AI and science frontiers · 2026-10-28

In safety-of-the-intended-functionality analysis under ISO 21448, an uncertainty threshold that keeps a LiDAR detector within a performance target under one condition had no procedure to check whether it still does under another. The threshold is transferred unchanged into the condition of use, and a statistical test decides at a stated error probability whether the target is applicable, not applicable or not determined, together with a lower bound on the independent units needed to show the target is met. On PointPillars ensembles over KITTI and nuScenes, counted in frames, the threshold transferred from nuScenes to KITTI was not applicable on all five splits; counted in recorded sequences, neither dataset held enough units to show any target limit met.

In safety-of-the-intended-functionality analysis under ISO 21448, an uncertainty threshold that keeps a LiDAR detector within a performance target under one condition had no procedure to check whether it still does under another. The threshold is transferred unchanged into the condition of use, and a statistical test decides at a stated error probability whether the target is applicable, not applicable or not determined, together with a lower bound on the independent units needed to show the target is met. On PointPillars ensembles over KITTI and nuScenes, counted in frames, the threshold transferred from nuScenes to KITTI was not applicable on all five splits; counted in recorded sequences, neither dataset held enough units to show any target limit met.

Physical hardware security evaluation has depended on a person choosing each laser fault-injection position, power and timing. A large-language-model agent now links firmware reverse engineering, chip imagery and experimental decisions into a closed loop, choosing the next shot inside human-approved safety bounds. On an ATmega328P it produced a skipped load instruction and a wrong-PIN acceptance confirmed by a separate state read; on an Artix-7 FPGA the same workflow extended to ciphertext-error classification in programmable logic.

Physical hardware security evaluation has depended on a person choosing each laser fault-injection position, power and timing. A large-language-model agent now links firmware reverse engineering, chip imagery and experimental decisions into a closed loop, choosing the next shot inside human-approved safety bounds. On an ATmega328P it produced a skipped load instruction and a wrong-PIN acceptance confirmed by a separate state read; on an Artix-7 FPGA the same workflow extended to ciphertext-error classification in programmable logic.

Turning an antibody into a drug means tuning binding, stability, aggregation and other properties that pull against one another, and lead optimization often takes months to years. Anthropic and BigHat had one researcher and Claude operate BigHat's automated lab via MCP to optimize one VHH and one single-chain Fv antibody across seven design rounds. In 10 weeks they synthesized and comprehensively tested 1,232 antibodies; the most improved variants of both bound their targets up to nine times more tightly, one reaching sub-nanomolar affinity, stayed stable in solution at temperatures up to 18°C higher, and aggregated far more slowly.

Turning an antibody into a drug means tuning binding, stability, aggregation and other properties that pull against one another, and lead optimization often takes months to years. Anthropic and BigHat had one researcher and Claude operate BigHat's automated lab via MCP to optimize one VHH and one single-chain Fv antibody across seven design rounds. In 10 weeks they synthesized and comprehensively tested 1,232 antibodies; the most improved variants of both bound their targets up to nine times more tightly, one reaching sub-nanomolar affinity, stayed stable in solution at temperatures up to 18°C higher, and aggregated far more slowly.

Code assistants often emit Python calls that belong to an older release and no longer exist in the version a user has installed. PyPI API Drift statically extracts, without importing, the public APIs of 55,565 releases across the 1,872 most-downloaded libraries, recording signatures and release-to-release changes while separating true removals from names that merely left the documented API. In a preregistered code-repair experiment, four models from one family raised API-compatible, intent-preserving migrations to 68.5%–69.9% with this API history, against 7.0%–56.6% unaided.

Code assistants often emit Python calls that belong to an older release and no longer exist in the version a user has installed. PyPI API Drift statically extracts, without importing, the public APIs of 55,565 releases across the 1,872 most-downloaded libraries, recording signatures and release-to-release changes while separating true removals from names that merely left the documented API. In a preregistered code-repair experiment, four models from one family raised API-compatible, intent-preserving migrations to 68.5%–69.9% with this API history, against 7.0%–56.6% unaided.

Manipulations in long videos often touch only a few seconds, while existing audio-visual deepfake benchmarks mostly use short talking-head clips and long-form forensics has been largely visual. AV-FABLE pairs 16,650 pristine–manipulated twins (998 hours, 27 editing configurations) with separate audio and visual timelines, so localization must report both the time boundaries and the affected modality. Reading frozen Qwen3-Omni perception features through separate audio and visual temporal decoders, NFSR reaches 91.30 in-domain AV-mAP against 84.76 for UMMAFormer on the same features, and the highest AV-mAP in all three out-of-distribution settings.

Manipulations in long videos often touch only a few seconds, while existing audio-visual deepfake benchmarks mostly use short talking-head clips and long-form forensics has been largely visual. AV-FABLE pairs 16,650 pristine–manipulated twins (998 hours, 27 editing configurations) with separate audio and visual timelines, so localization must report both the time boundaries and the affected modality. Reading frozen Qwen3-Omni perception features through separate audio and visual temporal decoders, NFSR reaches 91.30 in-domain AV-mAP against 84.76 for UMMAFormer on the same features, and the highest AV-mAP in all three out-of-distribution settings.

Liquid AI released LFM3, a family that puts text, vision, audio, omni and decision models on one shared backbone of roughly 2B parameters, with base and post-trained weights open-weight on Hugging Face and its API. The backbone mixes gated short convolutions, sliding-window attention and global attention, was pre-trained on 22T tokens, and a final 3.5T-token mid-training stage extends context to 256K. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.

Liquid AI released LFM3, a family that puts text, vision, audio, omni and decision models on one shared backbone of roughly 2B parameters, with base and post-trained weights open-weight on Hugging Face and its API. The backbone mixes gated short convolutions, sliding-window attention and global attention, was pre-trained on 22T tokens, and a final 3.5T-token mid-training stage extends context to 256K. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.

