and 41.9 percentage points on global-attribute and within-recording structure tasks, with about half or more of each gap lying in the recording gain. Training LoRA adapters with a content anchor in the readout-decline layers of Gemma-4 12B and Qwen2.5-Omni-7B raises attribute and structure accuracy by 12.8 to 44.4 points while mean lexical accuracy stays within 1 point of the base models.

PatternBloom changes the retrieval unit of agentic retrieval-augmented generation from what content to read to how to reason, using typed graph skeletons distilled from high-scoring Stage-I reinforcement-learning trajectories into an external Graph Pattern Memory, then coupling the policy to that frozen memory with a pattern-coverage reward. Across seven QA benchmarks, the 7B version ranks first on EM and F1 on all seven datasets, leads R1-Searcher on average out-of-distribution EM, and the same frozen pattern memory transfers to a Qwen3-8B backbone.

PatternBloom changes the retrieval unit of agentic retrieval-augmented generation from what content to read to how to reason, using typed graph skeletons distilled from high-scoring Stage-I reinforcement-learning trajectories into an external Graph Pattern Memory, then coupling the policy to that frozen memory with a pattern-coverage reward. Across seven QA benchmarks, the 7B version ranks first on EM and F1 on all seven datasets, leads R1-Searcher on average out-of-distribution EM, and the same frozen pattern memory transfers to a Qwen3-8B backbone.

For ℓ0 perturbations that change only a few pixels, FocusBranch rewrites the “at most k perturbed pixels” constraint as a node made of a focus pixel set and a level, so verification can branch across levels. It builds super nodes through covering designs and concrete nodes through upper shadows, and picks a starting level from sampling estimates. Across 13 networks and 3,189 candidate instances it verified 2,799; in a 740-instance comparison it verified 641, against 375 for CoVerD+top-k-GP.

For ℓ0 perturbations that change only a few pixels, FocusBranch rewrites the “at most k perturbed pixels” constraint as a node made of a focus pixel set and a level, so verification can branch across levels. It builds super nodes through covering designs and concrete nodes through upper shadows, and picks a starting level from sampling estimates. Across 13 networks and 3,189 candidate instances it verified 2,799; in a 740-instance comparison it verified 641, against 375 for CoVerD+top-k-GP.

Large language models refuse the same harmful request in one language but may comply when it is rephrased in another. Researchers first identified safety-related neurons independently within each language, then took the intersection of the candidate sets from six languages to obtain a cross-lingually shared "safety kernel," and tested its functional role with activation suppression and activation-state transfer. Building on it, SK-Masked LoRA restricts fine-tuning updates to the feed-forward network output dimensions corresponding to the safety kernel, lowering attack success rates across several models and multilingual safety benchmarks, with some effect on languages excluded from kernel discovery.

Large language models refuse the same harmful request in one language but may comply when it is rephrased in another. Researchers first identified safety-related neurons independently within each language, then took the intersection of the candidate sets from six languages to obtain a cross-lingually shared "safety kernel," and tested its functional role with activation suppression and activation-state transfer. Building on it, SK-Masked LoRA restricts fine-tuning updates to the feed-forward network output dimensions corresponding to the safety kernel, lowering attack success rates across several models and multilingual safety benchmarks, with some effect on languages excluded from kernel discovery.

Protein conformational ensemble generation has treated coordinated motion between residues as implicit in the learned distribution, which limits direct control. CoMoFlow builds protein-specific collective modes from the equilibrium structure, learns the distribution of their coefficients across the molecular dynamics ensemble, and uses sampled collective displacements to guide SE(3) flow matching through feature conditioning and prior recentering. On ATLAS and de novo proteins it raises DCCM correlation over BBFlow from 0.87/0.83 to 0.93/0.89 and lowers PCA distance from 1.33/0.67 to 1.21/0.59, while allowing generated conformations to be steered along a chosen mode in direction and magnitude.

Protein conformational ensemble generation has treated coordinated motion between residues as implicit in the learned distribution, which limits direct control. CoMoFlow builds protein-specific collective modes from the equilibrium structure, learns the distribution of their coefficients across the molecular dynamics ensemble, and uses sampled collective displacements to guide SE(3) flow matching through feature conditioning and prior recentering. On ATLAS and de novo proteins it raises DCCM correlation over BBFlow from 0.87/0.83 to 0.93/0.89 and lowers PCA distance from 1.33/0.67 to 1.21/0.59, while allowing generated conformations to be steered along a chosen mode in direction and magnitude.

Graph neural networks learn structure–property relationships from graphs of atoms and bonds, but those relationships stay in a latent space and are hard to read as scientific rules. Graph symbolic distillation (GSD) first uses a constrained message-passing architecture search to fix how information is transformed, propagated and aggregated, then fits each layer's learned transformation with a sparse symbolic expression while keeping graph connectivity and aggregation. Across halide-perovskite band gaps, molecular hydration free energies and multidentate adsorption on transition-metal surfaces, it rebuilds expert descriptors from elementary node and edge features and yields testable local-geometry, polarity–size and electronic–connectivity relationships.

Graph neural networks learn structure–property relationships from graphs of atoms and bonds, but those relationships stay in a latent space and are hard to read as scientific rules. Graph symbolic distillation (GSD) first uses a constrained message-passing architecture search to fix how information is transformed, propagated and aggregated, then fits each layer's learned transformation with a sparse symbolic expression while keeping graph connectivity and aggregation. Across halide-perovskite band gaps, molecular hydration free energies and multidentate adsorption on transition-metal surfaces, it rebuilds expert descriptors from elementary node and edge features and yields testable local-geometry, polarity–size and electronic–connectivity relationships.

Probabilistic PDE surrogates often turn overconfident in few-step inference, producing predictive spreads narrower than the real uncertainty. TORNADO recasts next-step prediction as stochastic transport between the variational-autoencoder posteriors of consecutive frames, and lets a neural network set the diffusion amplitude from the latent state instead of a fixed global noise schedule. On four synthetic PDE datasets, at a fixed budget of five network evaluations, it reaches the lowest calibration error, roughly 2x to over 5x lower than representative baselines, while keeping comparable predictive accuracy.

Probabilistic PDE surrogates often turn overconfident in few-step inference, producing predictive spreads narrower than the real uncertainty. TORNADO recasts next-step prediction as stochastic transport between the variational-autoencoder posteriors of consecutive frames, and lets a neural network set the diffusion amplitude from the latent state instead of a fixed global noise schedule. On four synthetic PDE datasets, at a fixed budget of five network evaluations, it reaches the lowest calibration error, roughly 2x to over 5x lower than representative baselines, while keeping comparable predictive accuracy.

Because an inverse map cannot be uniformly approximated by a continuous network near a discontinuity, the work learns a differentiable forward operator and its derivative and uses them as the surrogate inside Landweber iteration. On reaction-coefficient identification in an elliptic equation, four surrogates reconstruct both an in-distribution truth and one with sharp inclusions, but without preconditioning three of them semiconverge early and develop high-frequency spurious structure even on exact data; derivative supervision along fixed directions barely changes the best reconstruction.

Because an inverse map cannot be uniformly approximated by a continuous network near a discontinuity, the work learns a differentiable forward operator and its derivative and uses them as the surrogate inside Landweber iteration. On reaction-coefficient identification in an elliptic equation, four surrogates reconstruct both an in-distribution truth and one with sharp inclusions, but without preconditioning three of them semiconverge early and develop high-frequency spurious structure even on exact data; derivative supervision along fixed directions barely changes the best reconstruction.

A single triple-axis neutron measurement draws on papers, instrument logs, data files and specialist software, and that evidence was previously scattered across separate stages with no traceable link. TASCopilot records sources, processing steps and human review decisions as an inspectable evidence chain, while language models only extract information and explain results; scientific calculations, validation rules and human review decide whether evidence can support a claim. On the BT-7 spectrometer at the NIST Center for Neutron Research it recovered a sample rotation deliberately parked 40° away to within 0.08° of the orientation-matrix prediction, and completed 40 adaptive measurement points with all 40 counted and no refusals.

A single triple-axis neutron measurement draws on papers, instrument logs, data files and specialist software, and that evidence was previously scattered across separate stages with no traceable link. TASCopilot records sources, processing steps and human review decisions as an inspectable evidence chain, while language models only extract information and explain results; scientific calculations, validation rules and human review decide whether evidence can support a claim. On the BT-7 spectrometer at the NIST Center for Neutron Research it recovered a sample rotation deliberately parked 40° away to within 0.08° of the orientation-matrix prediction, and completed 40 adaptive measurement points with all 40 counted and no refusals.

Standard neural processes treat inputs as noise-free observations, so sensor measurement error degrades both predictions and uncertainty estimates. DnNP makes the clean input a latent variable inside the neural-process framework and trains a denoiser jointly with meta-learning through variational inference, predicting directly from noisy inputs. On PM2.5 spatial interpolation from 34 air-quality sensors in Brookline, Massachusetts, DnNP reduces RMSE by up to 30% and NLL by up to 73% relative to standard neural processes.

Standard neural processes treat inputs as noise-free observations, so sensor measurement error degrades both predictions and uncertainty estimates. DnNP makes the clean input a latent variable inside the neural-process framework and trains a denoiser jointly with meta-learning through variational inference, predicting directly from noisy inputs. On PM2.5 spatial interpolation from 34 air-quality sensors in Brookline, Massachusetts, DnNP reduces RMSE by up to 30% and NLL by up to 73% relative to standard neural processes.

In battery arbitrage, when a battery starts and ends the day at the same charge, a common shift in prices affects its schedule only through round-trip losses, so decision-focused losses identify the forecast level only weakly; clipping forecasts to a price cap can lower prediction error while removing profitable spreads. On 1,158 holdout days of the Indian day-ahead market, an equal-weight average of three statistical forecasts did better than the five decision-focused variants tested, and on 53 later heavily capped days no variant beat the regression or the average.

In battery arbitrage, when a battery starts and ends the day at the same charge, a common shift in prices affects its schedule only through round-trip losses, so decision-focused losses identify the forecast level only weakly; clipping forecasts to a price cap can lower prediction error while removing profitable spreads. On 1,158 holdout days of the Indian day-ahead market, an equal-weight average of three statistical forecasts did better than the five decision-focused variants tested, and on 53 later heavily capped days no variant beat the regression or the average.

Muse Spark Open decoding rises from a 517 tok/s baseline to 1,109 tok/s by fusing each layer's attention and MoE into a single kernel that stays resident on the GPU, overlapping weight loads, tensor-core work and cross-GPU communication. Previously SGLang launched about 1,450 kernels per decoding step, and the GPU drained and refilled between them, leaving some SMs and threads idle. The figure was measured on Blackwell GPUs at batch size one with DFlash V2 speculative decoding and has not yet been extended to larger batches.

