AI and science frontiers · 2026-10-13

The humanoid Reflex catches human-thrown boxes using only onboard RGB-D vision, reaching 85.7% catch success in simulation versus 87.9% with privileged box observations, and 65% and 40% on the real robot at 2.5 and 5 m.

The humanoid Reflex catches human-thrown boxes using only onboard RGB-D vision, reaching 85.7% catch success in simulation versus 87.9% with privileged box observations, and 65% and 40% on the real robot at 2.5 and 5 m.

Chaining machine-learning potentials, accelerated sampling and mass thermodynamic integration into one workflow brings water's autoionization constant to 14.31 against the experimental 13.93 once nuclear quantum effects are included explicitly, and the same workflow computes all six dissociation constants of phosphoric acid without changing the order-parameter definitions, cutting the mean absolute error from 2.19 to 0.45.

Chaining machine-learning potentials, accelerated sampling and mass thermodynamic integration into one workflow brings water's autoionization constant to 14.31 against the experimental 13.93 once nuclear quantum effects are included explicitly, and the same workflow computes all six dissociation constants of phosphoric acid without changing the order-parameter definitions, cutting the mean absolute error from 2.19 to 0.45.

An open benchmark evaluates five public PPG foundation models across nine datasets, 26 tasks and four downstream adaptation regimes under one protocol, finding that 35.3% of 68 task-regime combinations show no meaningful gain over a random or constant baseline and 83.8% show no PPG-specific model beating both classical pulse-rate-variability features and a generic time-series foundation model.

An open benchmark evaluates five public PPG foundation models across nine datasets, 26 tasks and four downstream adaptation regimes under one protocol, finding that 35.3% of 68 task-regime combinations show no meaningful gain over a random or constant baseline and 83.8% show no PPG-specific model beating both classical pulse-rate-variability features and a generic time-series foundation model.

On an 11-bit parity task, the Boolean function state learned by a Transformer first differentiates and then recoheres as data accumulate, while the connectivity of the training-set constraint graph rises monotonically, showing that the learned state keeps reorganizing after the identifiable support stops changing.

On an 11-bit parity task, the Boolean function state learned by a Transformer first differentiates and then recoheres as data accumulate, while the connectivity of the training-set constraint graph rises monotonically, showing that the learned state keeps reorganizing after the identifiable support stops changing.

Injecting random two-bit weight faults into agent executions turns model-level corruption into system-level outcomes: agents recover from malformed tool calls and still finish tasks, yet faulted runs raise mean output tokens by 38.0%, and some faults shift security and alignment behavior.

Injecting random two-bit weight faults into agent executions turns model-level corruption into system-level outcomes: agents recover from malformed tool calls and still finish tasks, yet faulted runs raise mean output tokens by 38.0%, and some faults shift security and alignment behavior.

DRYLAB-Bench puts 15 frontier models through 30 hazardous mutation-design scenarios and scores submitted amino-acid substitutions against published wet-lab measurements: 21.4% land in each assay's highest-risk decile without tools, rising to 23.5% with biological design tools and multi-round iteration.

DRYLAB-Bench puts 15 frontier models through 30 hazardous mutation-design scenarios and scores submitted amino-acid substitutions against published wet-lab measurements: 21.4% land in each assay's highest-risk decile without tools, rising to 23.5% with biological design tools and multi-round iteration.

On long-context tasks, self-consistency, the practice of sampling several answers and aggregating them, no longer improves performance and often lowers it, because positional bias makes the samples repeat the same mistake and the vote turns that mistake into a consensus.

On long-context tasks, self-consistency, the practice of sampling several answers and aggregating them, no longer improves performance and often lowers it, because positional bias makes the samples repeat the same mistake and the vote turns that mistake into a consensus.

PlacePulse compresses 1,055,537 US H3 resolution-7 cells into 256-dimensional unit-norm vectors using only static structural attributes such as demographics, businesses and environmental risk, beating a dimensionality-matched feature-selection baseline on 74 external targets and leading Google's Population Dynamics Foundation Model on health and economic outcomes.