Language models are often trained against a learned judge instead of the truth, but what that judge rewards could previously be seen only after training, from the policy's behavior. Three judges were trained from a 1B-parameter model on 500, 2,000 and 8,000 labeled answers, and a single-edit probe measured each judge's lift on wrong answers before any training. The probe predicted a climb onto flagged text against the weakest judge; after training, 115 of 128 final responses were one Sahara sentence ending "Answer: Africa", and the judge's mean score rose from 0.20 to 0.82. Against the strongest judge it predicted nothing climbable, and that run was not exploited.

Language models are often trained against a learned judge instead of the truth, but what that judge rewards could previously be seen only after training, from the policy's behavior. Three judges were trained from a 1B-parameter model on 500, 2,000 and 8,000 labeled answers, and a single-edit probe measured each judge's lift on wrong answers before any training. The probe predicted a climb onto flagged text against the weakest judge; after training, 115 of 128 final responses were one Sahara sentence ending "Answer: Africa", and the judge's mean score rose from 0.20 to 0.82. Against the strongest judge it predicted nothing climbable, and that run was not exploited.

When large language models apply a codebook, a structured annotation guide that maps theoretical concepts onto labels, high aggregate scores can hide label-level failures. Across four datasets and five models, controlled prompt perturbations, repeated runs and a domain shift left worst-label F1 about 26 to 30 points below micro-F1, and models leaned on label names more than on definitions. These failures stayed invisible in accuracy and micro-F1 and persisted after prompt optimization.

When large language models apply a codebook, a structured annotation guide that maps theoretical concepts onto labels, high aggregate scores can hide label-level failures. Across four datasets and five models, controlled prompt perturbations, repeated runs and a domain shift left worst-label F1 about 26 to 30 points below micro-F1, and models leaned on label names more than on definitions. These failures stayed invisible in accuracy and micro-F1 and persisted after prompt optimization.

Google DeepMind's Gemini Robotics ER 2, a high-level reasoning model for robots, is now in public preview on the Gemini platform, letting enterprises declare low-level controls such as vision-language-action models and navigation APIs as tools for the model to orchestrate. The model classifies task progress from continuous video and pinpoints critical event frames, with a stated 91.3% accuracy and 0.96-second mean absolute frame distance. It is described as working across embodiments from stationary arms to humanoids, with simulation, evaluation and deployment handled on Google Cloud.

Google DeepMind's Gemini Robotics ER 2, a high-level reasoning model for robots, is now in public preview on the Gemini platform, letting enterprises declare low-level controls such as vision-language-action models and navigation APIs as tools for the model to orchestrate. The model classifies task progress from continuous video and pinpoints critical event frames, with a stated 91.3% accuracy and 0.96-second mean absolute frame distance. It is described as working across embodiments from stationary arms to humanoids, with simulation, evaluation and deployment handled on Google Cloud.

Canelé used an AI-agent-orchestrated rational design loop to refactor the CAR intracellular signaling domain, compressing the second-generation BBζ domain into a 54-amino-acid design called Stoat and then a 45-amino-acid variant, Happy Stoat. At the same mRNA dose, Happy Stoat puts more CAR on the T-cell surface than both BBζ and Stoat and releases more cytokine. Across five donors, Happy Stoat matched BBζ's B-cell depletion at lower mRNA doses.

Canelé used an AI-agent-orchestrated rational design loop to refactor the CAR intracellular signaling domain, compressing the second-generation BBζ domain into a 54-amino-acid design called Stoat and then a 45-amino-acid variant, Happy Stoat. At the same mRNA dose, Happy Stoat puts more CAR on the T-cell surface than both BBζ and Stoat and releases more cytokine. Across five donors, Happy Stoat matched BBζ's B-cell depletion at lower mRNA doses.

Put a live user inside the negotiation loop and the agent bargaining for them gets better terms, while also saying more things that are not true. Prior negotiation evaluations represented the user with a static description of preferences and left the agent to bargain alone. Present compares the same game under matched conditions with and without a present user; across 7 games, 8 models and 11.2k runs, parties with a user earned 13.7% higher payouts on average, but counterparty-facing misalignment rose from 2.6% to 5.5% and user-facing misalignment appeared in 14.3% of games, uncorrelated with payout.

Put a live user inside the negotiation loop and the agent bargaining for them gets better terms, while also saying more things that are not true. Prior negotiation evaluations represented the user with a static description of preferences and left the agent to bargain alone. Present compares the same game under matched conditions with and without a present user; across 7 games, 8 models and 11.2k runs, parties with a user earned 13.7% higher payouts on average, but counterparty-facing misalignment rose from 2.6% to 5.5% and user-facing misalignment appeared in 14.3% of games, uncorrelated with payout.

On-policy distillation no longer treats the student's own samples as the only rollout distribution: previously each round sampled from the current student, forcing the teacher to supervise token by token on trajectories the student reaches but the teacher rarely does, so that supervision becomes noisy and degenerate patterns such as repetition are sampled again in later rounds. The new method reweights by relative teacher support, derives a geometric mixture of student and teacher inside a KL trust region, and trains a proposal policy to approximate that mixture in one forward pass. Across 16 math, knowledge QA and medical reasoning benchmarks, Qwen3-4B-Base gains 6.91 average accuracy over OPD on math while generating 41% shorter responses.

On-policy distillation no longer treats the student's own samples as the only rollout distribution: previously each round sampled from the current student, forcing the teacher to supervise token by token on trajectories the student reaches but the teacher rarely does, so that supervision becomes noisy and degenerate patterns such as repetition are sampled again in later rounds. The new method reweights by relative teacher support, derives a geometric mixture of student and teacher inside a KL trust region, and trains a proposal policy to approximate that mixture in one forward pass. Across 16 math, knowledge QA and medical reasoning benchmarks, Qwen3-4B-Base gains 6.91 average accuracy over OPD on math while generating 41% shorter responses.

Theory-of-mind benchmarks usually ask story questions about what others believe, and a correct answer does not show the model can use that judgment in a task. This evaluation extends household collaboration and calendar scheduling with asymmetric information and separates world understanding, information access, beliefs, decisions and execution into independent interfaces. On matched scheduling cases Claude recovers all agent beliefs yet produces subjectively optimal plans in only 67.9% of cases, while supplying beliefs raises Qwen-27B household task success from 55.0% to 75.0%.