Muse Spark Open decoding rises from a 517 tok/s baseline to 1,109 tok/s by fusing each layer's attention and MoE into a single kernel that stays resident on the GPU, overlapping weight loads, tensor-core work and cross-GPU communication. Previously SGLang launched about 1,450 kernels per decoding step, and the GPU drained and refilled between them, leaving some SMs and threads idle. The figure was measured on Blackwell GPUs at batch size one with DFlash V2 speculative decoding and has not yet been extended to larger batches.

Low-rank adaptation (LoRA) writes the weight update as a product of two small matrices, but their norms often differ sharply during training, producing imbalanced gradient flow and unstable optimization. SingLoRA instead writes the update as a single low-rank matrix times its own transpose, removing inter-matrix scale conflicts by construction and roughly halving the trainable parameters. Fine-tuning LLaMA-7B on MNLI, SingLoRA reaches 91.3% accuracy against 89.1% for LoRA and 90.2% for LoRA+, while using only 60% of their parameter budget.

Low-rank adaptation (LoRA) writes the weight update as a product of two small matrices, but their norms often differ sharply during training, producing imbalanced gradient flow and unstable optimization. SingLoRA instead writes the update as a single low-rank matrix times its own transpose, removing inter-matrix scale conflicts by construction and roughly halving the trainable parameters. Fine-tuning LLaMA-7B on MNLI, SingLoRA reaches 91.3% accuracy against 89.1% for LoRA and 90.2% for LoRA+, while using only 60% of their parameter budget.

Localizing pathology in 3D CT normally requires experts to outline lesions voxel by voxel, which is costly and hard to extend across findings. Uni-Loc replaces the classification head with similarity scores between a pathology-conditioned query and spatial image features, so the same scores yield both a scan-level diagnosis and a 3D localization map, with gradient-guided thresholding outlining lesion regions. Across five localization tasks covering 355 annotated CT scans, it reaches macro-averaged voxel-level AUROC/AUPR of 96.1%/34.5%, against 80.7%/13.9% for relevance propagation and 88.7%/8.1% for gradient saliency.

Localizing pathology in 3D CT normally requires experts to outline lesions voxel by voxel, which is costly and hard to extend across findings. Uni-Loc replaces the classification head with similarity scores between a pathology-conditioned query and spatial image features, so the same scores yield both a scan-level diagnosis and a 3D localization map, with gradient-guided thresholding outlining lesion regions. Across five localization tasks covering 355 annotated CT scans, it reaches macro-averaged voxel-level AUROC/AUPR of 96.1%/34.5%, against 80.7%/13.9% for relevance propagation and 88.7%/8.1% for gradient saliency.

HIV contact networks are prohibitively expensive to collect at scale, while peer referral networks can be gathered digitally but deviate from the true contact structure because referrals follow non-risk social ties and each participant accepts only one referral. R2C-Agent treats the referral process as incremental corruption of the contact network and learns to reverse it, then plans interventions over sampled contact networks. On real ICPSR HIV network data, its total regret is 54.9% lower on intensive treatment and 28.8% lower on targeted PrEP than the best baselines.

HIV contact networks are prohibitively expensive to collect at scale, while peer referral networks can be gathered digitally but deviate from the true contact structure because referrals follow non-risk social ties and each participant accepts only one referral. R2C-Agent treats the referral process as incremental corruption of the contact network and learns to reverse it, then plans interventions over sampled contact networks. On real ICPSR HIV network data, its total regret is 54.9% lower on intensive treatment and 28.8% lower on targeted PrEP than the best baselines.

An Apple team aligned wrist photoplethysmography (PPG, a green-light measure of microvascular blood-volume change) with meal descriptions to train the Nutrition PPG Language Model (NPLM) and test whether passively collected pulse waveforms carry dietary information. On 1,122,834 meal descriptions from 19,340 participants, near-meal PPG raised the likelihood of the correct meal text to a win rate of 0.69, against 0.60 for a control that kept the same participant and time of day but broke the meal pairing, and 0.50 for across-subject shuffling. Without retraining, the signal raised postprandial satiety prediction AUC from 0.64 to 0.71 in an independent cohort, which the authors frame as augmenting text-based logging rather than replacing it.

An Apple team aligned wrist photoplethysmography (PPG, a green-light measure of microvascular blood-volume change) with meal descriptions to train the Nutrition PPG Language Model (NPLM) and test whether passively collected pulse waveforms carry dietary information. On 1,122,834 meal descriptions from 19,340 participants, near-meal PPG raised the likelihood of the correct meal text to a win rate of 0.69, against 0.60 for a control that kept the same participant and time of day but broke the meal pairing, and 0.50 for across-subject shuffling. Without retraining, the signal raised postprandial satiety prediction AUC from 0.64 to 0.71 in an independent cohort, which the authors frame as augmenting text-based logging rather than replacing it.

Ai2 released ToyBlocks, an open simulation framework built around a robot arm and blocks on a tabletop and the first of its FoundationWorlds. It automatically generates tasks and demonstrations and reports metrics beyond success rates, such as execution speed, forces exerted and task progress, so researchers can change training data or model design, repeat a test and trace a failure. In zero-shot real-robot tests, ManiFlow's atomic-skill success rates were 80.0% in simulation and 81.1% in reality.

Ai2 released ToyBlocks, an open simulation framework built around a robot arm and blocks on a tabletop and the first of its FoundationWorlds. It automatically generates tasks and demonstrations and reports metrics beyond success rates, such as execution speed, forces exerted and task progress, so researchers can change training data or model design, repeat a test and trace a failure. In zero-shot real-robot tests, ManiFlow's atomic-skill success rates were 80.0% in simulation and 81.1% in reality.

NVIDIA released Alpamayo 2 Edge, an edge-oriented reasoning vision-language-action model that pairs a 2B-parameter Nemotron reasoning backbone with a diffusion action expert to produce Chain-of-Causation reasoning and future trajectories on DRIVE AGX Thor, generating reasoning only when needed. Closed-loop reinforcement learning post-training in AlpaGym improves closed-loop driving performance by 48.9% over the best-performing checkpoint without closed-loop RL, and the optimized model reaches 160–190 ms pixel-to-trajectory inference latency depending on the amount of reasoning.

NVIDIA released Alpamayo 2 Edge, an edge-oriented reasoning vision-language-action model that pairs a 2B-parameter Nemotron reasoning backbone with a diffusion action expert to produce Chain-of-Causation reasoning and future trajectories on DRIVE AGX Thor, generating reasoning only when needed. Closed-loop reinforcement learning post-training in AlpaGym improves closed-loop driving performance by 48.9% over the best-performing checkpoint without closed-loop RL, and the optimized model reaches 160–190 ms pixel-to-trajectory inference latency depending on the amount of reasoning.

Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.

Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.

EarthRecord hands multimodal agents the raw 1984–2026 Landsat, MODIS, VIIRS and Sentinel-2 archives of 56 sites, leaving cloud masking and cross-sensor harmonization to the agent, and poses 238 land-change questions. Among twelve models, the four leaders score 81.0% to 85.4% pass3 on the 226 computable and recognition questions; six leading models reach 92.0% on questions computed from record variables but 61.1% on those needing a map from imagery. In instrument pairs, 53.5% of previously right answers follow the altered satellite data into error.

EarthRecord hands multimodal agents the raw 1984–2026 Landsat, MODIS, VIIRS and Sentinel-2 archives of 56 sites, leaving cloud masking and cross-sensor harmonization to the agent, and poses 238 land-change questions. Among twelve models, the four leaders score 81.0% to 85.4% pass3 on the 226 computable and recognition questions; six leading models reach 92.0% on questions computed from record variables but 61.1% on those needing a map from imagery. In instrument pairs, 53.5% of previously right answers follow the altered satellite data into error.

Benchmarking single-cell methods has been limited by costly annotation of real data and by datasets that capture only one observed condition. scDeepSim encodes expression into latent subvectors assigned to cell type, batch and residual variation, learns their distribution with latent diffusion, and at generation time modifies only the target subvector, imposing batch effects or constructing branching trajectories at user-set strengths. On scIB human pancreas data at intervention strength α=1, the structured model reached a mean cell-type silhouette width of 0.519 versus 0.432 for a plain variational autoencoder.

Benchmarking single-cell methods has been limited by costly annotation of real data and by datasets that capture only one observed condition. scDeepSim encodes expression into latent subvectors assigned to cell type, batch and residual variation, learns their distribution with latent diffusion, and at generation time modifies only the target subvector, imposing batch effects or constructing branching trajectories at user-set strengths. On scIB human pancreas data at intervention strength α=1, the structured model reached a mean cell-type silhouette width of 0.519 versus 0.432 for a plain variational autoencoder.

An MIT team has developed a training-data optimization method that lets a climate emulator trained only on carefully selected synthetic scenarios predict outcomes of scenarios it never saw, while the traditional emulators evaluated failed to adapt to new situations. The method reduces the number of full earth-system model runs needed to generate training data from six to eight down to one, shortening production time for these fast surrogate models. Emulators trained this way also captured a local pollutant-control case in which one pollutant falls while other emissions keep rising.

An MIT team has developed a training-data optimization method that lets a climate emulator trained only on carefully selected synthetic scenarios predict outcomes of scenarios it never saw, while the traditional emulators evaluated failed to adapt to new situations. The method reduces the number of full earth-system model runs needed to generate training data from six to eight down to one, shortening production time for these fast surrogate models. Emulators trained this way also captured a local pollutant-control case in which one pollutant falls while other emissions keep rising.

P-glycoprotein efflux limits small-molecule bioavailability and central-nervous-system exposure. SF-GA embeds fragment synergy analysis in the mutation step of a genetic algorithm, preferentially swapping in substructures that co-occur with the non-substrate class, so predicted P-gp substrates can be rewritten into predicted non-substrates while the scaffold is retained. On the PMO benchmark under a fixed function-evaluation budget, SF-GA's optimization performance is comparable to GB-GA, REINVENT and FragGT, and it reaches the highest mean RAscore (0.932) among all methods tested. The molecules remain hypothetical computational predictions, and non-substrate behavior has not been experimentally validated.

P-glycoprotein efflux limits small-molecule bioavailability and central-nervous-system exposure. SF-GA embeds fragment synergy analysis in the mutation step of a genetic algorithm, preferentially swapping in substructures that co-occur with the non-substrate class, so predicted P-gp substrates can be rewritten into predicted non-substrates while the scaffold is retained. On the PMO benchmark under a fixed function-evaluation budget, SF-GA's optimization performance is comparable to GB-GA, REINVENT and FragGT, and it reaches the highest mean RAscore (0.932) among all methods tested. The molecules remain hypothetical computational predictions, and non-substrate behavior has not been experimentally validated.

In a heterogeneous medium, making a wave field take a prescribed shape at a set time means working backward to an initial pressure distribution, and previously every target and every medium required a fresh numerical inversion. JAX-WAVE trains a shared neural network to map the desired terminal field and a known wave-speed map directly to a bounded initial pressure, with training signal supplied by a differentiable finite-difference time-domain solver and no pre-optimized control labels. On an independent confirmation set, three trained models reach 25.24% error on unseen geometric targets and 4.77% on unseen reachable targets; within a 5 ms median online budget the learned map cuts mean error by 45% relative to accelerated numerical inversion on the heterogeneous benchmark.

