AI and science frontiers · 2026-10-21

Deciding which change is worth trying when a language model improves a language model has depended on costly training experiments. LM2LM uses a frozen LLM as a research world model that predicts a proposal's gain from real experimental records in context, turning those predictions into reinforcement-learning rewards, and periodically executes the current policy's top-rated proposals to write measured outcomes back into the records. Across eight autoregressive and diffusion language-model development environments, Qwen3.8-27B trained this way outperforms GPT-6-sol on held-out tasks, even when GPT-6-sol is given 300 real experimental records in context, at about 33% lower data-generation cost.

Deciding which change is worth trying when a language model improves a language model has depended on costly training experiments. LM2LM uses a frozen LLM as a research world model that predicts a proposal's gain from real experimental records in context, turning those predictions into reinforcement-learning rewards, and periodically executes the current policy's top-rated proposals to write measured outcomes back into the records. Across eight autoregressive and diffusion language-model development environments, Qwen3.8-27B trained this way outperforms GPT-6-sol on held-out tasks, even when GPT-6-sol is given 300 real experimental records in context, at about 33% lower data-generation cost.

Language models do not acquire in-context learning in one step. A controlled diagnostic separates copying the nearest demonstration from composing mappings spread across the prompt; across 17 models about a third prefer the copied answer and the rest the rule-consistent one, and in OLMo-2-7B dependence on the requested value overtakes the conflicting example hundreds of billions of tokens before accuracy rises above chance. Training on synthetic definition-and-query documents flips the preference, while exposure to definitions or suppressing copy-favoring heads does not.

Language models do not acquire in-context learning in one step. A controlled diagnostic separates copying the nearest demonstration from composing mappings spread across the prompt; across 17 models about a third prefer the copied answer and the rest the rule-consistent one, and in OLMo-2-7B dependence on the requested value overtakes the conflicting example hundreds of billions of tokens before accuracy rises above chance. Training on synthetic definition-and-query documents flips the preference, while exposure to definitions or suppressing copy-favoring heads does not.

Rare-variant studies now yield more significant gene-phenotype associations than teams can place against prior evidence, and that placement has become the tighter bottleneck. PESTO combines large language model literature curation with structured Open Targets genetic evidence in one agentic pipeline that assigns Novel, Hypothesized, Existing or Established with supporting citations. Across 940 unique associations from two rare-variant studies, 7% were called Novel and 21% Hypothesized; 86% of 50 randomly paired negative controls fell into those two tiers.

Rare-variant studies now yield more significant gene-phenotype associations than teams can place against prior evidence, and that placement has become the tighter bottleneck. PESTO combines large language model literature curation with structured Open Targets genetic evidence in one agentic pipeline that assigns Novel, Hypothesized, Existing or Established with supporting citations. Across 940 unique associations from two rare-variant studies, 7% were called Novel and 21% Hypothesized; 86% of 50 randomly paired negative controls fell into those two tiers.

Changing the GPU count during training usually changes the order in which a data loader feeds samples to the model, so an ablation can end up measuring hardware allocation instead of a data change. Zephon splits each loading operator into a single-threaded accumulator that makes stateful decisions and stateless parallel workers, and divides data into a fixed number of logical lanes; plugged into TorchTitan in place of the default loader, it resumed a 16-GPU checkpoint on 8, 32 and 64 GPUs with every global batch matching the original run's SHA-256 hash.

Changing the GPU count during training usually changes the order in which a data loader feeds samples to the model, so an ablation can end up measuring hardware allocation instead of a data change. Zephon splits each loading operator into a single-threaded accumulator that makes stateful decisions and stateless parallel workers, and divides data into a fixed number of logical lanes; plugged into TorchTitan in place of the default loader, it resumed a 16-GPU checkpoint on 8, 32 and 64 GPUs with every global batch matching the original run's SHA-256 hash.

For multimodal-LLM phone agents, MobileInject places one malicious objective in both the screenshot and the interface text and runs 108 executable risk instances on a resettable Android emulator. Prior evaluations mostly watched a single text channel or judged only whether an attack succeeded; with exposure counts matched, one appearance per channel reached 85.4% attack success and 0% benign completion, while repeating within one channel reached only 72.9% or 27.8%. In 30 of 108 paired instances, neither single channel succeeded but the aligned pair did.

For multimodal-LLM phone agents, MobileInject places one malicious objective in both the screenshot and the interface text and runs 108 executable risk instances on a resettable Android emulator. Prior evaluations mostly watched a single text channel or judged only whether an attack succeeded; with exposure counts matched, one appearance per channel reached 85.4% attack success and 0% benign completion, while repeating within one channel reached only 72.9% or 27.8%. In 30 of 108 paired instances, neither single channel succeeded but the aligned pair did.