Theory-of-mind benchmarks usually ask story questions about what others believe, and a correct answer does not show the model can use that judgment in a task. This evaluation extends household collaboration and calendar scheduling with asymmetric information and separates world understanding, information access, beliefs, decisions and execution into independent interfaces. On matched scheduling cases Claude recovers all agent beliefs yet produces subjectively optimal plans in only 67.9% of cases, while supplying beliefs raises Qwen-27B household task success from 55.0% to 75.0%.

Anthropic's Claude Platform release notes make the Claude Tag API available in beta to Claude Enterprise organizations, with /v1/organizations/claude_tag/ endpoints that read and change Claude Tag settings on Slack channels and workspaces. The endpoints show which settings apply in a channel and where each comes from, and can attach Access bundles; they require an Admin API key created by the primary owner and are not available to Claude Platform (Claude Console) organizations.

Anthropic's Claude Platform release notes make the Claude Tag API available in beta to Claude Enterprise organizations, with /v1/organizations/claude_tag/ endpoints that read and change Claude Tag settings on Slack channels and workspaces. The endpoints show which settings apply in a channel and where each comes from, and can attach Access bundles; they require an Admin API key created by the primary owner and are not available to Claude Platform (Claude Console) organizations.

The Sinkhorn divergence of entropic optimal transport is fast and statistically well behaved, but it is not a metric and has no Riemannian structure; the earlier tangent space and metric tensor came without covariant derivative, geodesic or parallel transport equations, and inference on them was open. Under finite-smoothness costs the metric tensor is shown Hadamard differentiable along curves of measures, which yields the Levi-Civita covariant derivative and turns geodesics and parallel transport into ODEs with smooth coefficients on a fixed RKHS. Linearizing those solutions with Banach-space ODE theory gives √n Gaussian limits for plug-in estimators of the exponential map, logarithmic map, parallel transport and squared Sinkhorn distance, with a consistent nonparametric bootstrap.

The Sinkhorn divergence of entropic optimal transport is fast and statistically well behaved, but it is not a metric and has no Riemannian structure; the earlier tangent space and metric tensor came without covariant derivative, geodesic or parallel transport equations, and inference on them was open. Under finite-smoothness costs the metric tensor is shown Hadamard differentiable along curves of measures, which yields the Levi-Civita covariant derivative and turns geodesics and parallel transport into ODEs with smooth coefficients on a fixed RKHS. Linearizing those solutions with Banach-space ODE theory gives √n Gaussian limits for plug-in estimators of the exponential map, logarithmic map, parallel transport and squared Sinkhorn distance, with a consistent nonparametric bootstrap.

In aboveground biomass estimation on the southeastern Tibetan Plateau, a spatial graph built from geographic nearest neighbors links valley forest to adjacent alpine meadow, because horizontal closeness does not imply ecological similarity. The study first quantifies these misconnections, then rebuilds the graph by keeping only edges whose endpoint elevation difference falls below a threshold, raising local-interpolation R² from 0.362 to 0.409 across 6,638 samples and ten seeds. Under a strict 30 km buffer holdout the same construction scores 0.054 lower than the ungated graph, so the gain depends on training neighbors remaining near the test points.

In aboveground biomass estimation on the southeastern Tibetan Plateau, a spatial graph built from geographic nearest neighbors links valley forest to adjacent alpine meadow, because horizontal closeness does not imply ecological similarity. The study first quantifies these misconnections, then rebuilds the graph by keeping only edges whose endpoint elevation difference falls below a threshold, raising local-interpolation R² from 0.362 to 0.409 across 6,638 samples and ten seeds. Under a strict 30 km buffer holdout the same construction scores 0.054 lower than the ungated graph, so the gain depends on training neighbors remaining near the test points.

As skill libraries grow, semantically similar skills become hard to tell apart from descriptions alone. SKALOS builds a multi-hierarchy capability graph of capability, method, artifact and constraint nodes, uses Max–Min witness learning to require that one implementation jointly covers every query requirement, and calibrates extraction errors with ranking feedback. On SkillRet its NDCG@10 is 2.68 points above a matched text-only reranker; on controlled ALFWorld tasks, witness-guided skill selection raises success from 22% to 29%.

As skill libraries grow, semantically similar skills become hard to tell apart from descriptions alone. SKALOS builds a multi-hierarchy capability graph of capability, method, artifact and constraint nodes, uses Max–Min witness learning to require that one implementation jointly covers every query requirement, and calibrates extraction errors with ranking feedback. On SkillRet its NDCG@10 is 2.68 points above a matched text-only reranker; on controlled ALFWorld tasks, witness-guided skill selection raises success from 22% to 29%.

In dense tiny-object detection, dynamic query designs previously entangled how many queries are decoded with which queries are kept, obscuring the source of gains. BUTR reframes dynamic query as allocating decoder query budget by object density: it sums encoder proposal confidences into a density proxy and maps it linearly to a per-image query count, keeping standard Top-k initialization without learned modules or heuristic filtering. On AI-TOD-v2, BUTR-L reaches 35.7 AP with 316 queries on average, versus 35.2 AP and 600 queries for same-scale D-FINE-L.

In dense tiny-object detection, dynamic query designs previously entangled how many queries are decoded with which queries are kept, obscuring the source of gains. BUTR reframes dynamic query as allocating decoder query budget by object density: it sums encoder proposal confidences into a density proxy and maps it linearly to a per-image query count, keeping standard Top-k initialization without learned modules or heuristic filtering. On AI-TOD-v2, BUTR-L reaches 35.7 AP with 316 queries on average, versus 35.2 AP and 600 queries for same-scale D-FINE-L.