PlacePulse compresses 1,055,537 US H3 resolution-7 cells into 256-dimensional unit-norm vectors using only static structural attributes such as demographics, businesses and environmental risk, beating a dimensionality-matched feature-selection baseline on 74 external targets and leading Google's Population Dynamics Foundation Model on health and economic outcomes.

Across four reasoners and six math benchmarks, injecting a correction instruction as a new user turn at a learned cut position raises average accuracy from 37.1% to 55.3% while the reasoner itself stays frozen.

Across four reasoners and six math benchmarks, injecting a correction instruction as a new user turn at a learned cut position raises average accuracy from 37.1% to 55.3% while the reasoner itself stays frozen.

The preconditioner memory of kernel ridge regression no longer grows with model size, letting Bloks scale to ten million parameters and match or beat state-of-the-art accuracy on large-scale benchmarks such as Higgs.

The preconditioner memory of kernel ridge regression no longer grows with model size, letting Bloks scale to ten million parameters and match or beat state-of-the-art accuracy on large-scale benchmarks such as Higgs.

On AlgoPerf v1's self-tuning track, where 13 submissions each had to run all nine workloads with a single fixed hyperparameter configuration, Schedule-Free AdamW v2 (PyTorch) ranked first on wall-clock time-to-target, indicating that schedule-free iterate averaging generalizes across heterogeneous architectures without per-task tuning.

On AlgoPerf v1's self-tuning track, where 13 submissions each had to run all nine workloads with a single fixed hyperparameter configuration, Schedule-Free AdamW v2 (PyTorch) ranked first on wall-clock time-to-target, indicating that schedule-free iterate averaging generalizes across heterogeneous architectures without per-task tuning.

Multimodal LLMs often report no text in images whose coarse structure carries a string a human recovers by squinting, and a training method that reads the original image alone lifts transcription from 26.50 to 88.61 on production moderation images.

Multimodal LLMs often report no text in images whose coarse structure carries a string a human recovers by squinting, and a training method that reads the original image alone lifts transcription from 26.50 to 88.61 on production moderation images.

A two-dimensional grain-growth simulation hands area evolution to the von Neumann–Mullins law and microstructure reconstruction to a graph neural network, reaching the same coarsening stage on five initial microstructures in 38.43 s on average versus 5,247.01 s for phase-field.

A two-dimensional grain-growth simulation hands area evolution to the von Neumann–Mullins law and microstructure reconstruction to a graph neural network, reaching the same coarsening stage on five initial microstructures in 38.43 s on average versus 5,247.01 s for phase-field.

Speech recognition decoding no longer has to wait token by token: a lightweight draft module that reuses the target model's acoustic and decoder features delivers up to 1.9x and 2.4x lossless speedups on Whisper and Conformer.

Speech recognition decoding no longer has to wait token by token: a lightweight draft module that reuses the target model's acoustic and decoder features delivers up to 1.9x and 2.4x lossless speedups on Whisper and Conformer.

Deleting the waveform region an audio-language model's claim points to overturns its "sound present" answer more often than deleting an equally sized random window, indicating that these hallucinations are at least partly grounded in locatable audio evidence, and the same contrast lifts Qwen2-Audio accuracy from 57.1% to 71.9%.

Deleting the waveform region an audio-language model's claim points to overturns its "sound present" answer more often than deleting an equally sized random window, indicating that these hallucinations are at least partly grounded in locatable audio evidence, and the same contrast lifts Qwen2-Audio accuracy from 57.1% to 71.9%.

A Bayes free energy computed from trained final-layer weights alone tracks model quality across convolutional networks and LLMs, and under a heavy-tailed spectral model it yields an optimal heaviness whose crossing marks the onset of overfitting.

A Bayes free energy computed from trained final-layer weights alone tracks model quality across convolutional networks and LLMs, and under a heavy-tailed spectral model it yields an optimal heaviness whose crossing marks the onset of overfitting.

Anisotropic hyperelastic homogenization of periodic cells now learns its directional features instead of fixing them, and six such features cut the stress error about sevenfold while keeping the constrained energies solvable beyond the data.