Ai2 released FairStress, a benchmark that moves demographic fairness evaluation for large language models from which candidate a model ultimately picks to whether it holds that pick under challenge. Its 300 two-candidate scenarios span 29 domains including hiring, healthcare and lending, paired with matched persuasive prefixes to form more than 13 million test cases. Across 15 models, the average gap in abandoning correct answers was 9.3 percentage points with explicit identity labels and 6.9 points with implied identity.

Ai2 released FairStress, a benchmark that moves demographic fairness evaluation for large language models from which candidate a model ultimately picks to whether it holds that pick under challenge. Its 300 two-candidate scenarios span 29 domains including hiring, healthcare and lending, paired with matched persuasive prefixes to form more than 13 million test cases. Across 15 models, the average gap in abandoning correct answers was 9.3 percentage points with explicit identity labels and 6.9 points with implied identity.

Ai2 released FairStress, a benchmark that measures demographic fairness in large language models beyond which candidate a model finally picks, by testing whether it abandons that choice under persuasion. Its 300 two-candidate scenarios span 29 domains and expand to more than 13 million test cases when identity presentation and matched adversarial prefixes are combined; across 15 models, the average gap in abandoning correct answers was 6.9 percentage points with implied identity and 9.3 points with explicit labels.

Ai2 released FairStress, a benchmark that measures demographic fairness in large language models beyond which candidate a model finally picks, by testing whether it abandons that choice under persuasion. Its 300 two-candidate scenarios span 29 domains and expand to more than 13 million test cases when identity presentation and matched adversarial prefixes are combined; across 15 models, the average gap in abandoning correct answers was 6.9 percentage points with implied identity and 9.3 points with explicit labels.

The SGLang team, with Google and Meta, released SGLang-TorchTPU, an open-source backend that runs the same SGLang serving engine on both GPUs and TPUs. TPU had not been a native PyTorch device, so a PyTorch-centric serving stack could not be reused directly; the new backend keeps the scheduler, RadixAttention and API server unchanged and places TPU work underneath. It is already running in production at Meta and serves as the rollout engine for reinforcement learning with TorchTitan through Miles.

The SGLang team, with Google and Meta, released SGLang-TorchTPU, an open-source backend that runs the same SGLang serving engine on both GPUs and TPUs. TPU had not been a native PyTorch device, so a PyTorch-centric serving stack could not be reused directly; the new backend keeps the scheduler, RadixAttention and API server unchanged and places TPU work underneath. It is already running in production at Meta and serves as the rollout engine for reinforcement learning with TorchTitan through Miles.

Particle transport in complex flows is usually obtained by solving the Navier–Stokes equations, and numerical or experimental routes demand deep physical insight and high computational cost, while existing machine-learning alternatives fail to generalize long-range particle trajectories under sparse data. TrajectoryFlowNet couples a Lagrangian view of particle tracks with a Eulerian view of full velocity and pressure fields inside one physics-constrained network, so a few trajectory measurements yield both the full-field flow and long-range paths for arbitrary particles. It is validated on numerical lid-driven cavity and complex cylinder flows and on experimental aortic and ventricle blood flows, keeping predictions physically consistent in flows with irregular boundaries.

Particle transport in complex flows is usually obtained by solving the Navier–Stokes equations, and numerical or experimental routes demand deep physical insight and high computational cost, while existing machine-learning alternatives fail to generalize long-range particle trajectories under sparse data. TrajectoryFlowNet couples a Lagrangian view of particle tracks with a Eulerian view of full velocity and pressure fields inside one physics-constrained network, so a few trajectory measurements yield both the full-field flow and long-range paths for arbitrary particles. It is validated on numerical lid-driven cavity and complex cylinder flows and on experimental aortic and ventricle blood flows, keeping predictions physically consistent in flows with irregular boundaries.

In multiple-choice evaluation, whether the space after the prompt's colon belongs to the prompt or to the option changes the probability the model assigns to the correct answer. Keeping the space with the option matches the tokenization seen in pre-training; keeping it with the prompt lowers the correct answer's log-probability by several orders of magnitude. Accuracy changes by less than half a point in six of nine model-benchmark pairs, but Qwen3.5 loses almost ten points on PIQA.

In multiple-choice evaluation, whether the space after the prompt's colon belongs to the prompt or to the option changes the probability the model assigns to the correct answer. Keeping the space with the option matches the tokenization seen in pre-training; keeping it with the prompt lowers the correct answer's log-probability by several orders of magnitude. Accuracy changes by less than half a point in six of nine model-benchmark pairs, but Qwen3.5 loses almost ten points on PIQA.