Single-cell epigenome sequencing has moved its barcoding step off the plate and into droplets. Long-read single-cell ATAC and CUT&Tag previously ran well by well, needing at least a week of library work for about 5,000 cells at high cost. Now nuclei are co-encapsulated with barcoded beads on a commercial microfluidic chip and read on Nanopore: about 5,000 cells per run, one day of library prep instead of a week, and roughly $0.39 per cell instead of $1.98. In GM12878, 10,000 input nuclei produced 66 Gb, with 82.8% of reads on valid cell barcodes.

Single-cell epigenome sequencing has moved its barcoding step off the plate and into droplets. Long-read single-cell ATAC and CUT&Tag previously ran well by well, needing at least a week of library work for about 5,000 cells at high cost. Now nuclei are co-encapsulated with barcoded beads on a commercial microfluidic chip and read on Nanopore: about 5,000 cells per run, one day of library prep instead of a week, and roughly $0.39 per cell instead of $1.98. In GM12878, 10,000 input nuclei produced 66 Gb, with 82.8% of reads on valid cell barcodes.

In modular assembly the robot's task order must respect assembly precedence and be reordered when conditions change after work begins. Constraint-Aware Genetic Task Planning (CGTP) keeps initialization and variation inside the precedence-feasible region and, on replanning, freezes completed operations and re-optimizes only the remaining task graph. On an exhaustively verifiable eight-task benchmark it returned the exact best feasible suffix in all 128 structurally recoverable scenarios and flagged all 3 unrecoverable ones; on a 90-instance corpus, per-class recovery ran from 0.881 to 1.000 with no budget-conditional search failures.

In modular assembly the robot's task order must respect assembly precedence and be reordered when conditions change after work begins. Constraint-Aware Genetic Task Planning (CGTP) keeps initialization and variation inside the precedence-feasible region and, on replanning, freezes completed operations and re-optimizes only the remaining task graph. On an exhaustively verifiable eight-task benchmark it returned the exact best feasible suffix in all 128 structurally recoverable scenarios and flagged all 3 unrecoverable ones; on a 90-instance corpus, per-class recovery ran from 0.881 to 1.000 with no budget-conditional search failures.

Google released Gemini 4 Argon, positioned as its most intelligent model for long-horizon, multi-step professional work built around reasoning, coding and tool coordination. In the official examples it wrote 5,000 lines of math from scratch, with zero assets or libraries, to render a 3D sunlit library, and in Antigravity orchestrated subagents and external tools to turn more than 80,000 raw NASA files from Mars missions into an executive-ready deck. The model is now available to developers and enterprises through the API in AI Studio and the Gemini platform.

Google released Gemini 4 Argon, positioned as its most intelligent model for long-horizon, multi-step professional work built around reasoning, coding and tool coordination. In the official examples it wrote 5,000 lines of math from scratch, with zero assets or libraries, to render a 3D sunlit library, and in Antigravity orchestrated subagents and external tools to turn more than 80,000 raw NASA files from Mars missions into an executive-ready deck. The model is now available to developers and enterprises through the API in AI Studio and the Gemini platform.

Splitting the Enron email network into 44 thirty-day snapshots and matching maximal cliques one-to-one by Jaccard similarity leaves about 60% of cliques without a successor, so next-month tension is undefined for most rows. Over 24 rolling origins, residual GCN, GraphSAGE-style and GCN+GRU models that start at persistence are equivalent to carrying the current value forward on surviving cliques, a plain GCN does worse, and neighbourhood and recurrent memory add nothing measurable. In the deployable setting, multiplying persistence by a simple hazard-model survival probability lowers error by about 17%.

Splitting the Enron email network into 44 thirty-day snapshots and matching maximal cliques one-to-one by Jaccard similarity leaves about 60% of cliques without a successor, so next-month tension is undefined for most rows. Over 24 rolling origins, residual GCN, GraphSAGE-style and GCN+GRU models that start at persistence are equivalent to carrying the current value forward on surviving cliques, a plain GCN does worse, and neighbourhood and recurrent memory add nothing measurable. In the deployable setting, multiplying persistence by a simple hazard-model survival probability lowers error by about 17%.

Classifying small cell lung cancer histopathology requires resolving high-frequency nuclear texture and long-range tissue organisation at once, which fixed-kernel convolutions and pure transformers each handle poorly. SCLC-SBT puts frequency-domain channel recalibration in its first three stages and a dual convolutional and 2D selective state-space branch in its deepest stage, reaching 0.9697 accuracy, 0.9728 F1 and 0.9971 AUC on a 1,974-patch binary benchmark. Against the six strongest baselines the difference is not statistically significant on the 297-sample test set, so its clear advantage is competitive accuracy at about 17M parameters and 7.0G FLOPs.

Classifying small cell lung cancer histopathology requires resolving high-frequency nuclear texture and long-range tissue organisation at once, which fixed-kernel convolutions and pure transformers each handle poorly. SCLC-SBT puts frequency-domain channel recalibration in its first three stages and a dual convolutional and 2D selective state-space branch in its deepest stage, reaching 0.9697 accuracy, 0.9728 F1 and 0.9971 AUC on a 1,974-patch binary benchmark. Against the six strongest baselines the difference is not statistically significant on the 297-sample test set, so its clear advantage is competitive accuracy at about 17M parameters and 7.0G FLOPs.

Recursive QAOA re-measures at every variable elimination, so shot cost compounds along the recursion even though steps differ in difficulty. The work casts step-wise shot allocation in depth-1 recursive QAOA as a sequential decision problem: a hand-crafted rule keyed to local difficulty sets a baseline, and a tabular Double Q-learning agent learns residual corrections under a Lagrangian constraint while the elimination rule stays fixed. Under a fixed per-step cap, on 130 held-out policy-instance pairs the heuristic cuts total shots by about 23% and the learned policy by 36.1% versus uniform allocation, with an effective shots-per-success ratio of 0.676, and the gain holds at about 36.6% on problem sizes unseen in training.