Anisotropic hyperelastic homogenization of periodic cells now learns its directional features instead of fixing them, and six such features cut the stress error about sevenfold while keeping the constrained energies solvable beyond the data.

A new backdoor makes a compromised model degrade its reasoning in four ordered stages as benign trigger words accumulate, going wrong before the final answer does, and existing defenses do not fully break it.

A new backdoor makes a compromised model degrade its reasoning in four ordered stages as benign trigger words accumulate, going wrong before the final answer does, and existing defenses do not fully break it.

Rubric construction now filters candidate decompositions by semantics first and then picks the final structure from how criteria behave on six responses of differing quality, reaching the highest pairwise accuracy on five of six preference benchmarks.

Rubric construction now filters candidate decompositions by semantics first and then picks the final structure from how criteria behave on six responses of differing quality, reaching the highest pairwise accuracy on five of six preference benchmarks.

A meta model pre-trained offline on synthetic Wiener-Hammerstein systems predicts a new system's output zero-shot from a short input-output context, matches state-of-the-art accuracy on nonlinear identification benchmarks, and cuts an out-of-distribution continuous stirred tank reactor's prediction error from 1.249 to 0.242 by adapting only the input side.

A meta model pre-trained offline on synthetic Wiener-Hammerstein systems predicts a new system's output zero-shot from a short input-output context, matches state-of-the-art accuracy on nonlinear identification benchmarks, and cuts an out-of-distribution continuous stirred tank reactor's prediction error from 1.249 to 0.242 by adapting only the input side.

During pretraining of Pythia language models, word-sense disambiguation improves in sharp steps around 1,000 to 2,000 training steps, two layer-3 attention heads in 14M track that improvement and their ablation measurably hurts it, while 410M spreads the function across more heads that hold up better under position and part-of-speech perturbations.

During pretraining of Pythia language models, word-sense disambiguation improves in sharp steps around 1,000 to 2,000 training steps, two layer-3 attention heads in 14M track that improvement and their ablation measurably hurts it, while 410M spreads the function across more heads that hold up better under position and part-of-speech perturbations.

Whether a generative model can be edited like a simulator depends on the environment: editability falls smoothly as Othello is simplified, while in the continuous disc world Rayworld probes fail to steer along the state yet an inverse map still sets the model from it.

Whether a generative model can be edited like a simulator depends on the environment: editability falls smoothly as Othello is simplified, while in the continuous disc world Rayworld probes fail to steer along the state yet an inverse map still sets the model from it.

NucleoNet evolves nuclei and electrons together in one network, so a single parameter set describes the N2 dissociation curve, Li displacement in LiCN, and the ethane and formamide bond-cleavage paths, matching established neural-wavefunction accuracy on equilibrium small molecules with far fewer parameters.

NucleoNet evolves nuclei and electrons together in one network, so a single parameter set describes the N2 dissociation curve, Li displacement in LiCN, and the ethane and formamide bond-cleavage paths, matching established neural-wavefunction accuracy on equilibrium small molecules with far fewer parameters.

Multi-turn LLM agent training now ties exploration and credit assignment to one hindsight score: turns that lower the reference-answer likelihood are revisited and re-branched from the same parent, and inverse weights fold the outcomes back into the original policy, giving the highest aggregate exact match across seven QA benchmarks and three Qwen backbones.

Multi-turn LLM agent training now ties exploration and credit assignment to one hindsight score: turns that lower the reference-answer likelihood are revisited and re-branched from the same parent, and inverse weights fold the outcomes back into the original policy, giving the highest aggregate exact match across seven QA benchmarks and three Qwen backbones.

In a 2048-GPU, renewable-direct-supply AI data center, embedding checkpoint waiting and feasible 3D-parallel GPU modes of LLM pre-training into day-ahead dispatch cuts operating cost by 20.02% and renewable curtailment from 1.077 to 0.014 MWh relative to fixed continuous training.

In a 2048-GPU, renewable-direct-supply AI data center, embedding checkpoint waiting and feasible 3D-parallel GPU modes of LLM pre-training into day-ahead dispatch cuts operating cost by 20.02% and renewable curtailment from 1.077 to 0.014 MWh relative to fixed continuous training.