DecisionRobust evaluates JEV-style decision models under text and image perturbations and prompt injection, and finds that 30 text models and 17 vision models take the same candidate set, with held-out models excluded from candidate selection also incurring errors while label-consistency checks sharply reduce apparent attack success.

DecisionRobust evaluates JEV-style decision models under text and image perturbations and prompt injection, and finds that 30 text models and 17 vision models take the same candidate set, with held-out models excluded from candidate selection also incurring errors while label-consistency checks sharply reduce apparent attack success.

Retrieval quality in dynamic product advertising could previously be compared only through sparse, biased engagement logs or costly online experiments. The new method compresses a user's engagement history into a small set of intent representatives grouped by semantic ID and stratified by event type, then has a large language model judge each candidate product against every intent and aggregate the results into a user-level score. On 30 days of production traffic, the score ranked six retrieval sources consistently with conversion precision and recall, and moved in the same direction as online CTR and CVR across four A/B tests.

Retrieval quality in dynamic product advertising could previously be compared only through sparse, biased engagement logs or costly online experiments. The new method compresses a user's engagement history into a small set of intent representatives grouped by semantic ID and stratified by event type, then has a large language model judge each candidate product against every intent and aggregate the results into a user-level score. On 30 days of production traffic, the score ranked six retrieval sources consistently with conversion precision and recall, and moved in the same direction as online CTR and CVR across four A/B tests.

Multi-turn agent distillation now grounds teacher supervision in the interaction state, keeping only executed actions and observations, and the Qwen3-4B student surpasses its 32B teacher on ALFWorld Seen at 43.39%.

Multi-turn agent distillation now grounds teacher supervision in the interaction state, keeping only executed actions and observations, and the Qwen3-4B student surpasses its 32B teacher on ALFWorld Seen at 43.39%.

IBM has open-sourced a dataflow scheduler for its Spyre accelerator that reads KTIR, an MLIR representation, and turns Inductor's already-settled core division, scratchpad residency and access tiles into a dataflow pipeline, which device-specific stages then lower to SpyreCode. Previously no interface carried the full data-parallel view, so the backend had to rediscover the mapping. Bringing up the first op was about 4x faster than on the previous stack and later ops about 8x; coverage is elementwise operations with broadcast and reduction dimensions.

IBM has open-sourced a dataflow scheduler for its Spyre accelerator that reads KTIR, an MLIR representation, and turns Inductor's already-settled core division, scratchpad residency and access tiles into a dataflow pipeline, which device-specific stages then lower to SpyreCode. Previously no interface carried the full data-parallel view, so the backend had to rediscover the mapping. Bringing up the first op was about 4x faster than on the previous stack and later ops about 8x; coverage is elementwise operations with broadcast and reduction dimensions.

Galaxy rotation curves have long favored cored dark-matter halos over the steeply rising cusps predicted by cold dark matter simulations. Fitting five SPARC galaxies with Bayesian inference while modeling correlated errors between neighboring radii largely closes that gap. In DDO 154 the reduced chi-square moves from 2.40 versus 9.41 to 0.94 versus 1.04, leaving one-dimensional rotation curves unable to separate the two halo shapes.

Galaxy rotation curves have long favored cored dark-matter halos over the steeply rising cusps predicted by cold dark matter simulations. Fitting five SPARC galaxies with Bayesian inference while modeling correlated errors between neighboring radii largely closes that gap. In DDO 154 the reduced chi-square moves from 2.40 versus 9.41 to 0.94 versus 1.04, leaving one-dimensional rotation curves unable to separate the two halo shapes.

In a specialist allergy and clinical immunology cohort, representing each patient as a node in a knowledge graph and clustering graph embeddings avoided collapse under random, conditional and value-dependent missingness. The common impute-then-reduce pipeline performed worst under value-dependent missingness, because imputation conflates an untested antibody with a confirmed negative.

In a specialist allergy and clinical immunology cohort, representing each patient as a node in a knowledge graph and clustering graph embeddings avoided collapse under random, conditional and value-dependent missingness. The common impute-then-reduce pipeline performed worst under value-dependent missingness, because imputation conflates an untested antibody with a confirmed negative.

Image factuality assessment no longer stops at whether a picture is AI-generated: the EviScene multimodal agent refines its own questions during investigation and reaches 79.05% balanced accuracy and 52.26% evidence sufficiency on 1,527 test images.

Image factuality assessment no longer stops at whether a picture is AI-generated: the EviScene multimodal agent refines its own questions during investigation and reaches 79.05% balanced accuracy and 52.26% evidence sufficiency on 1,527 test images.

Training-free generative compression no longer needs an explicit Gaussian codebook: H-DDCM swaps it for an implicit randomized Hadamard dictionary, matching complete-codebook fidelity at FLUX scale while cutting round-trip time from tens of seconds to seconds.