Recursive QAOA re-measures at every variable elimination, so shot cost compounds along the recursion even though steps differ in difficulty. The work casts step-wise shot allocation in depth-1 recursive QAOA as a sequential decision problem: a hand-crafted rule keyed to local difficulty sets a baseline, and a tabular Double Q-learning agent learns residual corrections under a Lagrangian constraint while the elimination rule stays fixed. Under a fixed per-step cap, on 130 held-out policy-instance pairs the heuristic cuts total shots by about 23% and the learned policy by 36.1% versus uniform allocation, with an effective shots-per-success ratio of 0.676, and the gain holds at about 36.6% on problem sizes unseen in training.

Handwritten Meitei Mayek has been recognised one character at a time, yet its most frequent confusions depend on where a character sits in a word. Training words are composed from isolated character images and read whole by a convolutional encoder, a bidirectional LSTM and CTC, giving 0.24% character and 1.57% word error on 5,000 synthetic test words; on 100 unseen real words written on a tablet by one native writer, character error is 2.96% and word error 16.0%.

Handwritten Meitei Mayek has been recognised one character at a time, yet its most frequent confusions depend on where a character sits in a word. Training words are composed from isolated character images and read whole by a convolutional encoder, a bidirectional LSTM and CTC, giving 0.24% character and 1.57% word error on 5,000 synthetic test words; on 100 unseen real words written on a tablet by one native writer, character error is 2.96% and word error 16.0%.

Construction robots working near people must anticipate human motion, yet existing social-navigation methods rely on hand-crafted interaction rules, labeled demonstrations, or robot-specific control policies. This LiDAR-only framework generates residual targets automatically by constant-velocity extrapolation, models relations among robot, pedestrians and goal with a graph neural network, and uses an extended Kalman filter to turn frame-wise detections into dynamic graphs, outputting planner-agnostic path corrections. In a bricklaying trial it reduced path irregularity by 39.2% and raised minimum time-to-collision; across seven Dynablox sequences it raised average minimum time-to-collision by 38.3% with shorter, smoother paths than the baselines.

Construction robots working near people must anticipate human motion, yet existing social-navigation methods rely on hand-crafted interaction rules, labeled demonstrations, or robot-specific control policies. This LiDAR-only framework generates residual targets automatically by constant-velocity extrapolation, models relations among robot, pedestrians and goal with a graph neural network, and uses an extended Kalman filter to turn frame-wise detections into dynamic graphs, outputting planner-agnostic path corrections. In a bricklaying trial it reduced path irregularity by 39.2% and raised minimum time-to-collision; across seven Dynablox sequences it raised average minimum time-to-collision by 38.3% with shorter, smoother paths than the baselines.

Megapixel image synthesis has concentrated acceleration on denoising, yet cascade pipelines also spend heavily on repeated VAE transitions between latent and pixel space. UltraFlash targets those transitions, replacing the intermediate decode-resize-encode round trip with a learned latent shortcut and making final decoding sparse with smaller overlap. Built on PixelRush, its SD configuration cuts 4K latency from 16.37s to 8.04s while improving FID from 56.50 to 52.38.

Megapixel image synthesis has concentrated acceleration on denoising, yet cascade pipelines also spend heavily on repeated VAE transitions between latent and pixel space. UltraFlash targets those transitions, replacing the intermediate decode-resize-encode round trip with a learned latent shortcut and making final decoding sparse with smaller overlap. Built on PixelRush, its SD configuration cuts 4K latency from 16.37s to 8.04s while improving FID from 56.50 to 52.38.

Grammar assessment for international Chinese education is recast as proof-carrying constrained inference: retrieved rules, error spans, minimal repairs and ordinal grades are coupled through one latent evidence graph, and every score-affecting judgment must carry a rule-level support certificate. On public corpora including MuCGEC, PC2G reaches an error-detection F0.5 of 71.4 against 67.6 for the strongest listed retrieval baseline, CRAG-style, while five-level grade MAE falls from 0.57 to 0.42 and evidence faithfulness reaches 91.8%. Those grades are measured against a silver rubric derived from reference edits, not independent teacher ratings.

Grammar assessment for international Chinese education is recast as proof-carrying constrained inference: retrieved rules, error spans, minimal repairs and ordinal grades are coupled through one latent evidence graph, and every score-affecting judgment must carry a rule-level support certificate. On public corpora including MuCGEC, PC2G reaches an error-detection F0.5 of 71.4 against 67.6 for the strongest listed retrieval baseline, CRAG-style, while five-level grade MAE falls from 0.57 to 0.42 and evidence faithfulness reaches 91.8%. Those grades are measured against a silver rubric derived from reference edits, not independent teacher ratings.

An occupational task catalogue for the US economy now has language models interview workers directly instead of relying only on fixed questionnaires. Standardized surveys like O*NET are slow to catch tasks that have just appeared, and human-led interviews are expensive to run at scale. The interviewer asks follow-up questions to clarify vague descriptions and confirms the resulting task bundles with the worker; on a holdout of 1,184 task reports from 57 workers, TaskNet's lists covered 89.3% of reported tasks on average across six occupations, against 78.6% for O*NET.

An occupational task catalogue for the US economy now has language models interview workers directly instead of relying only on fixed questionnaires. Standardized surveys like O*NET are slow to catch tasks that have just appeared, and human-led interviews are expensive to run at scale. The interviewer asks follow-up questions to clarify vague descriptions and confirms the resulting task bundles with the worker; on a holdout of 1,184 task reports from 57 workers, TaskNet's lists covered 89.3% of reported tasks on average across six occupations, against 78.6% for O*NET.

Adding a morphology input channel to an organic-photovoltaic prediction pipeline that previously held only molecular descriptors raises the blended PCE prediction for the 3HT decamer:JSM5 blend from about 4.0% to about 4.5%, and the PCNN alone from about 2.4% to about 3.7%. The channel is a set of latent placeholder slots learned from device-target residuals and filled at prediction time with molecular-dynamics morphology descriptors; because the latents are target-derived, the shift reflects the pipeline's response to MD inputs and cannot be attributed to a physical morphology effect.