Memory retrieval no longer compares only whole-memory vectors: sparse matching over token states a frozen encoder already computes, fused with whole-memory similarity, raises average Recall@10 on Qwen3-Embedding-4B from 62.5 to 75.7 without additional training.

Memory retrieval no longer compares only whole-memory vectors: sparse matching over token states a frozen encoder already computes, fused with whole-memory similarity, raises average Recall@10 on Qwen3-Embedding-4B from 62.5 to 75.7 without additional training.

The end-to-end planner ManeuverFlow organizes candidate trajectories into eight motion maneuvers, generates them in one step, and scores and locally refines them with shared future BEV evidence, reaching 90.58 PDMS and 90.14 EPDMS on NAVSIM v1/v2 navtest.

The end-to-end planner ManeuverFlow organizes candidate trajectories into eight motion maneuvers, generates them in one step, and scores and locally refines them with shared future BEV evidence, reaching 90.58 PDMS and 90.14 EPDMS on NAVSIM v1/v2 navtest.

Topic models can describe topics with interpretable directions learned by sparse autoencoders instead of word lists, and an LLM-based topic judge prefers that representation on eight datasets, with the largest margins on short and abstract texts.

Topic models can describe topics with interpretable directions learned by sparse autoencoders instead of word lists, and an LLM-based topic judge prefers that representation on eight datasets, with the largest margins on short and abstract texts.

SAM3 segments lesions from text or image-exemplar concept prompts across 13 datasets and 11 lesion types, beating SAM2 on Dice in 10 datasets, by 4.41 on ISLES22 and 5.39 on ISIC2018.

SAM3 segments lesions from text or image-exemplar concept prompts across 13 datasets and 11 lesion types, beating SAM2 on Dice in 10 datasets, by 4.41 on ISLES22 and 5.39 on ISIC2018.

Across 2,160 sessions of a modified rock-paper-scissors game, six frontier LLMs became insensitive to changing outcomes once a rule was supplied, and two sentences of direct tracking induction in the system prompt significantly reduced that insensitivity in all six.

Across 2,160 sessions of a modified rock-paper-scissors game, six frontier LLMs became insensitive to changing outcomes once a rule was supplied, and two sentences of direct tracking induction in the system prompt significantly reduced that insensitivity in all six.

Chest X-ray report generation moves to question-answer output: predicted disease labels align image and text representations, a medical knowledge graph supplies retrieved evidence, and a fine-tuned LLM rewrites the result, outperforming prior methods on most metrics on MIMIC-CXR and IU X-Ray.

Chest X-ray report generation moves to question-answer output: predicted disease labels align image and text representations, a medical knowledge graph supplies retrieved evidence, and a fine-tuned LLM rewrites the result, outperforming prior methods on most metrics on MIMIC-CXR and IU X-Ray.

Terminal-agent training data shifts from short command demonstrations to long-horizon trajectories averaging 52.8 turns, and fine-tuning Qwen3.5-27B on roughly 3K such trajectories raises Terminal-Bench 2.1 from 46.82 to 60.67 while beating the base model on all 12 benchmarks.

Terminal-agent training data shifts from short command demonstrations to long-horizon trajectories averaging 52.8 turns, and fine-tuning Qwen3.5-27B on roughly 3K such trajectories raises Terminal-Bench 2.1 from 46.82 to 60.67 while beating the base model on all 12 benchmarks.

QRAFT splits fact-checking article writing into evidence gathering, drafting and a simulated editorial review run by multiple LLM agents, beating prior text-generation methods on automatic metrics across 1,227 examples while still scoring well below expert-written articles with professional fact-checkers.

QRAFT splits fact-checking article writing into evidence gathering, drafting and a simulated editorial review run by multiple LLM agents, beating prior text-generation methods on automatic metrics across 1,227 examples while still scoring well below expert-written articles with professional fact-checkers.