In a heterogeneous medium, making a wave field take a prescribed shape at a set time means working backward to an initial pressure distribution, and previously every target and every medium required a fresh numerical inversion. JAX-WAVE trains a shared neural network to map the desired terminal field and a known wave-speed map directly to a bounded initial pressure, with training signal supplied by a differentiable finite-difference time-domain solver and no pre-optimized control labels. On an independent confirmation set, three trained models reach 25.24% error on unseen geometric targets and 4.77% on unseen reachable targets; within a 5 ms median online budget the learned map cuts mean error by 45% relative to accelerated numerical inversion on the heterogeneous benchmark.

On severe tuberculosis CT, cavities and consolidations have soft-tissue density and are left out of lung masks by segmenters such as R231. LungMend treats mask correction as shape completion: a latent diffusion prior trained on 57 healthy lungs, conditioned on a synthetically corrupted mask and on bone anatomy that disease does not alter, with no pathological data or CT intensities. On 200 radiologist-corrected slices from 20 severe TB cases it reaches 0.986 Dice and 0.57 mm average surface distance, with cavity and consolidation recall of 0.993 and 0.981.

On severe tuberculosis CT, cavities and consolidations have soft-tissue density and are left out of lung masks by segmenters such as R231. LungMend treats mask correction as shape completion: a latent diffusion prior trained on 57 healthy lungs, conditioned on a synthetically corrupted mask and on bone anatomy that disease does not alter, with no pathological data or CT intensities. On 200 radiologist-corrected slices from 20 severe TB cases it reaches 0.986 Dice and 0.57 mm average surface distance, with cavity and consolidation recall of 0.993 and 0.981.

When path integral diffusion (PID) generates a finite Ising spin glass from a point source, each trajectory first chooses one of the two mirror halves of the zero-field measure, then decides among glassy classes inside that half. Theory shows every isotropic PID is Gaussian stochastic localization seen on a different clock, so the protocol sets when decisions happen and what they cost, not what is decided; on exactly enumerable Bethe and SK glasses, nested glassy levels are often settled in one or two distinct events, and these classes appear near the glass transition, keep their identity on cooling, and are absent in a ferromagnet on the same graphs.

When path integral diffusion (PID) generates a finite Ising spin glass from a point source, each trajectory first chooses one of the two mirror halves of the zero-field measure, then decides among glassy classes inside that half. Theory shows every isotropic PID is Gaussian stochastic localization seen on a different clock, so the protocol sets when decisions happen and what they cost, not what is decided; on exactly enumerable Bethe and SK glasses, nested glassy levels are often settled in one or two distinct events, and these classes appear near the glass transition, keep their identity on cooling, and are absent in a ferromagnet on the same graphs.

Drex turns choosing into a single forward pass: given a context, a question and a set of options, the model returns a probability for every option instead of parsing text written by a large model. Drex Max, the largest of four open decision models, scores 67.52 on the public Decision Index 0.4.1 suite against 60.70 for Jev 1.14.0, and drops less than a point when private tests are added. On Decision Bench v4 with raw image input it is statistically tied with d3-turbo and GPT-6 Sol, answering a typical request in 183 ms.

Drex turns choosing into a single forward pass: given a context, a question and a set of options, the model returns a probability for every option instead of parsing text written by a large model. Drex Max, the largest of four open decision models, scores 67.52 on the public Decision Index 0.4.1 suite against 60.70 for Jev 1.14.0, and drops less than a point when private tests are added. On Decision Bench v4 with raw image input it is statistically tied with d3-turbo and GPT-6 Sol, answering a typical request in 183 ms.

Frozen pathology foundation-model embeddings transfer poorly to single-cell spatial gene expression prediction from H&E: on the 5K gene panel, mean per-gene Pearson correlation falls from 0.375 at pseudo-spot resolution to 0.065 at single-cell resolution. A set of single-cell-oriented lightweight adaptations, namely cell-localized token pooling, 40x field-of-view inputs, Poisson count supervision and LoRA fine-tuning, raises the same panel to 0.174 and beats SpatialEx on the 5K panel with AUROC 0.800 versus 0.642. Gains concentrate in spatially variable and disease-relevant genes, while absolute correlation across the full panel stays low.

Frozen pathology foundation-model embeddings transfer poorly to single-cell spatial gene expression prediction from H&E: on the 5K gene panel, mean per-gene Pearson correlation falls from 0.375 at pseudo-spot resolution to 0.065 at single-cell resolution. A set of single-cell-oriented lightweight adaptations, namely cell-localized token pooling, 40x field-of-view inputs, Poisson count supervision and LoRA fine-tuning, raises the same panel to 0.174 and beats SpatialEx on the 5K panel with AUROC 0.800 versus 0.642. Gains concentrate in spatially variable and disease-relevant genes, while absolute correlation across the full panel stays low.

Meta's Intelligent Delivery routes compute across the ads serving stack through an ML-driven adaptive optimization layer that targets the marginal return of each request, model, hardware fleet and ranking stage. From 2017 through the first half of 2026 it contributed about 10% additional ads ranking performance on top of model and system advancements. A shared Resource Core Model reads each request, layered decision engines place traffic, hardware and per-host capacity, and AI agents propose and test configuration changes that engineers approve before shipping.

Meta's Intelligent Delivery routes compute across the ads serving stack through an ML-driven adaptive optimization layer that targets the marginal return of each request, model, hardware fleet and ranking stage. From 2017 through the first half of 2026 it contributed about 10% additional ads ranking performance on top of model and system advancements. A shared Resource Core Model reads each request, layered decision engines place traffic, hardware and per-host capacity, and AI agents propose and test configuration changes that engineers approve before shipping.

In aerial drone imagery, a referred target often covers only tens of pixels and sits among visually similar vehicles or people, so a generative editor can draw a plausible mask around the wrong instance. DroneSegEdit freezes the open-weight LongCat-Image-Edit, trains only a lightweight LoRA to render candidate regions, then uses a language-grounded spatial verifier to strip foreground inside that candidate that the referring expression does not support. On the complete official RIS-LAD test set it reaches 52.78 mIoU, 7.16 points above the previous best ReSaP-8B, and stays competitive overall on UAVid-RIS and VDD-RIS, with the clearest gains on compact vehicles and humans.

In aerial drone imagery, a referred target often covers only tens of pixels and sits among visually similar vehicles or people, so a generative editor can draw a plausible mask around the wrong instance. DroneSegEdit freezes the open-weight LongCat-Image-Edit, trains only a lightweight LoRA to render candidate regions, then uses a language-grounded spatial verifier to strip foreground inside that candidate that the referring expression does not support. On the complete official RIS-LAD test set it reaches 52.78 mIoU, 7.16 points above the previous best ReSaP-8B, and stays competitive overall on UAVid-RIS and VDD-RIS, with the clearest gains on compact vehicles and humans.

The Malawi multi-disease agent-based simulation Thanzi La Onse replaces step-by-step event simulation of its road traffic injury module with a conditional tabular generative adversarial network that samples outcomes directly at the agent level. Across 50 independent runs, mean CPU runtime improves by 86.6% (95% bootstrap confidence interval [85.2, 88.1]), mean disability-adjusted life years averaged over age, sex and year carry a 13% error, and multi-disease dynamics are still recovered.

The Malawi multi-disease agent-based simulation Thanzi La Onse replaces step-by-step event simulation of its road traffic injury module with a conditional tabular generative adversarial network that samples outcomes directly at the agent level. Across 50 independent runs, mean CPU runtime improves by 86.6% (95% bootstrap confidence interval [85.2, 88.1]), mean disability-adjusted life years averaged over age, sex and year carry a 13% error, and multi-disease dynamics are still recovered.

The conversion efficiency of gas high-harmonic generation traces a hyperbolic ridge in the pressure–medium-length plane, and standard Bayesian optimization, which models correlations by Euclidean distance, struggles to reproduce that structure. Writing known physical scaling relations into nonlinear coordinate transformations of the search space lets a Gaussian-process surrogate raise structural similarity to a reference grid scan from 0.77 to 0.98 within 60 iterations in two dimensions, and test reconstruction accuracy against 1024 independent reference points in four dimensions. A threshold-based acquisition function then maps extended high-efficiency operating regions rather than a single optimum.

The conversion efficiency of gas high-harmonic generation traces a hyperbolic ridge in the pressure–medium-length plane, and standard Bayesian optimization, which models correlations by Euclidean distance, struggles to reproduce that structure. Writing known physical scaling relations into nonlinear coordinate transformations of the search space lets a Gaussian-process surrogate raise structural similarity to a reference grid scan from 0.77 to 0.98 within 60 iterations in two dimensions, and test reconstruction accuracy against 1024 independent reference points in four dimensions. A threshold-based acquisition function then maps extended high-efficiency operating regions rather than a single optimum.

A team at the University of Zurich and ETH Zurich built CURRENT, a recurrent neural network that performs spatial, temporal and spatiotemporal super-resolution of aortic 4D flow MRI within one architecture. On synthetic data its spatial super-resolution directional error was 0.072±0.037 versus 0.097±0.053 for 4DFlowNet, and its temporal directional error was 0.062±0.031 versus 0.064±0.031 for T4DFlowNet. Joint spatiotemporal super-resolution gave relative pressure estimates closer to the high-resolution reference than either direction alone, and the model, trained only on synthetic data, showed promising generalization to in vivo acquisitions.

A team at the University of Zurich and ETH Zurich built CURRENT, a recurrent neural network that performs spatial, temporal and spatiotemporal super-resolution of aortic 4D flow MRI within one architecture. On synthetic data its spatial super-resolution directional error was 0.072±0.037 versus 0.097±0.053 for 4DFlowNet, and its temporal directional error was 0.062±0.031 versus 0.064±0.031 for T4DFlowNet. Joint spatiotemporal super-resolution gave relative pressure estimates closer to the high-resolution reference than either direction alone, and the model, trained only on synthetic data, showed promising generalization to in vivo acquisitions.

ElecLoad reads weather alerts, grid notices and news event descriptions alongside historical load, judging each event's relevance to the target load, the timing of its effect, and whether events reinforce or offset one another, then decides whether to forecast the whole horizon at once or split it where event effects change. On the segmented path, sub-agents draft each segment independently and the main agent reviews them and sends feedback to selected segments only. Across hourly load data including GEFCom2014, PJM and LCL it ranks near the top in most evaluation scenarios; under multiple events on PJM its WAPE is 3.69% versus 4.31% for a direct-only variant using the same model and inputs.

ElecLoad reads weather alerts, grid notices and news event descriptions alongside historical load, judging each event's relevance to the target load, the timing of its effect, and whether events reinforce or offset one another, then decides whether to forecast the whole horizon at once or split it where event effects change. On the segmented path, sub-agents draft each segment independently and the main agent reviews them and sends feedback to selected segments only. Across hourly load data including GEFCom2014, PJM and LCL it ranks near the top in most evaluation scenarios; under multiple events on PJM its WAPE is 3.69% versus 4.31% for a direct-only variant using the same model and inputs.