Adding a morphology input channel to an organic-photovoltaic prediction pipeline that previously held only molecular descriptors raises the blended PCE prediction for the 3HT decamer:JSM5 blend from about 4.0% to about 4.5%, and the PCNN alone from about 2.4% to about 3.7%. The channel is a set of latent placeholder slots learned from device-target residuals and filled at prediction time with molecular-dynamics morphology descriptors; because the latents are target-derived, the shift reflects the pipeline's response to MD inputs and cannot be attributed to a physical morphology effect.

Local video stylization used to work in two steps: a mask marked where to edit, and a separate text prompt named the style, leaving the model to link the two. GRNStyle, built on Generative Refinement Networks, instead writes each style into the mask as a fixed polar direction of equal length, so region and style arrive as one conditioning field. Trained on 100K pairs built from global stylization, object removal, matting and compositing, it reaches a Style Prototype Similarity of 0.705 for background editing and 0.407 for foreground editing on the 168-case LociArt Benchmark, the highest in the comparison.

Local video stylization used to work in two steps: a mask marked where to edit, and a separate text prompt named the style, leaving the model to link the two. GRNStyle, built on Generative Refinement Networks, instead writes each style into the mask as a fixed polar direction of equal length, so region and style arrive as one conditioning field. Trained on 100K pairs built from global stylization, object removal, matting and compositing, it reaches a Style Prototype Similarity of 0.705 for background editing and 0.407 for foreground editing on the 168-case LociArt Benchmark, the highest in the comparison.

LangChain has published how it built Content Studio, its in-house design agent: the model could already write HTML, CSS and animation timelines but lacked the company's own design judgment, so output drifted toward familiar patterns. The team wrote designers' principles, brand rules and the reasoning behind feedback into skills and a wiki, and gave the agent a sandbox computer that keeps a project across a Slack thread plus asynchronous subagents. The agent has delivered more than 3,000 assets, and median model spend per request fell from about $8 in mid-September to about 50 cents.

LangChain has published how it built Content Studio, its in-house design agent: the model could already write HTML, CSS and animation timelines but lacked the company's own design judgment, so output drifted toward familiar patterns. The team wrote designers' principles, brand rules and the reasoning behind feedback into skills and a wiki, and gave the agent a sandbox computer that keeps a project across a Slack thread plus asynchronous subagents. The agent has delivered more than 3,000 assets, and median model spend per request fell from about $8 in mid-September to about 50 cents.

Marine drill cores no longer need a geologist's rock name as a training target. The usual practice treats shipboard lithology as the ground truth a model should reproduce, yet the six measurements logged on every core (bulk density, magnetic susceptibility, natural gamma radiation and three reflectance channels) are the properties geologists read when they assign that name. A masked variational autoencoder trained on 236,963 observations from 296 boreholes in the LILY database, without labels, separates lithologies as well as a semi-supervised variant (adjusted Rand index 0.237 against 0.247), and raising the label weight a thousand-fold only drops reconstruction from 0.999 to 0.836 while classification barely moves.

Marine drill cores no longer need a geologist's rock name as a training target. The usual practice treats shipboard lithology as the ground truth a model should reproduce, yet the six measurements logged on every core (bulk density, magnetic susceptibility, natural gamma radiation and three reflectance channels) are the properties geologists read when they assign that name. A masked variational autoencoder trained on 236,963 observations from 296 boreholes in the LILY database, without labels, separates lithologies as well as a semi-supervised variant (adjusted Rand index 0.237 against 0.247), and raising the label weight a thousand-fold only drops reconstruction from 0.999 to 0.836 while classification barely moves.

A silicon-photonic matrix-vector multiplication engine is designed around mode-division multiplexing, using the fundamental and first-order transverse magnetic modes as two parallel computing channels and weighting signals through the amorphous-to-crystalline contrast of a Ge2Sb2Se4Te1 phase-change material. Weights persist without static power, the 2×2 optical core measures 185 µm × 21 µm, and total insertion loss at 1550 nm is 2.76 dB for TM0 and 3.64 dB for TM1 across a 40 nm band from 1540 to 1580 nm. All results come from optical and electro-thermal numerical simulation and have not been experimentally validated.

A silicon-photonic matrix-vector multiplication engine is designed around mode-division multiplexing, using the fundamental and first-order transverse magnetic modes as two parallel computing channels and weighting signals through the amorphous-to-crystalline contrast of a Ge2Sb2Se4Te1 phase-change material. Weights persist without static power, the 2×2 optical core measures 185 µm × 21 µm, and total insertion loss at 1550 nm is 2.76 dB for TM0 and 3.64 dB for TM1 across a 40 nm band from 1540 to 1580 nm. All results come from optical and electro-thermal numerical simulation and have not been experimentally validated.

Long-tailed recognition no longer treats every class alike. Prior methods pulled all features onto one fixed equiangular tight frame, a shared target whose directions make equal angles, and that full collapse over-contracts tail samples and weakens generalization. Head classes still collapse onto the frame while tail classes keep within-class diversity: a feature-conditioned gating classifier specializes nodes so tail classes take their own directions, and a frequency-adaptive margin shrinks for head classes and grows for tail classes. On CIFAR100-LT, ImageNet-LT, Places-LT and iNaturalist-2018, overall accuracy beats prior methods, and within-class spread rises monotonically from head to tail.

Long-tailed recognition no longer treats every class alike. Prior methods pulled all features onto one fixed equiangular tight frame, a shared target whose directions make equal angles, and that full collapse over-contracts tail samples and weakens generalization. Head classes still collapse onto the frame while tail classes keep within-class diversity: a feature-conditioned gating classifier specializes nodes so tail classes take their own directions, and a frequency-adaptive margin shrinks for head classes and grows for tail classes. On CIFAR100-LT, ImageNet-LT, Places-LT and iNaturalist-2018, overall accuracy beats prior methods, and within-class spread rises monotonically from head to tail.

Helical released a Model Context Protocol (MCP) integration for OpenAI's Codex that lets researchers call biological foundation models from the agentic tools they already use, and use or post-train those models on their own data through natural-language instructions. Using such models previously required researchers to configure infrastructure and build the workflows for an experiment. One model available at launch is PerturbGen, developed at the Wellcome Sanger Institute to predict how perturbations reshape cellular trajectories, and Helical with Nebius is offering free compute credits that academic researchers can request.