Feeding chromatin accessibility into a sequence model at nucleotide resolution raised CRISPR off-target AUC-PR above the sequence-only version on all three held-out test sets, topping out at 0.587.

Feeding chromatin accessibility into a sequence model at nucleotide resolution raised CRISPR off-target AUC-PR above the sequence-only version on all three held-out test sets, topping out at 0.587.

An agentic generative-AI framework produced disease concept diagrams and model structure diagrams for breast cancer and hATTR-PN that were structurally concordant with the core disease-staging logic of the NICE appraisal accepted for each indication, with automated runs of about 1 hour, or 1–2 days including expert validation, against up to 3 months of fully manual conceptualisation.

An agentic generative-AI framework produced disease concept diagrams and model structure diagrams for breast cancer and hATTR-PN that were structurally concordant with the core disease-staging logic of the NICE appraisal accepted for each indication, with automated runs of about 1 hour, or 1–2 days including expert validation, against up to 3 months of fully manual conceptualisation.

In partially observable reinforcement learning, freezing the recurrent memory module and training only the output and mixer layers still matches fully trained counterparts on delayed recall, discrete and continuous control, and pixel-based tasks, while cutting trainable parameters by more than half.

In partially observable reinforcement learning, freezing the recurrent memory module and training only the output and mixer layers still matches fully trained counterparts on delayed recall, discrete and continuous control, and pixel-based tasks, while cutting trainable parameters by more than half.

Across the text-to-image Transformers FLUX.1, Z-Image and Infinity-2B, attention sinks land in architecture-specific places: the autoregressive model concentrates on its first global token, while the bidirectional diffusion models concentrate near modality boundaries, inside text regions and at image corners.

Across the text-to-image Transformers FLUX.1, Z-Image and Infinity-2B, attention sinks land in architecture-specific places: the autoregressive model concentrates on its first global token, while the bidirectional diffusion models concentrate near modality boundaries, inside text regions and at image corners.

Multi-agent collaboration is now sized by how competent the backbone is on the current query: 93.81% average across six benchmarks, and on higher-competence backbones average accuracy rises from 87.22 to 89.97 while token cost per query drops 36.26%.

Multi-agent collaboration is now sized by how competent the backbone is on the current query: 93.81% average across six benchmarks, and on higher-competence backbones average accuracy rises from 87.22 to 89.97 while token cost per query drops 36.26%.

From one RGB image and a task prompt, a video diffusion model generates virtual camera views that ORB-SLAM3 and aligned dense depth turn into a plannable 3D space; planning-feasible scenes rise from 13/19 at P20 to 15/19 at P81 across a 19-scene study.

From one RGB image and a task prompt, a video diffusion model generates virtual camera views that ORB-SLAM3 and aligned dense depth turn into a plannable 3D space; planning-feasible scenes rise from 13/19 at P20 to 15/19 at P81 across a 19-scene study.

A multi-scale biophysical model couples sleep homeostasis, circadian drive, a hypothalamic switch and a cortical spiking network to simulate five sleep disorders under one 72-hour photoperiod; paradoxical insomnia reaches the same total sleep time as controls yet scores lowest on restoration.

A multi-scale biophysical model couples sleep homeostasis, circadian drive, a hypothalamic switch and a cortical spiking network to simulate five sleep disorders under one 72-hour photoperiod; paradoxical insomnia reaches the same total sleep time as controls yet scores lowest on restoration.

Zero-order fine-tuning previously treated training examples alike; SAZO learns each example's importance online and allocates queries accordingly, improving average downstream performance across four model families and cutting the queries needed to reach the same training loss by up to 6x.

Zero-order fine-tuning previously treated training examples alike; SAZO learns each example's importance online and allocates queries accordingly, improving average downstream performance across four model families and cutting the queries needed to reach the same training loss by up to 6x.

SkillForge turns an agent's skill library from an append-only store into a fitness-driven lifecycle of trial, active, stable, and retired states, so skills and policy co-evolve during reinforcement learning and reach the highest aggregate success rate across ALFWorld, WebShop, and Search-Augmented QA, up to 7.8% relative over the strongest baseline.