Full-waveform LiDAR (FWL) records a complete return histogram for every pixel, so its data volume far exceeds a point cloud, and Transformer-based FWL models have run at only a few frames per second, short of real time. FWL-ToPM prunes low-intensity background patches by signal intensity and then merges the remaining patches along a curve that preserves spatial proximity, raising mean F1 on the Ghost-FWL benchmark from 0.503 to 0.592 while reaching 53.2 FPS. The companion FWLAug expands training samples while preserving each ray's time-of-flight waveform and the frustum geometry, adding a 4.6-point F1 gain on an out-of-distribution set of unseen scenes.

Full-waveform LiDAR (FWL) records a complete return histogram for every pixel, so its data volume far exceeds a point cloud, and Transformer-based FWL models have run at only a few frames per second, short of real time. FWL-ToPM prunes low-intensity background patches by signal intensity and then merges the remaining patches along a curve that preserves spatial proximity, raising mean F1 on the Ghost-FWL benchmark from 0.503 to 0.592 while reaching 53.2 FPS. The companion FWLAug expands training samples while preserving each ray's time-of-flight waveform and the frustum geometry, adding a 4.6-point F1 gain on an out-of-distribution set of unseen scenes.

In two-view data, each view has its own covariance and a cross-view covariance links them, yet covariance neural networks handled only single-view data and could not distinguish these structures once the views were concatenated. CANE splits the model into an in-view covariance filtering stage and a cross-view covariance filtering stage, replacing the whitening and singular-vector selection of canonical correlation analysis with learnable polynomial responses on the covariance spectrum, and trains directly on the downstream task. Across seven brain-imaging and motion-sensing datasets, CANE is best or tied for best on six, reaching a mean absolute error of 4.97 years for age prediction on ADNI1, often with fewer parameters than a two-view VNN.

In two-view data, each view has its own covariance and a cross-view covariance links them, yet covariance neural networks handled only single-view data and could not distinguish these structures once the views were concatenated. CANE splits the model into an in-view covariance filtering stage and a cross-view covariance filtering stage, replacing the whitening and singular-vector selection of canonical correlation analysis with learnable polynomial responses on the covariance spectrum, and trains directly on the downstream task. Across seven brain-imaging and motion-sensing datasets, CANE is best or tied for best on six, reaching a mean absolute error of 4.97 years for age prediction on ADNI1, often with fewer parameters than a two-view VNN.

Arbitrary-scale super-resolution for remote sensing images has mostly been conditioned on the relative upscaling factor between low- and high-resolution images, leaving out absolute spatial resolution in meters per pixel, a physical prior under which the same factor can correspond to very different texture distributions. SpatialSR instead conditions reconstruction on a target spatial resolution: it estimates the low-resolution image's absolute resolution to derive the required factor, draws resolution-specific texture style from a pretrained resolution-to-image diffusion model, and folds the diffusion process into a single-step training constraint. On Vaihingen at a 0.1 m/pixel target from 0.5 m/pixel sources, its segmentation mIoU is 66.7, above the listed comparison methods.

Arbitrary-scale super-resolution for remote sensing images has mostly been conditioned on the relative upscaling factor between low- and high-resolution images, leaving out absolute spatial resolution in meters per pixel, a physical prior under which the same factor can correspond to very different texture distributions. SpatialSR instead conditions reconstruction on a target spatial resolution: it estimates the low-resolution image's absolute resolution to derive the required factor, draws resolution-specific texture style from a pretrained resolution-to-image diffusion model, and folds the diffusion process into a single-step training constraint. On Vaihingen at a 0.1 m/pixel target from 0.5 m/pixel sources, its segmentation mIoU is 66.7, above the listed comparison methods.

IC-GNN regroups the same atlas-defined brain regions into discrete, multiscale functional communities for each person, then uses those communities to guide edge refinement and region-specific weights in a graph neural network. Group-level atlases had smoothed over individual differences in community boundaries, while voxel-level individualized parcellations demand long scans and complicate cross-subject comparison. Across more than 30,000 participants in ABIDE, ADNI and the UK Biobank, IC-GNN reached the highest mean AUROC and accuracy among 15 baselines, with AUROC of 75.41% on ABIDE, 86.46% on ADNI and 60.62% on the UK Biobank.

IC-GNN regroups the same atlas-defined brain regions into discrete, multiscale functional communities for each person, then uses those communities to guide edge refinement and region-specific weights in a graph neural network. Group-level atlases had smoothed over individual differences in community boundaries, while voxel-level individualized parcellations demand long scans and complicate cross-subject comparison. Across more than 30,000 participants in ABIDE, ADNI and the UK Biobank, IC-GNN reached the highest mean AUROC and accuracy among 15 baselines, with AUROC of 75.41% on ABIDE, 86.46% on ADNI and 60.62% on the UK Biobank.

EDP-PI-DOSnet embeds classical energy-stable schemes into a physics-informed operator-splitting network so that the learned discrete operators satisfy a discrete energy dissipation law. The earlier PI-DOSnet did not guarantee that its learned linear operator was symmetric negative semidefinite, energy could rise between steps, and long-time extrapolation needed two or three rounds of retraining. On the Allen–Cahn equation, both EDP-PI-DOSnet variants keep the modified energy non-increasing at every observed time level and give smaller long-time prediction errors without retraining.

EDP-PI-DOSnet embeds classical energy-stable schemes into a physics-informed operator-splitting network so that the learned discrete operators satisfy a discrete energy dissipation law. The earlier PI-DOSnet did not guarantee that its learned linear operator was symmetric negative semidefinite, energy could rise between steps, and long-time extrapolation needed two or three rounds of retraining. On the Allen–Cahn equation, both EDP-PI-DOSnet variants keep the modified energy non-increasing at every observed time level and give smaller long-time prediction errors without retraining.

In power-system dynamic simulation, physics-informed neural networks built on multilayer perceptrons are limited by spectral bias toward smooth functions and struggle to capture electromechanical oscillations, while B-spline physics-informed Kolmogorov-Arnold networks are costly to train. mac-PIKAN embeds the closed-form modal response of the linearized swing equation as a fixed term in the trial solution and uses a Chebyshev-basis KAN to learn a nonlinear correction that vanishes at the initial time, so both initial rotor angle and speed hold exactly. On the single-machine infinite-bus benchmark its mean rotor-angle and frequency errors are 0.29% and 0.99%, against 1.59% and 5.26% for the PINN, and the advantage persists with less training data and at lower damping.

In power-system dynamic simulation, physics-informed neural networks built on multilayer perceptrons are limited by spectral bias toward smooth functions and struggle to capture electromechanical oscillations, while B-spline physics-informed Kolmogorov-Arnold networks are costly to train. mac-PIKAN embeds the closed-form modal response of the linearized swing equation as a fixed term in the trial solution and uses a Chebyshev-basis KAN to learn a nonlinear correction that vanishes at the initial time, so both initial rotor angle and speed hold exactly. On the single-machine infinite-bus benchmark its mean rotor-angle and frequency errors are 0.29% and 0.99%, against 1.59% and 5.26% for the PINN, and the advantage persists with less training data and at lower damping.

Mass-conservation constraints had been embedded only in lumped LSTMs, which yield basin-averaged time series and cannot test whether internal states are hydrologically meaningful. PECAN recasts the constraint in a distributed convolutional LSTM, with a strict precipitation-only variant, PECANMC, and a relaxed variant, PECANMCR, trained solely on streamflow at two gauges. In the snow-dominated North Platte Headwaters in Colorado, both variants match lumped deep-learning baselines and the distributed process model VIC on outlet streamflow and expose spatially explicit internal states, yet diagnostics show the model removes winter snow water storage prematurely instead of holding it for spring release.

Mass-conservation constraints had been embedded only in lumped LSTMs, which yield basin-averaged time series and cannot test whether internal states are hydrologically meaningful. PECAN recasts the constraint in a distributed convolutional LSTM, with a strict precipitation-only variant, PECANMC, and a relaxed variant, PECANMCR, trained solely on streamflow at two gauges. In the snow-dominated North Platte Headwaters in Colorado, both variants match lumped deep-learning baselines and the distributed process model VIC on outlet streamflow and expose spatially explicit internal states, yet diagnostics show the model removes winter snow water storage prematurely instead of holding it for spring release.

Databricks announced general availability of Genie Ontology, a layer of enterprise business context that agents learn automatically and people govern. Enterprise ontologies were previously built and maintained by specialists, covering only a narrow slice of the business and going stale quickly; Genie Ontology extracts reusable ontology snippets from dashboards, SQL queries, notebooks and semantic models, then ranks them by authority before handing context to an agent. On a benchmark built with Turing around a fictional UK energy supplier, Genie drawing on the ontology answered correctly in 91% of runs, against 45% and 58% for two frontier agents.

Databricks announced general availability of Genie Ontology, a layer of enterprise business context that agents learn automatically and people govern. Enterprise ontologies were previously built and maintained by specialists, covering only a narrow slice of the business and going stale quickly; Genie Ontology extracts reusable ontology snippets from dashboards, SQL queries, notebooks and semantic models, then ranks them by authority before handing context to an agent. On a benchmark built with Turing around a fictional UK energy supplier, Genie drawing on the ontology answered correctly in 91% of runs, against 45% and 58% for two frontier agents.

When a traffic forecasting model enforces junction flow conservation at its output, a training sample's influence on predictions no longer travels only through the loss but also through the constraint projection. Classical influence functions assume unconstrained learning and cannot separate perturbations the physics keeps from those the projection removes. This work computes influence inside the feasible set, ranks station sensitivity on hourly PeMS District 3 data covering 318 freeway links, and shows that withholding the ten highest-ranked stations raises the constrained model's MAE by 44.2%.

When a traffic forecasting model enforces junction flow conservation at its output, a training sample's influence on predictions no longer travels only through the loss but also through the constraint projection. Classical influence functions assume unconstrained learning and cannot separate perturbations the physics keeps from those the projection removes. This work computes influence inside the feasible set, ranks station sensitivity on hourly PeMS District 3 data covering 318 freeway links, and shows that withholding the ten highest-ranked stations raises the constrained model's MAE by 44.2%.

Cross-modal remote sensing retrieval must match semantically identical scenes between synthetic aperture radar and multispectral imagery, yet earlier JEPA-style methods regress a single deterministic latent for each masked patch and assume every region is equally predictable. CAPE-JEPA instead outputs a per-patch Gaussian mean and variance, trains it by negative log-likelihood, and adds adversarial cross-modal distributional regularization plus in-training contrastive learning. On BEN-14K and fMoW it beats baselines including X-JEPA, SkySense and CROMA across all eight retrieval directions, gaining up to 13.74 points over X-JEPA on BEN-14K cross-modal directions and up to 15.59 on fMoW, with 407.30M parameters versus 432.72M.