Helical released a Model Context Protocol (MCP) integration for OpenAI's Codex that lets researchers call biological foundation models from the agentic tools they already use, and use or post-train those models on their own data through natural-language instructions. Using such models previously required researchers to configure infrastructure and build the workflows for an experiment. One model available at launch is PerturbGen, developed at the Wellcome Sanger Institute to predict how perturbations reshape cellular trajectories, and Helical with Nebius is offering free compute credits that academic researchers can request.

A consortium of 18 institutions, C-BRAIN, has begun feeding Alzheimer's data that AI tools and human researchers could not previously reach, including unpublished findings, negative results and governed proprietary data, into model training. That data had stayed inside individual institutions, where it could not be synthesized into usable knowledge. A hybrid centralized and federated design lets models learn where the data sits, and two open-source agentic tools have already come out of it; in published case studies OpenScientist finished in minutes analyses that domain scientists estimated would take weeks to months, at under $10 per run.

A consortium of 18 institutions, C-BRAIN, has begun feeding Alzheimer's data that AI tools and human researchers could not previously reach, including unpublished findings, negative results and governed proprietary data, into model training. That data had stayed inside individual institutions, where it could not be synthesized into usable knowledge. A hybrid centralized and federated design lets models learn where the data sits, and two open-source agentic tools have already come out of it; in published case studies OpenScientist finished in minutes analyses that domain scientists estimated would take weeks to months, at under $10 per run.

Knowledge distillation for driving vision-language models previously matched teacher distributions uniformly across answer tokens, diluting planning-relevant semantics and passing on confident teacher errors. RSKD first localizes task-critical phrases in the ground-truth answer, such as traffic participants, states and spatial relations, then uses a frozen sentence encoder to score how semantically consistent the teacher's prediction is with each phrase, so consistent positions take more teacher-distribution supervision and inconsistent ones take more ground-truth supervision. On DrivingVQA, LingoQA and MapLMv2, RSKD reaches 16.60, 0.672 and 81.82, beating the compared SFT, GKD and ToDi student baselines on the two QA benchmarks and nearly matching SFT's 81.85 on MapLMv2.

Knowledge distillation for driving vision-language models previously matched teacher distributions uniformly across answer tokens, diluting planning-relevant semantics and passing on confident teacher errors. RSKD first localizes task-critical phrases in the ground-truth answer, such as traffic participants, states and spatial relations, then uses a frozen sentence encoder to score how semantically consistent the teacher's prediction is with each phrase, so consistent positions take more teacher-distribution supervision and inconsistent ones take more ground-truth supervision. On DrivingVQA, LingoQA and MapLMv2, RSKD reaches 16.60, 0.672 and 81.82, beating the compared SFT, GKD and ToDi student baselines on the two QA benchmarks and nearly matching SFT's 81.85 on MapLMv2.

Chinese regions use separate support policies to compete for one-person companies, solo founders who work through digital platforms and AI agents and can compare locations, and the two sides used to act on their own separate estimates. PolicyMediator turns that into a cognition-constrained leader-multi-follower Stackelberg framework: a mediator agent explains verified policies to the companies to build awareness and trust, then reports their needs and unresolved concerns to governments for policy learning. Across six-region simulations in China with five OPC types, five scenarios and three simulator backbones, it ranks first in all 15 combinations, with the aggregate government objective 42.4% above the eight-baseline mean.

Chinese regions use separate support policies to compete for one-person companies, solo founders who work through digital platforms and AI agents and can compare locations, and the two sides used to act on their own separate estimates. PolicyMediator turns that into a cognition-constrained leader-multi-follower Stackelberg framework: a mediator agent explains verified policies to the companies to build awareness and trust, then reports their needs and unresolved concerns to governments for policy learning. Across six-region simulations in China with five OPC types, five scenarios and three simulator backbones, it ranks first in all 15 combinations, with the aggregate government objective 42.4% above the eight-baseline mean.

At the 2026 OCP Global Summit, Meta published the network designs behind its latest AI models and its gigawatt-scale Prometheus cluster. Scale-out had been organized as a three-tier Clos, and the switch platform layer stayed proprietary, with each new platform repeating months of bring-up work. Meta now connects each GPU to eight network planes and flattens Clos to two tiers, and shows IcePack, its first liquid-cooled network rack, holding IceCube, a 102.4T switch with 128 800G ports; MetaRoCE goes open source and the OSPA 1.0 specification goes to OCP.

At the 2026 OCP Global Summit, Meta published the network designs behind its latest AI models and its gigawatt-scale Prometheus cluster. Scale-out had been organized as a three-tier Clos, and the switch platform layer stayed proprietary, with each new platform repeating months of bring-up work. Meta now connects each GPU to eight network planes and flattens Clos to two tiers, and shows IcePack, its first liquid-cooled network rack, holding IceCube, a 102.4T switch with 128 800G ports; MetaRoCE goes open source and the OSPA 1.0 specification goes to OCP.

RGB-D salient object detection has traded the limited receptive field of CNNs against the quadratic cost of Transformers; Mamba models global context at linear cost but flattens 2D features into 1D sequences and loses local structure. LWNet adds a Vision Liquid Time-Constant module on a VMamba backbone, using input-dependent decay to restore detail weakened by scanning, and a cross-modal wavelet fusion that splits RGB and depth features into low- and high-frequency bands for separate alignment. Against 18 methods on STERE, NLPR, NJU2K, DUTLF and SIP, LWNet ranks first on all five datasets.

RGB-D salient object detection has traded the limited receptive field of CNNs against the quadratic cost of Transformers; Mamba models global context at linear cost but flattens 2D features into 1D sequences and loses local structure. LWNet adds a Vision Liquid Time-Constant module on a VMamba backbone, using input-dependent decay to restore detail weakened by scanning, and a cross-modal wavelet fusion that splits RGB and depth features into low- and high-frequency bands for separate alignment. Against 18 methods on STERE, NLPR, NJU2K, DUTLF and SIP, LWNet ranks first on all five datasets.