Training-free generative compression no longer needs an explicit Gaussian codebook: H-DDCM swaps it for an implicit randomized Hadamard dictionary, matching complete-codebook fidelity at FLUX scale while cutting round-trip time from tens of seconds to seconds.

Reinforcement learning rewards for software engineering agents no longer stop at pass or fail: the current policy compares debugging traces from the same task and verifier outcome, turning those preferences into bounded auxiliary rewards and lifting Qwen3-14B average accuracy on SWE-bench Verified from 32.1% to 34.7%.

Reinforcement learning rewards for software engineering agents no longer stop at pass or fail: the current policy compares debugging traces from the same task and verifier outcome, turning those preferences into bounded auxiliary rewards and lifting Qwen3-14B average accuracy on SWE-bench Verified from 32.1% to 34.7%.

Mistral released Voxtral 3, which puts speech recognition, reasoning, tool calling and speech synthesis into a single real-time streaming pipeline, alongside a voice agent SDK. Previously, making an agent listen accurately, reason, speak naturally and act across systems in multiple steps remained difficult. Its Small transcription model beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription and diarization in English, French and Spanish, and realtime transcription runs below 500 ms latency.

Mistral released Voxtral 3, which puts speech recognition, reasoning, tool calling and speech synthesis into a single real-time streaming pipeline, alongside a voice agent SDK. Previously, making an agent listen accurately, reason, speak naturally and act across systems in multiple steps remained difficult. Its Small transcription model beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription and diarization in English, French and Spanish, and realtime transcription runs below 500 ms latency.

Chess engines play far beyond human ability yet cannot say why they prefer a move. A study first builds an engine-grounded score for explanations and finds that eight language models prompted directly lean on surface board features; it then introduces BinProp, a training-free method in which models iteratively propose and filter binary chess motifs on winning and losing positions, then use them to explain specific moves. On chess, the eight-model average rises to nearly three times the direct-prompting baseline, and the same setup extends to Go, draughts and shogi.

Chess engines play far beyond human ability yet cannot say why they prefer a move. A study first builds an engine-grounded score for explanations and finds that eight language models prompted directly lean on surface board features; it then introduces BinProp, a training-free method in which models iteratively propose and filter binary chess motifs on winning and losing positions, then use them to explain specific moves. On chess, the eight-model average rises to nearly three times the direct-prompting baseline, and the same setup extends to Go, draughts and shogi.

Professional motion video editing now has an agent retrieve and self-check a demonstration clip, then hand it with a grounded prompt to a frozen editor; across 314 professional motion tasks, Seedance 2.0 rises from 55.1 to 65.6 and MiniMax-H3 from 51.2 to 64.5.

Professional motion video editing now has an agent retrieve and self-check a demonstration clip, then hand it with a grounded prompt to a frozen editor; across 314 professional motion tasks, Seedance 2.0 rises from 55.1 to 65.6 and MiniMax-H3 from 51.2 to 64.5.

Long-Horizon-Terminal-Bench 2.0 extends agent evaluation from 90 minutes to three hours and, on 30 reconstruction tasks, reports binary verifier scores averaging 0.221 against dense ladder scores averaging 0.653.

Long-Horizon-Terminal-Bench 2.0 extends agent evaluation from 90 minutes to three hours and, on 30 reconstruction tasks, reports binary verifier scores averaging 0.221 against dense ladder scores averaging 0.653.

A training-free diffusion caching rule spends its exact denoiser evaluations where they matter by dividing shallow-probe output change by input displacement, lowering median LPIPS at matched exact-compute budgets on FLUX.1-dev, Wan2.1 and Open-Sora 1.2.

A training-free diffusion caching rule spends its exact denoiser evaluations where they matter by dividing shallow-probe output change by input displacement, lowering median LPIPS at matched exact-compute budgets on FLUX.1-dev, Wan2.1 and Open-Sora 1.2.

Multimodal large language models associate visual cues with category concepts too weakly during fine-grained recognition; LIVE tightens the candidate category list and adds category contrast hints at inference time, raising Qwen2.5-VL-7B average accuracy by 14.96 points across 13 benchmarks.

Multimodal large language models associate visual cues with category concepts too weakly during fine-grained recognition; LIVE tightens the candidate category list and adds category contrast hints at inference time, raising Qwen2.5-VL-7B average accuracy by 14.96 points across 13 benchmarks.

An evaluation across 220 data science tasks finds most agentic solutions contain at least one methodological failure, and that data leakage failures correlate positively with predictive performance, so accuracy alone misses method problems.