Cross-modal remote sensing retrieval must match semantically identical scenes between synthetic aperture radar and multispectral imagery, yet earlier JEPA-style methods regress a single deterministic latent for each masked patch and assume every region is equally predictable. CAPE-JEPA instead outputs a per-patch Gaussian mean and variance, trains it by negative log-likelihood, and adds adversarial cross-modal distributional regularization plus in-training contrastive learning. On BEN-14K and fMoW it beats baselines including X-JEPA, SkySense and CROMA across all eight retrieval directions, gaining up to 13.74 points over X-JEPA on BEN-14K cross-modal directions and up to 15.59 on fMoW, with 407.30M parameters versus 432.72M.

MPSi replaces the data-dependent, variable-iteration nonlinear refinement of classical channel sounding with a small path-count network plus a fixed-depth unfolded Levenberg–Marquardt refiner, estimating each multipath component's complex gain, angle of arrival and delay from a single pilot observation. In two ray-traced scenarios, a small cell (Bahen) and a dense urban area (Dundas), its initial-access and full-channel-estimation performance is comparable to an enhanced RIMAX baseline while using a fixed and substantially smaller number of refinement iterations; the reconstructed channels give bit-error-rate performance close to the perfect-channel reference.

MPSi replaces the data-dependent, variable-iteration nonlinear refinement of classical channel sounding with a small path-count network plus a fixed-depth unfolded Levenberg–Marquardt refiner, estimating each multipath component's complex gain, angle of arrival and delay from a single pilot observation. In two ray-traced scenarios, a small cell (Bahen) and a dense urban area (Dundas), its initial-access and full-channel-estimation performance is comparable to an enhanced RIMAX baseline while using a fixed and substantially smaller number of refinement iterations; the reconstructed channels give bit-error-rate performance close to the perfect-channel reference.

Computational hyperspectral imaging must hold highlights and shadows in one shot, but the encoding mask's channels have unequal spectral responsivities, so no single exposure keeps every channel in its reliable range. A Zhejiang University team moves multi-exposure fusion ahead of spectral inversion, into the encoded measurement domain, where an analytical reliability prior and a learned channel-specific adjustment jointly set fusion weights, and measurement-domain spectral-angle supervision constrains relative channel responses. On a 9-channel broadband-encoded camera, an optical-density target gives a spectral reconstruction dynamic range of 42.8 dB, 16.0 dB above the best fixed single exposure at 26.8 dB.

Computational hyperspectral imaging must hold highlights and shadows in one shot, but the encoding mask's channels have unequal spectral responsivities, so no single exposure keeps every channel in its reliable range. A Zhejiang University team moves multi-exposure fusion ahead of spectral inversion, into the encoded measurement domain, where an analytical reliability prior and a learned channel-specific adjustment jointly set fusion weights, and measurement-domain spectral-angle supervision constrains relative channel responses. On a 9-channel broadband-encoded camera, an optical-density target gives a spectral reconstruction dynamic range of 42.8 dB, 16.0 dB above the best fixed single exposure at 26.8 dB.

Power swings at AI data centers make grid-import smoothing hard, and with limited battery power and energy, forecast errors of similar size can lead to very different scheduling costs. BS-DFL writes the battery scheduling objective into training: a constraint-aware neural optimization layer first learns schedules without optimal-solution labels, then is frozen so scheduling costs and ramping risk train the demand predictor, and deployment needs no optimization solver. Over a 10,000-minute simulation using loads derived from measured AI workloads, BS-DFL cut the mean absolute minute-to-minute grid-import change by 71.70% and total scheduling cost by 18.01% relative to an MSE-trained predictor with a Gurobi solver, at an average decision time of 18.77 ms.

Power swings at AI data centers make grid-import smoothing hard, and with limited battery power and energy, forecast errors of similar size can lead to very different scheduling costs. BS-DFL writes the battery scheduling objective into training: a constraint-aware neural optimization layer first learns schedules without optimal-solution labels, then is frozen so scheduling costs and ramping risk train the demand predictor, and deployment needs no optimization solver. Over a 10,000-minute simulation using loads derived from measured AI workloads, BS-DFL cut the mean absolute minute-to-minute grid-import change by 71.70% and total scheduling cost by 18.01% relative to an MSE-trained predictor with a Gurobi solver, at an average decision time of 18.77 ms.

Fairness in medical image classification is usually enforced only on output-level group gaps, yet with a linear classification head the output score gap equals the signed sum of feature-attribution differences, so opposing contributions can cancel and make a model look fair. Adding the same attribution-disparity penalty to three equalized-odds surrogates produced positive paired hypervolume gains in all six comparisons on HAM10000 gender fairness and Harvard-GF race fairness, with relative gains of 1.49% to 10.57%. Under a predefined utility constraint, the augmented model families also cut equalized-odds difference by 9.0% to 38.7% while attribution disparity and the cancellation ratio fell.

Fairness in medical image classification is usually enforced only on output-level group gaps, yet with a linear classification head the output score gap equals the signed sum of feature-attribution differences, so opposing contributions can cancel and make a model look fair. Adding the same attribution-disparity penalty to three equalized-odds surrogates produced positive paired hypervolume gains in all six comparisons on HAM10000 gender fairness and Harvard-GF race fairness, with relative gains of 1.49% to 10.57%. Under a predefined utility constraint, the augmented model families also cut equalized-odds difference by 9.0% to 38.7% while attribution disparity and the cancellation ratio fell.

Extreme mass ratio inspirals (EMRIs) have long, multi-harmonic waveforms that make search and parameter estimation expensive. This work trains simulation-based inference networks directly on frequency tracks from a time-frequency search: harmonic pairs are identified correctly in 98% of cases among physically plausible candidates, and sequential inference shrinks the intrinsic-parameter prior volume, with a detection-statistic refinement cutting up to four further orders of magnitude so heavily masked tracks approach the precision of the least-masked cases.

Extreme mass ratio inspirals (EMRIs) have long, multi-harmonic waveforms that make search and parameter estimation expensive. This work trains simulation-based inference networks directly on frequency tracks from a time-frequency search: harmonic pairs are identified correctly in 98% of cases among physically plausible candidates, and sequential inference shrinks the intrinsic-parameter prior volume, with a detection-statistic refinement cutting up to four further orders of magnitude so heavily masked tracks approach the precision of the least-masked cases.

Databricks has extended Genie Agents from turning business questions into SQL queries to parsing documents, connecting to enterprise tools and being shared across organizations. Builders attach Unity Catalog volumes holding PDFs, slide decks or images to an agent; custom document parsing keeps tables, figures, headers and footers intact, the agent retrieves the most relevant file content and reasons over it together with structured data, and answers can cite the exact source page. Unstructured data analysis is in public preview, native MCP connectors are generally available, and OpenSharing and Genie Clean Rooms are in beta.

Databricks has extended Genie Agents from turning business questions into SQL queries to parsing documents, connecting to enterprise tools and being shared across organizations. Builders attach Unity Catalog volumes holding PDFs, slide decks or images to an agent; custom document parsing keeps tables, figures, headers and footers intact, the agent retrieves the most relevant file content and reasons over it together with structured data, and answers can cite the exact source page. Unstructured data analysis is in public preview, native MCP connectors are generally available, and OpenSharing and Genie Clean Rooms are in beta.

Databricks released Genie plugins that let agents in ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot reach enterprise context through the Genie MCP server. Previously an agent facing a business question started a live search, hopping between MCP servers, opening documents and trying tables with little sense of where the right answer lived. With the plugin, Genie interprets business terms through Genie Ontology, finds data the user is allowed to see, writes and runs SQL, and returns an answer with citations, routing to domain Genie Agents when needed.

Databricks released Genie plugins that let agents in ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot reach enterprise context through the Genie MCP server. Previously an agent facing a business question started a live search, hopping between MCP servers, opening documents and trying tables with little sense of where the right answer lived. With the plugin, Genie interprets business terms through Genie Ontology, finds data the user is allowed to see, writes and runs SQL, and returns an answer with citations, routing to domain Genie Agents when needed.

Rare genetic disorders are often missed at first presentation, and most facial diagnostic models are end-to-end black boxes that cannot say what drove a call. ProtoGestalt adds facial-landmark constraints to PIP-Net prototype learning, so each prototype draws evidence only from one of ten anatomical regions and predictions are sums of those local prototypes. On controls matched by age, sex and ethnicity it separated syndromic from unaffected faces at 0.85 accuracy, versus 0.73 for the region-masked variant; AUPRC fell as controls increased but stayed above 0.70.

Rare genetic disorders are often missed at first presentation, and most facial diagnostic models are end-to-end black boxes that cannot say what drove a call. ProtoGestalt adds facial-landmark constraints to PIP-Net prototype learning, so each prototype draws evidence only from one of ten anatomical regions and predictions are sums of those local prototypes. On controls matched by age, sex and ethnicity it separated syndromic from unaffected faces at 0.85 accuracy, versus 0.73 for the region-masked variant; AUPRC fell as controls increased but stayed above 0.70.

High-pressure phase identification has relied on dynamic graph convolutional networks (DGCNNs) to classify each atom's local environment, which is costly to run. A Gaussian kernel model instead places a set of atom detectors around the origin, sums the isotropic Gaussian density contributed by every atom at each detector, and feeds those values to an ordinary neural network. On test sets with 16, 32 and 64 neighbors it reaches 78%, 86% and 90% accuracy, close to DGCNN's 75%, 85% and 92%, while predicting a full simulation in about 20-30 minutes versus about 5-6 hours for the comparison model on the same HPC environment.

High-pressure phase identification has relied on dynamic graph convolutional networks (DGCNNs) to classify each atom's local environment, which is costly to run. A Gaussian kernel model instead places a set of atom detectors around the origin, sums the isotropic Gaussian density contributed by every atom at each detector, and feeds those values to an ordinary neural network. On test sets with 16, 32 and 64 neighbors it reaches 78%, 86% and 90% accuracy, close to DGCNN's 75%, 85% and 92%, while predicting a full simulation in about 20-30 minutes versus about 5-6 hours for the comparison model on the same HPC environment.

The pretrained time-series foundation model Chronos-2 was tested as a surrogate for power-system dynamic simulation. On 1,800 load-disturbance samples from an IEEE 9-bus system, retrieval-augmented in-context learning beat random reference selection, and fine-tuned Chronos-2 had lower errors than a traditional spatiotemporal graph neural network baseline on the 178 held-out samples, with power-balance regularization improving every reported metric except voltage-angle MAE. Its 35.5 ms inference was 563 times faster than the 20 s ANDES simulation.

The pretrained time-series foundation model Chronos-2 was tested as a surrogate for power-system dynamic simulation. On 1,800 load-disturbance samples from an IEEE 9-bus system, retrieval-augmented in-context learning beat random reference selection, and fine-tuned Chronos-2 had lower errors than a traditional spatiotemporal graph neural network baseline on the 178 held-out samples, with power-balance regularization improving every reported metric except voltage-angle MAE. Its 35.5 ms inference was 563 times faster than the 20 s ANDES simulation.