Anthropic released Claude for Financial Advisors, a plugin that connects Claude to the custodians, CRM, portfolio reporting, planning, estate and tax tools advisors already use, and adds skills for meeting prep, rebalance review, post-meeting notes and follow-up, and compliance checks. Anthropic cites Kitces research that a typical advisory practice spends only about a sixth of its time in client meetings, with the rest going to cross-system preparation and documentation. The plugin is available today, and regulated activities such as investment recommendations, client communications and compliance determinations remain subject to human review and approval.

Anthropic released Claude for Financial Advisors, a plugin that connects Claude to the custodians, CRM, portfolio reporting, planning, estate and tax tools advisors already use, and adds skills for meeting prep, rebalance review, post-meeting notes and follow-up, and compliance checks. Anthropic cites Kitces research that a typical advisory practice spends only about a sixth of its time in client meetings, with the rest going to cross-system preparation and documentation. The plugin is available today, and regulated activities such as investment recommendations, client communications and compliance determinations remain subject to human review and approval.

TradingAgents 0.7.0 lets users run the framework on their own ChatGPT Plus or Pro subscription and scores backtests from the first close after each decision. Earlier versions let backtests see only data published by each analysis date, so the numbers stopped at the analysis day. The framework splits a trading firm's roles across specialized LLM-driven agents, from fundamental, sentiment and technical analysts to a trader and a risk management team; 0.7.0 also notes at start-up when a newer release is out and ships a wheel with each GitHub release.

TradingAgents 0.7.0 lets users run the framework on their own ChatGPT Plus or Pro subscription and scores backtests from the first close after each decision. Earlier versions let backtests see only data published by each analysis date, so the numbers stopped at the analysis day. The framework splits a trading firm's roles across specialized LLM-driven agents, from fundamental, sentiment and technical analysts to a trader and a risk management team; 0.7.0 also notes at start-up when a newer release is out and ships a wheel with each GitHub release.

Genspark has launched Media Studio, an AI studio that moves video production from a pipeline of script, cast, sets, voice, score and editor into one workspace run by an agent director. A user describes the piece and answers a few questions; the agent settles concept, script, storyboard and budget first, then orchestrates image, video, audio and 3D generation models from more than a dozen providers to generate, check quality, edit and composite. Output ranges from a 15-second ad to a 10-minute short drama, and from a poster to an interactive film.

Genspark has launched Media Studio, an AI studio that moves video production from a pipeline of script, cast, sets, voice, score and editor into one workspace run by an agent director. A user describes the piece and answers a few questions; the agent settles concept, script, storyboard and budget first, then orchestrates image, video, audio and 3D generation models from more than a dozen providers to generate, check quality, edit and composite. Output ranges from a 15-second ad to a 10-minute short drama, and from a poster to an interactive film.

Anthropic released the Claude for Nonprofits plugin, a set of skills and connectors built with about 30 nonprofits, covering organization mapping, AI-policy checks and the fundraising cycle. Before this, what Anthropic offered nonprofits was price relief and data connections, and nonprofits said cost was only one barrier, the other two being too few people with AI skills and too few tools built for how nonprofits operate. Anthropic also plans to extend Claude Corps, its program placing early-career fellows in US nonprofits, through a partnership with Taproot.

Anthropic released the Claude for Nonprofits plugin, a set of skills and connectors built with about 30 nonprofits, covering organization mapping, AI-policy checks and the fundraising cycle. Before this, what Anthropic offered nonprofits was price relief and data connections, and nonprofits said cost was only one barrier, the other two being too few people with AI skills and too few tools built for how nonprofits operate. Anthropic also plans to extend Claude Corps, its program placing early-career fellows in US nonprofits, through a partnership with Taproot.

Neural solvers for the damped nonlinear elastic wave equation now train differently: Lie symmetry analysis reduces it to four ordinary differential equations in a single similarity variable, the network learns the reduced unknown, a symmetry loss constrains invariance, and residuals are weighted in time so early times are learned first. A standard physics-informed network relies only on the residual and boundary conditions, and on time-dependent nonlinear systems it tends to hit gradient imbalance and slow convergence. Across four cases, profiles, surfaces and contours stay closer to the analytical solutions of the reduced equations, and the generalized damped case shows a 1.53× total wall-clock speed-up including derivation and training.

Neural solvers for the damped nonlinear elastic wave equation now train differently: Lie symmetry analysis reduces it to four ordinary differential equations in a single similarity variable, the network learns the reduced unknown, a symmetry loss constrains invariance, and residuals are weighted in time so early times are learned first. A standard physics-informed network relies only on the residual and boundary conditions, and on time-dependent nonlinear systems it tends to hit gradient imbalance and slow convergence. Across four cases, profiles, surfaces and contours stay closer to the analytical solutions of the reduced equations, and the generalized damped case shows a 1.53× total wall-clock speed-up including derivation and training.

Vercel's observability console now lets teams define their own alerting monitors from their own query results, not only from the built-in alerts that Observability Plus has provided. Each monitor tracks a single metric, such as response time, or a ratio of two, such as errors divided by total requests, and uses either statistical anomaly detection or a fixed threshold with a configurable detection window, filters and grouping. Monitors can be started from a chart query, a preset or the team's alerts settings, or defined in JSON and added with the CLI; alerts go to email, Slack or team webhooks, and the feature requires Observability Plus on Pro or Enterprise plans.

Vercel's observability console now lets teams define their own alerting monitors from their own query results, not only from the built-in alerts that Observability Plus has provided. Each monitor tracks a single metric, such as response time, or a ratio of two, such as errors divided by total requests, and uses either statistical anomaly detection or a fixed threshold with a configurable detection window, filters and grouping. Monitors can be started from a chart query, a preset or the team's alerts settings, or defined in JSON and added with the CLI; alerts go to email, Slack or team webhooks, and the feature requires Observability Plus on Pro or Enterprise plans.