An evaluation across 220 data science tasks finds most agentic solutions contain at least one methodological failure, and that data leakage failures correlate positively with predictive performance, so accuracy alone misses method problems.

Generated meshes often carry holes, stray faces and duplicated faces; MeshRepair reads the draft as a prefix and autoregressively emits only a deletion-then-addition edit sequence, raising quad-equivalent face F1 on LATO2 drafts from 0.294 to 0.340 where rule-based repair tools leave it flat or lower.

Generated meshes often carry holes, stray faces and duplicated faces; MeshRepair reads the draft as a prefix and autoregressively emits only a deletion-then-addition edit sequence, raising quad-equivalent face F1 on LATO2 drafts from 0.294 to 0.340 where rule-based repair tools leave it flat or lower.

Model routing for multi-turn agents is studied at the user-turn boundary, where a router picks one model for the whole turn before it runs. At a matched offloading rate of roughly a fifth of turns, a router that reads session history passes 28.3% of SheetCopilot tasks against 25.0% for the same router seeing only the current request, a 3.3-point gap, while staying within one point of always using the expensive model. A closed-form cache analysis finds downward switches repay their cache cost in about two calls, while upward switches do not break even under the evaluated conditions.

Model routing for multi-turn agents is studied at the user-turn boundary, where a router picks one model for the whole turn before it runs. At a matched offloading rate of roughly a fifth of turns, a router that reads session history passes 28.3% of SheetCopilot tasks against 25.0% for the same router seeing only the current request, a 3.3-point gap, while staying within one point of always using the expensive model. A closed-form cache analysis finds downward switches repay their cache cost in about two calls, while upward switches do not break even under the evaluated conditions.

On structural MRI from three independent cohorts, Alzheimer's disease classifiers use age as a shortcut: older patients receive higher predicted risk. Residualizing age linearly in feature space (PMDN) brings the age-prediction correlation near zero in cognitively normal and mild cognitive impairment groups, with balanced accuracy 0.71 and Brier score about 0.22, and far less fold-to-fold spread than the adversarial CF-Net. Post hoc calibration further lowers probability error, with nonparametric isotonic regression and Bayesian binning performing best.

On structural MRI from three independent cohorts, Alzheimer's disease classifiers use age as a shortcut: older patients receive higher predicted risk. Residualizing age linearly in feature space (PMDN) brings the age-prediction correlation near zero in cognitively normal and mild cognitive impairment groups, with balanced accuracy 0.71 and Brier score about 0.22, and far less fold-to-fold spread than the adversarial CF-Net. Post hoc calibration further lowers probability error, with nonparametric isotonic regression and Bayesian binning performing best.

In imperfect-information games with ordered state spaces, cumulative regrets are no longer stored cell by cell for every information set; they are transformed into Haar wavelet coefficients and only the largest are kept. Tabular CFR+ previously grew memory linearly with the game tree under fine discretization, while Deep CFR traded memory for neural inference latency. On Deep-Stack Kuhn poker, growing the state count from one thousand to ten million raised memory only from 0.6 MB to 2.1 MB, a compression ratio above 99.9%; on Flop Hold'em, reaching comparable exploitability took about one-sixth the wall-clock time of Deep CFR.

In imperfect-information games with ordered state spaces, cumulative regrets are no longer stored cell by cell for every information set; they are transformed into Haar wavelet coefficients and only the largest are kept. Tabular CFR+ previously grew memory linearly with the game tree under fine discretization, while Deep CFR traded memory for neural inference latency. On Deep-Stack Kuhn poker, growing the state count from one thousand to ten million raised memory only from 0.6 MB to 2.1 MB, a compression ratio above 99.9%; on Flop Hold'em, reaching comparable exploitability took about one-sixth the wall-clock time of Deep CFR.

Full Capacity Protocols in emergency departments usually start only after crowding thresholds are crossed, yet hallway beds, temporary nurses and transfer restrictions take time to arrange. A discrete-event simulation evaluation embedded a machine learning model forecasting waiting-room counts six hours ahead into the emergency flow and used those forecasts to trigger the protocol early. Over a simulated year, early activation significantly shortened time above the crowding threshold and average protocol duration at all three threshold settings, while changes in average emergency length of stay and wait time were not statistically significant.

Full Capacity Protocols in emergency departments usually start only after crowding thresholds are crossed, yet hallway beds, temporary nurses and transfer restrictions take time to arrange. A discrete-event simulation evaluation embedded a machine learning model forecasting waiting-room counts six hours ahead into the emergency flow and used those forecasts to trigger the protocol early. Over a simulated year, early activation significantly shortened time above the crowding threshold and average protocol duration at all three threshold settings, while changes in average emergency length of stay and wait time were not statistically significant.