GRACE and GRACE-FO satellites sense only total column water, so groundwater must be separated by subtracting soil moisture, snow and canopy storage, and the models used for that step underestimate the largest trends. A physics-constrained decomposition LSTM splits each cell's groundwater signal into a trend, a seasonal cycle and a residual, holding the trend inside the net atmospheric supply and enforcing basin mass balance during training. The reconstruction reached a root mean squared error of 6.97 centimeters and a Nash-Sutcliffe efficiency of 0.69 on withheld months, and its trend error against monitoring wells was lower than an independent hydrological model and a data-assimilation product at every scale tested.

GRACE and GRACE-FO satellites sense only total column water, so groundwater must be separated by subtracting soil moisture, snow and canopy storage, and the models used for that step underestimate the largest trends. A physics-constrained decomposition LSTM splits each cell's groundwater signal into a trend, a seasonal cycle and a residual, holding the trend inside the net atmospheric supply and enforcing basin mass balance during training. The reconstruction reached a root mean squared error of 6.97 centimeters and a Nash-Sutcliffe efficiency of 0.69 on withheld months, and its trend error against monitoring wells was lower than an independent hydrological model and a data-assimilation product at every scale tested.

OpenRouter released a new generation of Auto Router and added Auto Code, a router tuned for coding work. The previous router relied on a few imprecise signals such as task type; the new routers weigh eleven factors about each request together with live pricing, provider health and cache status to pick both model and reasoning effort. Across five benchmarks Auto Code scores in line with Claude Opus 5.5 while costing 74% less per task, and the routing step runs at 0.65s at p90 in a live test against 1.96s for the old router.

OpenRouter released a new generation of Auto Router and added Auto Code, a router tuned for coding work. The previous router relied on a few imprecise signals such as task type; the new routers weigh eleven factors about each request together with live pricing, provider health and cache status to pick both model and reasoning effort. Across five benchmarks Auto Code scores in line with Claude Opus 5.5 while costing 74% less per task, and the routing step runs at 0.65s at p90 in a live test against 1.96s for the old router.

Event-driven learning turns continuous signals into sparse events, but the triggering threshold has long been chosen empirically. This work maps signed threshold encoding onto discretized learning systems, uses the ratio of normal-condition increment standard deviation to threshold to define an information-sufficiency boundary, then picks the sparsest threshold retaining a prescribed fraction of peak discriminative information. On power-system fault classification, retaining 75% of peak information cut event density about 7.6-fold relative to that boundary while overall accuracy was 94.14%.

Event-driven learning turns continuous signals into sparse events, but the triggering threshold has long been chosen empirically. This work maps signed threshold encoding onto discretized learning systems, uses the ratio of normal-condition increment standard deviation to threshold to define an information-sufficiency boundary, then picks the sparsest threshold retaining a prescribed fraction of peak discriminative information. On power-system fault classification, retaining 75% of peak information cut event density about 7.6-fold relative to that boundary while overall accuracy was 94.14%.

Molecular property prediction has been scattered across separate datasets and evaluation protocols, making it hard to tell whether a molecular representation transfers across properties and chemical systems. Chem World consolidates 17 public datasets with more than 800,000 molecular and chemical-system records into 10 property tracks, paired with a physics-aware compositional modeling framework, PACM. Under the unified evaluation, MolFormer with PACM reaches the lowest macro normalized RMSE of 0.367, and physical-boundary violation falls from 9.7% for the baseline to 1.8%.

Molecular property prediction has been scattered across separate datasets and evaluation protocols, making it hard to tell whether a molecular representation transfers across properties and chemical systems. Chem World consolidates 17 public datasets with more than 800,000 molecular and chemical-system records into 10 property tracks, paired with a physics-aware compositional modeling framework, PACM. Under the unified evaluation, MolFormer with PACM reaches the lowest macro normalized RMSE of 0.367, and physical-boundary violation falls from 9.7% for the baseline to 1.8%.

Deep learning models for electrocardiograms are mostly trained with supervision tied to specific diagnostic labels, which makes them fragile when patient populations or acquisition settings change. This self-supervised pretraining framework learns features from 918,915 unlabeled 12-lead ECG recordings, with one path explicitly measuring each lead's deviation from normal waveform morphology and another reconstructing cross-lead associations under those deviation cues. With the feature extractor frozen and no downstream fine-tuning, it improves AUROC by up to 6.5% and F1 by up to 12.8% over prior methods on four public datasets.

Deep learning models for electrocardiograms are mostly trained with supervision tied to specific diagnostic labels, which makes them fragile when patient populations or acquisition settings change. This self-supervised pretraining framework learns features from 918,915 unlabeled 12-lead ECG recordings, with one path explicitly measuring each lead's deviation from normal waveform morphology and another reconstructing cross-lead associations under those deviation cues. With the feature extractor frozen and no downstream fine-tuning, it improves AUROC by up to 6.5% and F1 by up to 12.8% over prior methods on four public datasets.

Swiss forest composition was previously shown as one dominant species per pixel, erasing the share of accompanying species. A deep-learning model anchored on 0.1 m aerial RGB imagery and combined with airborne laser scanning, annual AlphaEarth satellite embeddings and climate variables was trained on basal-area observations from 3,992 Swiss National Forest Inventory sites to output basal-area shares for ten main tree species and an Other class on a 30 m grid. Spatial cross-validation gave a macro coefficient of determination of 0.485; on forest-reserve data not used in model development, the macro value was 0.552 at 30 m pixel support and 0.733 at plot support, with dominant-class identification in 73.2% of pixels and 81.1% of plots.

Swiss forest composition was previously shown as one dominant species per pixel, erasing the share of accompanying species. A deep-learning model anchored on 0.1 m aerial RGB imagery and combined with airborne laser scanning, annual AlphaEarth satellite embeddings and climate variables was trained on basal-area observations from 3,992 Swiss National Forest Inventory sites to output basal-area shares for ten main tree species and an Other class on a 30 m grid. Spatial cross-validation gave a macro coefficient of determination of 0.485; on forest-reserve data not used in model development, the macro value was 0.552 at 30 m pixel support and 0.733 at plot support, with dominant-class identification in 73.2% of pixels and 81.1% of plots.

In bulk RNA-seq differential expression, general-purpose LLM agents tend to pick DESeq2 or edgeR without checking whether those methods' statistical assumptions fit the data at hand. DE-Rigor-Agent splits the analysis into input standardization, method filtering and method execution, and before running anything it evaluates six common methods against expert-curated method cards and dataset-specific diagnostics, recording why each is kept or excluded. Across 140 simulated datasets its filtering decisions largely matched the incompatibilities built into the simulations and stayed stable across random seeds, while two Codex baselines chose DESeq2 or edgeR for nearly all datasets.

In bulk RNA-seq differential expression, general-purpose LLM agents tend to pick DESeq2 or edgeR without checking whether those methods' statistical assumptions fit the data at hand. DE-Rigor-Agent splits the analysis into input standardization, method filtering and method execution, and before running anything it evaluates six common methods against expert-curated method cards and dataset-specific diagnostics, recording why each is kept or excluded. Across 140 simulated datasets its filtering decisions largely matched the incompatibilities built into the simulations and stayed stable across random seeds, while two Codex baselines chose DESeq2 or edgeR for nearly all datasets.

Most knowledge-work benchmarks can be solved with text alone, without ever looking at an image or page layout, yet everyday professional work such as architectural blueprints, engineering schematics and polished reports is deeply visual. Google and Artificial Analysis launched MKA, a multimodal knowledge agent benchmark whose 200 tasks were written by domain experts across 10 verticals and 83 occupations, each paired with roughly 70 yes/no rubric checks weighted 75% on factual accuracy and 25% on visual design. When domain experts graded model outputs themselves, they agreed with the autorater 96% of the time; the 200-task set stays private, and 10 public tasks are released on Hugging Face so anyone can inspect the format.

Most knowledge-work benchmarks can be solved with text alone, without ever looking at an image or page layout, yet everyday professional work such as architectural blueprints, engineering schematics and polished reports is deeply visual. Google and Artificial Analysis launched MKA, a multimodal knowledge agent benchmark whose 200 tasks were written by domain experts across 10 verticals and 83 occupations, each paired with roughly 70 yes/no rubric checks weighted 75% on factual accuracy and 25% on visual design. When domain experts graded model outputs themselves, they agreed with the autorater 96% of the time; the 200-task set stays private, and 10 public tasks are released on Hugging Face so anyone can inspect the format.

Sleep foundation models compress multi-second EEG into compact tokens, leaving too little temporal detail to trace the onset and offset of sleep spindles. One approach keeps the SleepFM encoder frozen and adds a local time-frequency branch that preserves 11–16 Hz sigma-band energy and intra-token position, then uses residual cross-attention to bring in pretrained temporal context. On the MODA benchmark across six folds, event-level F1 is 0.8080±0.0040 at IoU 0.2 and 0.7766±0.0050 at IoU 0.5, with a faster decline at stricter boundary overlap.

Sleep foundation models compress multi-second EEG into compact tokens, leaving too little temporal detail to trace the onset and offset of sleep spindles. One approach keeps the SleepFM encoder frozen and adds a local time-frequency branch that preserves 11–16 Hz sigma-band energy and intra-token position, then uses residual cross-attention to bring in pretrained temporal context. On the MODA benchmark across six folds, event-level F1 is 0.8080±0.0040 at IoU 0.2 and 0.7766±0.0050 at IoU 0.5, with a faster decline at stricter boundary overlap.

Nace released Drex, a decision model for document processing that, after the parsing engine returns layout-aware Markdown, judges page by page whether text matches the page image, whether a table kept its shape and whether a ticked box survived, and triggers correction only where something is in doubt. On Nace's own parsing leaderboard, NDI plus Drex scores a Parse Index of 0.834, ahead of all 12 parsers tested, including LlamaParse Agentic at 0.820; it improves on the engine alone on OmniDocBench, MDPBench and ExtractBench, and is within noise of it on ParseBench and RealDocBench.

Nace released Drex, a decision model for document processing that, after the parsing engine returns layout-aware Markdown, judges page by page whether text matches the page image, whether a table kept its shape and whether a ticked box survived, and triggers correction only where something is in doubt. On Nace's own parsing leaderboard, NDI plus Drex scores a Parse Index of 0.834, ahead of all 12 parsers tested, including LlamaParse Agentic at 0.820; it improves on the engine alone on OmniDocBench, MDPBench and ExtractBench, and is within noise of it on ParseBench and RealDocBench.

Ancient-book pages often carry paper ageing, water-induced distortion, insect holes and mold at once, while existing public benchmarks are limited in scale and stack isolated degradations, which weakens supervised restoration. AncBench supplies 100,379 pixel-aligned clean–degraded pairs in which chemical, physical and biological damage are composed in sequence; trained on it, AncRestore first estimates page-wide and local degradation, then dynamically selects and distributes visual priors according to those cues. On the AncBench test set AncRestore reaches 26.70 dB PSNR, 0.9191 SSIM and 0.1346 LPIPS, with 89.58% character-recognition accuracy.