SkillForge turns an agent's skill library from an append-only store into a fitness-driven lifecycle of trial, active, stable, and retired states, so skills and policy co-evolve during reinforcement learning and reach the highest aggregate success rate across ALFWorld, WebShop, and Search-Augmented QA, up to 7.8% relative over the strongest baseline.

Biomedical retrieval sentence encoders can be converted into typed decision models that return thresholdable probabilities in one forward pass, with a cross head reaching 66–73 seen-task points, close to fine-tuned Laya-large at 75.0 with about three times the parameters.

Biomedical retrieval sentence encoders can be converted into typed decision models that return thresholdable probabilities in one forward pass, with a cross head reaching 66–73 seen-task points, close to fine-tuned Laya-large at 75.0 with about three times the parameters.

Evo-World adds optical-flow-supervised future motion prediction to a lightweight vision-language-action policy, reaching 91.1% tier-averaged success on Meta-World MT50, a 13.0% relative gain over the non-predictive baseline, and beating the evaluated baselines on AgileX single-arm and bimanual tasks with 0.87B parameters.

Evo-World adds optical-flow-supervised future motion prediction to a lightweight vision-language-action policy, reaching 91.1% tier-averaged success on Meta-World MT50, a 13.0% relative gain over the non-predictive baseline, and beating the evaluated baselines on AgileX single-arm and bimanual tasks with 0.87B parameters.

Three-dimensional Gaussian reconstruction for building inspection no longer aims only at photorealistic views: with region and point-wise depth supervision, Gaussian centers are pulled toward physical surfaces, cutting depth error by 58.1% and 86.8% on synthetic UAV and smartphone-captured data, with a mean relative height-measurement error of 0.33%.

Three-dimensional Gaussian reconstruction for building inspection no longer aims only at photorealistic views: with region and point-wise depth supervision, Gaussian centers are pulled toward physical surfaces, cutting depth error by 58.1% and 86.8% on synthetic UAV and smartphone-captured data, with a mean relative height-measurement error of 0.33%.

Transformer attention is interpreted as max-product belief propagation on a Markov random field that varies from instance to instance, while the parameter change between layers acts as the parameter-update step of EM, a view that groups self-attention, mixture-of-experts, gated units and selective state-space models under one meta-model principle in which the input generates the parameters.

Transformer attention is interpreted as max-product belief propagation on a Markov random field that varies from instance to instance, while the parameter change between layers acts as the parameter-update step of EM, a view that groups self-attention, mixture-of-experts, gated units and selective state-space models under one meta-model principle in which the input generates the parameters.

A distributed feedback built on harmonic edge signals turns hole information into robot forces in planar swarms whose sensing disks cover a region; under stated assumptions, model and initial state alone yield offline rational lower bounds on cumulative hole area, and in a 15-second numerical demo both initial holes close by 9.9s.

A distributed feedback built on harmonic edge signals turns hole information into robot forces in planar swarms whose sensing disks cover a region; under stated assumptions, model and initial state alone yield offline rational lower bounds on cumulative hole area, and in a 15-second numerical demo both initial holes close by 9.9s.

After GeneNetwork's 1TB structured database was remodeled as an RDF graph alongside a literature corpus, a system that runs document, graph and agentic retrieval-augmented generation in parallel and merges their answers raised the satisfaction rate from 36% for plain document retrieval to 86% on 40 expert questions, while delegating most work to a local model still reached 59%.

After GeneNetwork's 1TB structured database was remodeled as an RDF graph alongside a literature corpus, a system that runs document, graph and agentic retrieval-augmented generation in parallel and merges their answers raised the satisfaction rate from 36% for plain document retrieval to 86% on 40 expert questions, while delegating most work to a local model still reached 59%.

Relay-side scheduling and dropping in deep-space DTNs are driven by destination-side information freshness, and in DTNSim simulations with a shared Moon–Mars relay buffer, masked Lagrangian PPO reaches the lowest mean aggregate excess AoI while meeting drop, expiry and inter-source fairness budgets.