ElevenLabs released Eleven Realtime, a family of speech-to-speech models that each act as a complete voice agent, replacing separate choices for speech recognition, voice activity detection, turn-taking, the underlying LLM and text to speech. Both support more than 90 languages and the library's 17,000+ voices. In independent evaluation, v1 Max answered 98.8% of 1,000 spoken reasoning questions correctly and began speaking in 0.57 seconds on average; v1 Mini scored 75.9% and 0.62 seconds.

ElevenLabs released Eleven Realtime, a family of speech-to-speech models that each act as a complete voice agent, replacing separate choices for speech recognition, voice activity detection, turn-taking, the underlying LLM and text to speech. Both support more than 90 languages and the library's 17,000+ voices. In independent evaluation, v1 Max answered 98.8% of 1,000 spoken reasoning questions correctly and began speaking in 0.57 seconds on average; v1 Mini scored 75.9% and 0.62 seconds.

Underwater images suffer color shifts, low contrast and blurred edges from light attenuation, particle scattering and uneven illumination, while public datasets cover only a limited range of degradation types and intensities. GDH-Det extracts water color, turbidity and blur level from real images as degradation conditions, uses a conditional diffusion network with bounding-box masks to generate degraded images at continuously varying intensities, and screens candidates by object-region structure and detection credibility. During detector training, sample difficulty is computed from classification uncertainty, localization deviation and perturbation response to set sampling probabilities, reaching mAP@0.5 of 85.3%, 79.6%, 66.2% and 84.7% on URPC2020, DUO, UODD and Brackish.

Underwater images suffer color shifts, low contrast and blurred edges from light attenuation, particle scattering and uneven illumination, while public datasets cover only a limited range of degradation types and intensities. GDH-Det extracts water color, turbidity and blur level from real images as degradation conditions, uses a conditional diffusion network with bounding-box masks to generate degraded images at continuously varying intensities, and screens candidates by object-region structure and detection credibility. During detector training, sample difficulty is computed from classification uncertainty, localization deviation and perturbation response to set sampling probabilities, reaching mAP@0.5 of 85.3%, 79.6%, 66.2% and 84.7% on URPC2020, DUO, UODD and Brackish.

Automated angle-of-progression (AoP) measurement from intrapartum ultrasound is accurate to a few degrees on average, yet a small fraction of frames fails silently by tens of degrees. On the 501 test files of the public IUGC2025 benchmark, an ensemble of six landmark detectors ranked frames by disagreement between detectors; referring the most suspicious 5% for manual review lowered the share of errors above 20 degrees among retained frames from 1.2% to 0.2%, catching five of six. The missed case was one all six detectors underestimated with little spread between them; no difference was detected between angle-space and coordinate-space disagreement at the primary endpoint, and thresholds set on validation frames did not hold their intended referral rate on test frames.

Automated angle-of-progression (AoP) measurement from intrapartum ultrasound is accurate to a few degrees on average, yet a small fraction of frames fails silently by tens of degrees. On the 501 test files of the public IUGC2025 benchmark, an ensemble of six landmark detectors ranked frames by disagreement between detectors; referring the most suspicious 5% for manual review lowered the share of errors above 20 degrees among retained frames from 1.2% to 0.2%, catching five of six. The missed case was one all six detectors underestimated with little spread between them; no difference was detected between angle-space and coordinate-space disagreement at the primary endpoint, and thresholds set on validation frames did not hold their intended referral rate on test frames.

Claude Managed Agents on the Claude Platform can now write their own workflows in beta for work with many pieces, such as reviewing hundreds of documents. A workflow is a program that runs many agents in phases and combines their results, and the server runs it in the background as a workflow run. It is enabled by setting the agent's multiagent field to multiagent_20261001 with workflows enabled, and each run is followed through workflow_run.* events on the session's event stream.

Claude Managed Agents on the Claude Platform can now write their own workflows in beta for work with many pieces, such as reviewing hundreds of documents. A workflow is a program that runs many agents in phases and combines their results, and the server runs it in the background as a workflow run. It is enabled by setting the agent's multiagent field to multiagent_20261001 with workflows enabled, and each run is followed through workflow_run.* events on the session's event stream.

Vercel released durable-harness, an open-source set of adapters that lets coding agents such as Pi, OpenCode, Codex and Claude Code resume a turn from a log after a cloud function crashes or times out. Running such agents in the cloud previously meant a VM with persistent storage, which does not record which model and tool calls finished or resume interrupted turns, so one failure lost work. The tool records an intent before each model or tool call and an outcome after it, and a queue message hands a dead turn to a new function that reads the log and continues.