Ancient-book pages often carry paper ageing, water-induced distortion, insect holes and mold at once, while existing public benchmarks are limited in scale and stack isolated degradations, which weakens supervised restoration. AncBench supplies 100,379 pixel-aligned clean–degraded pairs in which chemical, physical and biological damage are composed in sequence; trained on it, AncRestore first estimates page-wide and local degradation, then dynamically selects and distributes visual priors according to those cues. On the AncBench test set AncRestore reaches 26.70 dB PSNR, 0.9191 SSIM and 0.1346 LPIPS, with 89.58% character-recognition accuracy.

Sequence models for oral resistome proteins lacked an evaluation that rules out homology-driven memorization. The study clustered 26,324 proteins into 1,437 clusters at 40% sequence identity, trained eight causal architectures under one protocol, and validated externally on CARD and NCBI AMRFinderPlus. Hyena reached the lowest held-out perplexity at 14.93, but only four of eight architectures beat the residue-composition baseline of 17.99 externally, where parameter-matched MoR led at 17.40 and the Transformer degraded by 27%.

Sequence models for oral resistome proteins lacked an evaluation that rules out homology-driven memorization. The study clustered 26,324 proteins into 1,437 clusters at 40% sequence identity, trained eight causal architectures under one protocol, and validated externally on CARD and NCBI AMRFinderPlus. Hyena reached the lowest held-out perplexity at 14.93, but only four of eight architectures beat the residue-composition baseline of 17.99 externally, where parameter-matched MoR led at 17.40 and the Transformer degraded by 27%.

ElevenLabs and NVIDIA are expanding their partnership to bring NVIDIA's Nemotron open models into ElevenAgents, ElevenLabs' conversational agent platform, and to add 15 MW of Vera Rubin capacity on top of existing Blackwell Ultra systems. On the ElevenAgents benchmark for realistic customer-support tasks, Nemotron 3.5 Super VL reached leading end-to-end task-completion accuracy at 70% lower cost than other leading open models. The two companies are also applying voice agents to Poland's national e-Health central appointment system, with an on-premises deployment that keeps citizen data and audio within government infrastructure.

ElevenLabs and NVIDIA are expanding their partnership to bring NVIDIA's Nemotron open models into ElevenAgents, ElevenLabs' conversational agent platform, and to add 15 MW of Vera Rubin capacity on top of existing Blackwell Ultra systems. On the ElevenAgents benchmark for realistic customer-support tasks, Nemotron 3.5 Super VL reached leading end-to-end task-completion accuracy at 70% lower cost than other leading open models. The two companies are also applying voice agents to Poland's national e-Health central appointment system, with an on-premises deployment that keeps citizen data and audio within government infrastructure.

Cardiac motion tracking from cine magnetic resonance previously either lacked anatomical correspondence across subjects or relied on a slow, error-prone manual atlas fit. CardioDrift generalises coherent point drift to jointly register the left-ventricular endocardium, epicardium and right-ventricular atlas surfaces, and estimates per-slice translations inside the same probabilistic model to correct slice misalignment, fitting and propagating the atlas automatically from short-axis segmentation contours alone. On 1,301 UK Biobank subjects and the M&M dataset, its rigid fitting and non-rigid propagation lead every distance and overlap metric against the strongest baseline.

Cardiac motion tracking from cine magnetic resonance previously either lacked anatomical correspondence across subjects or relied on a slow, error-prone manual atlas fit. CardioDrift generalises coherent point drift to jointly register the left-ventricular endocardium, epicardium and right-ventricular atlas surfaces, and estimates per-slice translations inside the same probabilistic model to correct slice misalignment, fitting and propagating the atlas automatically from short-axis segmentation contours alone. On 1,301 UK Biobank subjects and the M&M dataset, its rigid fitting and non-rigid propagation lead every distance and overlap metric against the strongest baseline.

Causaly has released Causaly Work in public preview for knowledge workers at life-sciences enterprises, combining computational analysis, evidence investigation and professional work-product creation inside a single long-running assignment. Previously such work moved between retrieval, code and documents, making every line of evidence hard to track. The system can run analytical code on supplied tables, compare reported trial endpoints and populations, and assemble findings with citations and open questions into a proposal and presentation, while people review the results and remain responsible for decisions.

Causaly has released Causaly Work in public preview for knowledge workers at life-sciences enterprises, combining computational analysis, evidence investigation and professional work-product creation inside a single long-running assignment. Previously such work moved between retrieval, code and documents, making every line of evidence hard to track. The system can run analytical code on supplied tables, compare reported trial endpoints and populations, and assemble findings with citations and open questions into a proposal and presentation, while people review the results and remain responsible for decisions.

Neural surrogates for structural dynamic response no longer fit the whole frequency-response curve directly; they first predict frequency-independent modal quantities and then reconstruct the response by analytical modal superposition. Across 20,000 designs each for a steel truss tower and a tapered beam, with supplied modal frequencies, 5% reference damping and a declared band, the modal configurations give mean displacement-history errors of 0.73–0.76% and 0.029–0.116% respectively. These figures cover fixed topology and conditional inference, and do not yet establish design-only deployment or an earthquake-engineering assessment.

Neural surrogates for structural dynamic response no longer fit the whole frequency-response curve directly; they first predict frequency-independent modal quantities and then reconstruct the response by analytical modal superposition. Across 20,000 designs each for a steel truss tower and a tapered beam, with supplied modal frequencies, 5% reference damping and a declared band, the modal configurations give mean displacement-history errors of 0.73–0.76% and 0.029–0.116% respectively. These figures cover fixed topology and conditional inference, and do not yet establish design-only deployment or an earthquake-engineering assessment.

For home health consultation and triage in traditional Chinese medicine gynecology, MAGR-TCM turns 741 clinical case records into a knowledge graph of 10,231 entities and 32,051 relationships, then has four specialized agents reason step by step along the disease-pattern-principle-formula-medicinal chain. On 30 held-out test cases its RAGAS faithfulness is 0.821, against 0.766 for GraphRAG, 0.651 for LightRAG and 0.590 for Naive RAG; on a safety set of 10 emergency and 20 routine cases, the red-flag mechanism intercepted and passed every case, with sensitivity and specificity both 1.000. These results come from retrospective, expert-annotated records and have not yet been validated prospectively with real home users.

For home health consultation and triage in traditional Chinese medicine gynecology, MAGR-TCM turns 741 clinical case records into a knowledge graph of 10,231 entities and 32,051 relationships, then has four specialized agents reason step by step along the disease-pattern-principle-formula-medicinal chain. On 30 held-out test cases its RAGAS faithfulness is 0.821, against 0.766 for GraphRAG, 0.651 for LightRAG and 0.590 for Naive RAG; on a safety set of 10 emergency and 20 routine cases, the red-flag mechanism intercepted and passed every case, with sensitivity and specificity both 1.000. These results come from retrospective, expert-annotated records and have not yet been validated prospectively with real home users.

At its New York Summit, ElevenLabs announced a set of enterprise interaction products: a plan to help businesses fine-tune certain open-source language models and its own voice models on live-deployment data in the coming months, a preview of Interaction Engine that manages turn-taking and interruptions and picks a large model per part of an exchange, Memory and Routines for ElevenAgents that can run multi-step work before, during and after a conversation, and free early access to the workplace assistant ElevenAI. The company also announced a partner program, a Marketplace, and an expanded NVIDIA partnership bringing Nemotron 3.5 Super VL into ElevenAgents. These are product announcements and previews; the material reports no comparative performance data.

At its New York Summit, ElevenLabs announced a set of enterprise interaction products: a plan to help businesses fine-tune certain open-source language models and its own voice models on live-deployment data in the coming months, a preview of Interaction Engine that manages turn-taking and interruptions and picks a large model per part of an exchange, Memory and Routines for ElevenAgents that can run multi-step work before, during and after a conversation, and free early access to the workplace assistant ElevenAI. The company also announced a partner program, a Marketplace, and an expanded NVIDIA partnership bringing Nemotron 3.5 Super VL into ElevenAgents. These are product announcements and previews; the material reports no comparative performance data.

Solar filaments in Hα images break into fragments as they evolve, and projection effects can separate pieces of one filament while bringing pieces of different filaments close together, so distance and orientation alone cannot decide which fragments belong together. The study treats each fragment as a graph node and each candidate pair as an edge, predicts at the edge level with a graph neural network, and replaces conventional additive aggregation with multiplicative aggregation so that inconsistent neighbors contribute less. On a dataset built from about 100 Kanzelhöhe Observatory Hα images and about 1,700 fragments, the method reaches AUC 0.867 and AP 0.792 at kNN 10, above GCN at 0.832/0.747 and GAT at 0.842/0.766.

Solar filaments in Hα images break into fragments as they evolve, and projection effects can separate pieces of one filament while bringing pieces of different filaments close together, so distance and orientation alone cannot decide which fragments belong together. The study treats each fragment as a graph node and each candidate pair as an edge, predicts at the edge level with a graph neural network, and replaces conventional additive aggregation with multiplicative aggregation so that inconsistent neighbors contribute less. On a dataset built from about 100 Kanzelhöhe Observatory Hα images and about 1,700 fragments, the method reaches AUC 0.867 and AP 0.792 at kNN 10, above GCN at 0.832/0.747 and GAT at 0.842/0.766.

bAIes originally drew structural information only from AlphaFold2 distograms to build atomic-resolution conformational ensembles of intrinsically disordered proteins. It now accepts six predictors, AlphaFold2, AlphaFold3, Boltz-1, Boltz-2, ESMFold and ESMFold2, and was compared on NMR and SAXS data for 17 disordered proteins. Ensembles from every predictor reach average accuracy comparable to the original AlphaFold2 version, best on local observables such as chemical shifts, RDCs and 3J couplings, and weaker on long-range and global observables such as PREs and SAXS.

bAIes originally drew structural information only from AlphaFold2 distograms to build atomic-resolution conformational ensembles of intrinsically disordered proteins. It now accepts six predictors, AlphaFold2, AlphaFold3, Boltz-1, Boltz-2, ESMFold and ESMFold2, and was compared on NMR and SAXS data for 17 disordered proteins. Ensembles from every predictor reach average accuracy comparable to the original AlphaFold2 version, best on local observables such as chemical shifts, RDCs and 3J couplings, and weaker on long-range and global observables such as PREs and SAXS.

Spatial transcriptomics models usually exchange representations only between cells or spots, with gene information arriving through the attached expression profile. SpaAxFormer keeps every cell–gene entry as its own state, first letting genes within a cell reference each other, then exchanging the same gene's state across spatial neighbors, with a separate cell-level summary. On six DLPFC slices its mean L1 imputation distance was lower than Impeller's, for example 0.239 versus 0.254 on slice 151510; on Stereo-seq and Slide-seq V2, Impeller had lower mean L1 and RMSE.