Relay-side scheduling and dropping in deep-space DTNs are driven by destination-side information freshness, and in DTNSim simulations with a shared Moon–Mars relay buffer, masked Lagrangian PPO reaches the lowest mean aggregate excess AoI while meeting drop, expiry and inter-source fairness budgets.

Focused ion beam milling variance is no longer predicted as a single deterministic result: a conditional diffusion model learns the statistical distribution of process outcomes given the input design, yielding distributional predictions of edge morphology, linewidth variation, and defect probability.

Focused ion beam milling variance is no longer predicted as a single deterministic result: a conditional diffusion model learns the statistical distribution of process outcomes given the input design, yielding distributional predictions of edge morphology, linewidth variation, and defect probability.

Treating undiagnosed hypertension screening as a positive-unlabeled problem in 43,105 Chinese community adults reached a sensitivity of 0.817 and 729 false negatives in an independent test set, 41.1% fewer than standard XGBoost.

Treating undiagnosed hypertension screening as a positive-unlabeled problem in 43,105 Chinese community adults reached a sensitivity of 0.817 and 729 false negatives in an independent test set, 41.1% fewer than standard XGBoost.

Google Research released a workshop report on agentic privacy and security that argues for grounding agent behavior in Contextual Integrity, turning high-level social norms into machine-readable policy that judges whether an information flow is appropriate before an action runs.

Google Research released a workshop report on agentic privacy and security that argues for grounding agent behavior in Contextual Integrity, turning high-level social norms into machine-readable policy that judges whether an information flow is appropriate before an action runs.

RegulaTutor links historical evidence, learner interpretation and teaching policy into one inspectable chain for long-term personalized tutoring, raising historical-evidence accuracy by 13.2 percentage points and learner diagnosis by 11.6 points on average across nine fixed-backbone comparisons on LongTutor.

RegulaTutor links historical evidence, learner interpretation and teaching policy into one inspectable chain for long-term personalized tutoring, raising historical-evidence accuracy by 13.2 percentage points and learner diagnosis by 11.6 points on average across nine fixed-backbone comparisons on LongTutor.

NVIDIA and Microsoft said they are co-engineering hardware and software for AI agents on Windows PCs, with RTX Spark laptops on preorder now and shipping October 16, offering up to 128GB unified memory and a petaflop of FP4 compute so a 125B-class model runs locally, unmetered.

NVIDIA and Microsoft said they are co-engineering hardware and software for AI agents on Windows PCs, with RTX Spark laptops on preorder now and shipping October 16, offering up to 128GB unified memory and a petaflop of FP4 compute so a 125B-class model runs locally, unmetered.

Cloudflare's Agents SDK adds a web_fetch tool that lets an agent pull webpages, PDFs, and supported Office documents by URL and turn them into model-readable text through Workers AI Markdown Conversion.

Cloudflare's Agents SDK adds a web_fetch tool that lets an agent pull webpages, PDFs, and supported Office documents by URL and turn them into model-readable text through Workers AI Markdown Conversion.

Meta's Muse Wonder is now callable through AI Gateway for coding, long-running agents, and computer use, with a 1M-token context window and text, image, and PDF input, so developers reach it with one AI Gateway key instead of a separate Meta account.

Meta's Muse Wonder is now callable through AI Gateway for coding, long-running agents, and computer use, with a 1M-token context window and text, image, and PDF input, so developers reach it with one AI Gateway key instead of a separate Meta account.

Cohere released North 2, an enterprise agent platform that puts reusable skills, shared knowledge libraries, cross-session memory and per-agent guardrails into one privately deployable system, so regulated industries no longer have to choose between security, intelligence, cost and control.

Cohere released North 2, an enterprise agent platform that puts reusable skills, shared knowledge libraries, cross-session memory and per-agent guardrails into one privately deployable system, so regulated industries no longer have to choose between security, intelligence, cost and control.

Microsoft's Decision-1 is now callable through AI Gateway, answering typed questions about text with probabilities, choices, and scores using only a gateway API key.

Microsoft's Decision-1 is now callable through AI Gateway, answering typed questions about text with probabilities, choices, and scores using only a gateway API key.