Vercel released durable-harness, an open-source set of adapters that lets coding agents such as Pi, OpenCode, Codex and Claude Code resume a turn from a log after a cloud function crashes or times out. Running such agents in the cloud previously meant a VM with persistent storage, which does not record which model and tool calls finished or resume interrupted turns, so one failure lost work. The tool records an intent before each model or tool call and an outcome after it, and a queue message hands a dead turn to a new function that reads the log and continues.

Google Cloud has configured its general-purpose Gemini work agent into a version for government agencies, offered in preview. Government work runs on statute, regulation and agency guidance spread across siloed legacy systems under data-sovereignty and confidentiality rules. The version adds government-task skills written by domain experts, secure connectors to existing case management, ERP and HR systems, and pre-built partner agents, all under one governance control plane, with outputs grounded in source documents and citations.

Google Cloud has configured its general-purpose Gemini work agent into a version for government agencies, offered in preview. Government work runs on statute, regulation and agency guidance spread across siloed legacy systems under data-sovereignty and confidentiality rules. The version adds government-task skills written by domain experts, secure connectors to existing case management, ERP and HR systems, and pre-built partner agents, all under one governance control plane, with outputs grounded in source documents and citations.

Whether diffusion language models can compete with autoregressive models on complex tasks had been an open question. Inception's Mercury 3 scales parameters more than 40x over Mercury 1 and runs at 1,068 tokens per second on NVIDIA GPUs; the company reports it matches or beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while running 4-8x faster. In 50 live Mind2Web navigation tasks it completed 50%, well above GPT-6.1-Sol, using the same token volume and agent steps.

Whether diffusion language models can compete with autoregressive models on complex tasks had been an open question. Inception's Mercury 3 scales parameters more than 40x over Mercury 1 and runs at 1,068 tokens per second on NVIDIA GPUs; the company reports it matches or beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while running 4-8x faster. In 50 live Mind2Web navigation tasks it completed 50%, well above GPT-6.1-Sol, using the same token volume and agent steps.

Automated screening for voice disorders is limited by scarce, imbalanced clinical speech data. Using standardized sentence recordings from the Saarbrücken Voice Database, two text-to-speech models each generated one synthetic sample per training recording, and several architectures and fusion strategies were compared. In blinded listening only 52.6% of Kartoffelbox samples were identified as synthetic versus 89.5% for XTTS; with Kartoffelbox augmentation HuBERT reached 0.88 sensitivity and 0.87 accuracy, while laryngitis remained hardest in the four-class task.

Automated screening for voice disorders is limited by scarce, imbalanced clinical speech data. Using standardized sentence recordings from the Saarbrücken Voice Database, two text-to-speech models each generated one synthetic sample per training recording, and several architectures and fusion strategies were compared. In blinded listening only 52.6% of Kartoffelbox samples were identified as synthetic versus 89.5% for XTTS; with Kartoffelbox augmentation HuBERT reached 0.88 sensitivity and 0.87 accuracy, while laryngitis remained hardest in the four-class task.

Equivalent-circuit fitting of electrochemical impedance spectra has depended on a human choosing a circuit topology and starting values, even though different topologies can produce nearly identical spectra. A University of Zagreb team expanded a few manually fitted spectra into a synthetic reference library, projected it into principal-component space, and used nearest-neighbour retrieval to shortlist topologies and supply physically plausible starting values at once, then refined each candidate by complex nonlinear least squares and chose by fit quality. Against independent manual fits, most parameters agreed within 1%, reduced chi-squared was lower in every case, and one spectrum took about 5 to 6 seconds.

Equivalent-circuit fitting of electrochemical impedance spectra has depended on a human choosing a circuit topology and starting values, even though different topologies can produce nearly identical spectra. A University of Zagreb team expanded a few manually fitted spectra into a synthetic reference library, projected it into principal-component space, and used nearest-neighbour retrieval to shortlist topologies and supply physically plausible starting values at once, then refined each candidate by complex nonlinear least squares and chose by fit quality. Against independent manual fits, most parameters agreed within 1%, reduced chi-squared was lower in every case, and one spectrum took about 5 to 6 seconds.

Rafiq is an Arabic conversational system for early depression detection that combines per-message symptom monitoring, a three-phase PHQ-9 questionnaire delivered inside the chat, and retrieval-augmented answers grounded in a clinical knowledge base. On 415 held-out instances symptom detection reached 0.98 accuracy, while retrieval quality depended heavily on embedder choice: an Arabic-specific model scored 80% MAP and a general multilingual mixture-of-experts model only 4%.

Rafiq is an Arabic conversational system for early depression detection that combines per-message symptom monitoring, a three-phase PHQ-9 questionnaire delivered inside the chat, and retrieval-augmented answers grounded in a clinical knowledge base. On 415 held-out instances symptom detection reached 0.98 accuracy, while retrieval quality depended heavily on embedder choice: an Arabic-specific model scored 80% MAP and a general multilingual mixture-of-experts model only 4%.