Spatial transcriptomics models usually exchange representations only between cells or spots, with gene information arriving through the attached expression profile. SpaAxFormer keeps every cell–gene entry as its own state, first letting genes within a cell reference each other, then exchanging the same gene's state across spatial neighbors, with a separate cell-level summary. On six DLPFC slices its mean L1 imputation distance was lower than Impeller's, for example 0.239 versus 0.254 on slice 151510; on Stereo-seq and Slide-seq V2, Impeller had lower mean L1 and RMSE.

McKinsey and ElevenLabs announced an alliance that pairs ElevenLabs' frontier conversational AI technology and agents platform with McKinsey's industry transformation experience and QuantumBlack's AI engineering, helping organizations prove the value of conversational AI through initial use cases before scaling across the enterprise. Each deployment, the two firms say, combines multilingual conversational technology with agent design, systems integration and operating-model changes, starting from settings such as telecommunications and extending to insurance for claims intake, policy servicing and everyday customer questions.

McKinsey and ElevenLabs announced an alliance that pairs ElevenLabs' frontier conversational AI technology and agents platform with McKinsey's industry transformation experience and QuantumBlack's AI engineering, helping organizations prove the value of conversational AI through initial use cases before scaling across the enterprise. Each deployment, the two firms say, combines multilingual conversational technology with agent design, systems integration and operating-model changes, starting from settings such as telecommunications and extending to insurance for claims intake, policy servicing and everyday customer questions.

Inside the unified OneFormer3D query decoder, regional feature aggregation previously relied on fixed statistics such as mean or max; COC replaces that step with a learned positive scalar weight per sample, adds contracted first- and second-order geometric descriptors, and evolves states bidirectionally along decoder depth. On ScanNetV2 instance mAP rises from 59.3 to 59.8, semantic mIoU from 76.6 to 77.4 and panoptic PQ from 71.2 to 72.0; on S3DIS Area-5 mIoU rises from 72.4 to 73.8. Gains vary by dataset and metric, and the authors note cross-paper numbers are reference points only.

Inside the unified OneFormer3D query decoder, regional feature aggregation previously relied on fixed statistics such as mean or max; COC replaces that step with a learned positive scalar weight per sample, adds contracted first- and second-order geometric descriptors, and evolves states bidirectionally along decoder depth. On ScanNetV2 instance mAP rises from 59.3 to 59.8, semantic mIoU from 76.6 to 77.4 and panoptic PQ from 71.2 to 72.0; on S3DIS Area-5 mIoU rises from 72.4 to 73.8. Gains vary by dataset and metric, and the authors note cross-paper numbers are reference points only.

Google has brought the Gemini app to Apple CarPlay, letting iOS users hold natural, hands-free conversations with Gemini Live from the car display and pick up a recent chat. The app had already rolled out on Android Auto and cars with Google built-in; this extends it to CarPlay. Conversations sync automatically to the Gemini app on the iPhone, so summaries, notes and brainstorms are waiting in chat history once the driver parks.

Google has brought the Gemini app to Apple CarPlay, letting iOS users hold natural, hands-free conversations with Gemini Live from the car display and pick up a recent chat. The app had already rolled out on Android Auto and cars with Google built-in; this extends it to CarPlay. Conversations sync automatically to the Gemini app on the iPhone, so summaries, notes and brainstorms are waiting in chat history once the driver parks.

ElevenLabs' Ads Engine can now route ads through human review before they are finalized, with reviewers drawn from ElevenProductions localization experts and in-house ElevenLabs creative specialists. Marketing teams asked to ship more creative into more markets need translators, localization professionals and designers that many teams do not have, so some brands run original-language ads globally or make fewer variants. Ads Engine connects to ad platforms, pulls existing creative, translates image text, adapts images and dubs video with Dubbing v2, after which reviewers check translation accuracy and dub naturalness before the finished ad returns for approval or is pushed to the platforms.

ElevenLabs' Ads Engine can now route ads through human review before they are finalized, with reviewers drawn from ElevenProductions localization experts and in-house ElevenLabs creative specialists. Marketing teams asked to ship more creative into more markets need translators, localization professionals and designers that many teams do not have, so some brands run original-language ads globally or make fewer variants. Ads Engine connects to ad platforms, pulls existing creative, translates image text, adapts images and dubs video with Dubbing v2, after which reviewers check translation accuracy and dub naturalness before the finished ad returns for approval or is pushed to the platforms.

Mastra has added tasks to its MCP server and client so that long-running tool calls run in the background. Previously MCP tool calls were synchronous: a single HTTP request stayed open until the call finished, blocking the agent loop, and although Mastra's existing background tasks primitive could free the loop, a call that outlasted a proxy timeout or a load balancer's idle limit would still fail. Now the server returns a task ID that ends the request, the client polls that ID for working, input_required, completed, failed or cancelled status, and a notification signal reaches the agent when the task completes.

Mastra has added tasks to its MCP server and client so that long-running tool calls run in the background. Previously MCP tool calls were synchronous: a single HTTP request stayed open until the call finished, blocking the agent loop, and although Mastra's existing background tasks primitive could free the loop, a call that outlasted a proxy timeout or a load balancer's idle limit would still fail. Now the server returns a task ID that ends the request, the client polls that ID for working, input_required, completed, failed or cancelled status, and a notification signal reaches the agent when the task completes.

ElevenLabs launched an enterprise marketplace listing industry-specific AI agents and voice applications built on ElevenLabs by partners including Salesforce, IBM, NiCE, Twilio and Zendesk, spanning financial services, insurance, healthcare, telecommunications, retail and government. Enterprises can browse solutions and contact the partners behind them to evaluate deployment, while partners gain a new channel to showcase their applications. ElevenLabs says that from next year enterprises will be able to buy eligible partner solutions through the marketplace using their existing ElevenLabs commitments.

ElevenLabs launched an enterprise marketplace listing industry-specific AI agents and voice applications built on ElevenLabs by partners including Salesforce, IBM, NiCE, Twilio and Zendesk, spanning financial services, insurance, healthcare, telecommunications, retail and government. Enterprises can browse solutions and contact the partners behind them to evaluate deployment, while partners gain a new channel to showcase their applications. ElevenLabs says that from next year enterprises will be able to buy eligible partner solutions through the marketplace using their existing ElevenLabs commitments.

StoneWise's small-molecule early-discovery agent MolVortex has completed local deployment on Lianyungang's "AI + medicine" public service platform and been listed in its AI tool marketplace, open to local pharma companies, hospital researchers, universities and innovation service providers. The platform, invested in and built by Jiangsu XinHailian Group, went live at the 8th Lianyungang Pharmaceutical Technology Conference, aiming to let companies call AI capabilities and data assets on demand without building their own compute facilities. MolVortex covers target research, pocket analysis, molecule generation and dynamics evaluation across the early-discovery path from hit to preclinical candidate.

StoneWise's small-molecule early-discovery agent MolVortex has completed local deployment on Lianyungang's "AI + medicine" public service platform and been listed in its AI tool marketplace, open to local pharma companies, hospital researchers, universities and innovation service providers. The platform, invested in and built by Jiangsu XinHailian Group, went live at the 8th Lianyungang Pharmaceutical Technology Conference, aiming to let companies call AI capabilities and data assets on demand without building their own compute facilities. MolVortex covers target research, pocket analysis, molecule generation and dynamics evaluation across the early-discovery path from hit to preclinical candidate.

NVIDIA's open-source AICR now includes Google's k8s-aibom as a qualified component in its recipes, so a GPU-accelerated Kubernetes platform can deploy a known-good configuration while recording which AI workloads actually run in the cluster. Previously a recipe fixed only the platform configuration, and workloads updated independently by application teams, plus mutable tags like :latest, could leave two nodes running different code under the same tag. Two GKE recipes select the component by default, and others can opt in at generation time.

NVIDIA's open-source AICR now includes Google's k8s-aibom as a qualified component in its recipes, so a GPU-accelerated Kubernetes platform can deploy a known-good configuration while recording which AI workloads actually run in the cluster. Previously a recipe fixed only the platform configuration, and workloads updated independently by application teams, plus mutable tags like :latest, could leave two nodes running different code under the same tag. Two GKE recipes select the component by default, and others can opt in at generation time.

Sierra has released its first Accelerators in Agent Studio, packaging high-volume customer-service workflows such as order tracking, appointment scheduling, first notice of loss and ID verification together with basic integrations and guardrails, which Ghostwriter then customizes to a company's own policies. Unlike one-time templates, Accelerators record a company's changes separately from the underlying workflow, so those changes carry over when Sierra ships an update, and the company reviews the full set of changes before deciding whether to apply them. The official example is a media company adding the retention Accelerator and adapting it to its own cancellation policies and offers, with the agent ready for testing in minutes.

Sierra has released its first Accelerators in Agent Studio, packaging high-volume customer-service workflows such as order tracking, appointment scheduling, first notice of loss and ID verification together with basic integrations and guardrails, which Ghostwriter then customizes to a company's own policies. Unlike one-time templates, Accelerators record a company's changes separately from the underlying workflow, so those changes carry over when Sierra ships an update, and the company reviews the full set of changes before deciding whether to apply them. The official example is a media company adding the retention Accelerator and adapting it to its own cancellation policies and offers, with the agent ready for testing in minutes.

Mastra agents can now receive Ring doorbell events, each arriving in the agent's thread as a notification signal with a configurable priority. Agents previously responded mainly to software events from Slack, Discord, iMessage and other channel providers; now they can fetch a snapshot or video clip from the Ring API and decide what to do when someone is at the door. By default each event carries a snapshot image; setting visibleToModel to true on tools.clip and using a model that accepts video input lets the agent watch clips too.

Mastra agents can now receive Ring doorbell events, each arriving in the agent's thread as a notification signal with a configurable priority. Agents previously responded mainly to software events from Slack, Discord, iMessage and other channel providers; now they can fetch a snapshot or video clip from the Ring API and decide what to do when someone is at the door. By default each event carries a snapshot image; setting visibleToModel to true on tools.clip and using a model that accepts video input lets the agent watch clips too.

ModRetro released a Chromatic plugin for Codex that lets the model drive a set of Game Boy Color development tools from chat and turn a prompt into a playable game. The plugin exposes animation, maps, ROM playtesting, on-screen annotation and button-sequence testing as tools the model can call, and it can write the compiled ROM to a rewritable cartridge on the Chromatic handheld. In the example, the top-down adventure game Lodge Podge was generated from a prompt and then revised on request to change the character's build and the materials needed for construction.

ModRetro released a Chromatic plugin for Codex that lets the model drive a set of Game Boy Color development tools from chat and turn a prompt into a playable game. The plugin exposes animation, maps, ROM playtesting, on-screen annotation and button-sequence testing as tools the model can call, and it can write the compiled ROM to a rewritable cartridge on the Chromatic handheld. In the example, the top-down adventure game Lodge Podge was generated from a prompt and then revised on request to change the character's build and the materials needed for construction.