Distribution-network self-healing has traded decision quality, real-time speed and explainability against one another: mathematical programming nears the optimum but is slow, while deep reinforcement learning decides in milliseconds but stays opaque. Embedding a graph attention network as the shared backbone of a soft actor–critic agent and writing a resilience index into the reward brings online decisions to 2.3 ms on the IEEE 33-bus and 123-bus systems, within 2.8% of the mixed-integer linear programming optimum. Attention heatmaps and SHAP attributions show where the agent looks and why it decides.

Distribution-network self-healing has traded decision quality, real-time speed and explainability against one another: mathematical programming nears the optimum but is slow, while deep reinforcement learning decides in milliseconds but stays opaque. Embedding a graph attention network as the shared backbone of a soft actor–critic agent and writing a resilience index into the reward brings online decisions to 2.3 ms on the IEEE 33-bus and 123-bus systems, within 2.8% of the mixed-integer linear programming optimum. Attention heatmaps and SHAP attributions show where the agent looks and why it decides.

DeepSeek V4.1 Flash 1017 is now available on Cloudflare Workers AI. The multimodal Mixture-of-Experts model has a 552B-parameter backbone but activates only 8B parameters per token during prefill and 16B during decode, which keeps input-heavy agentic workloads fast and low-cost. The 1017 release improves reasoning, coding, agentic tasks and visual benchmarks over the prior Flash version and completes tasks with shorter outputs, with image-plus-text requests, reasoning on by default, multi-turn function calling and a 1,048,576-token context window.

DeepSeek V4.1 Flash 1017 is now available on Cloudflare Workers AI. The multimodal Mixture-of-Experts model has a 552B-parameter backbone but activates only 8B parameters per token during prefill and 16B during decode, which keeps input-heavy agentic workloads fast and low-cost. The 1017 release improves reasoning, coding, agentic tasks and visual benchmarks over the prior Flash version and completes tasks with shorter outputs, with image-plus-text requests, reasoning on by default, multi-turn function calling and a 1,048,576-token context window.

Google Antigravity's model picker now includes Claude Opus 5.5 and Claude Sonnet 5.5 for Google Cloud customers, selectable by developers once an administrator enables them. Previously the models available inside Antigravity came from Google's own Gemini line, including Gemini 3.8 Flash and Gemini 4 Argon; Opus 5.5 targets complex work while Sonnet 5.5 is a faster, lower-cost complement. Both apply across Antigravity 2.0, the Antigravity CLI and the IDE extensions for VS Code, JetBrains IntelliJ and Visual Studio, and the active model can be switched between turns in one conversation thread while preserving workspace state, open artifacts and terminal context.

Google Antigravity's model picker now includes Claude Opus 5.5 and Claude Sonnet 5.5 for Google Cloud customers, selectable by developers once an administrator enables them. Previously the models available inside Antigravity came from Google's own Gemini line, including Gemini 3.8 Flash and Gemini 4 Argon; Opus 5.5 targets complex work while Sonnet 5.5 is a faster, lower-cost complement. Both apply across Antigravity 2.0, the Antigravity CLI and the IDE extensions for VS Code, JetBrains IntelliJ and Visual Studio, and the active model can be switched between turns in one conversation thread while preserving workspace state, open artifacts and terminal context.

Optimizing biomass torrefaction experimentally is limited by cost and time, since one-factor-at-a-time trials miss synergistic effects and full factorial designs quickly become prohibitive. Using 486 torrefaction observations, Gaussian process regression with automatic relevance determination kernels reached cross-validated R² of 0.854 for mass yield, 0.971 for higher heating value and 0.869 for energy yield, outperforming random forest, support vector regression and a neural network. An acquisition function weighting uncertainty, mass-yield improvement and HHV improvement at 0.4/0.3/0.3, plus a diversity penalty, then recommended 10 experiments spanning 200–315 °C and 30–1089 s, all inside the conventional torrefaction window.

Optimizing biomass torrefaction experimentally is limited by cost and time, since one-factor-at-a-time trials miss synergistic effects and full factorial designs quickly become prohibitive. Using 486 torrefaction observations, Gaussian process regression with automatic relevance determination kernels reached cross-validated R² of 0.854 for mass yield, 0.971 for higher heating value and 0.869 for energy yield, outperforming random forest, support vector regression and a neural network. An acquisition function weighting uncertainty, mass-yield improvement and HHV improvement at 0.4/0.3/0.3, plus a diversity penalty, then recommended 10 experiments spanning 200–315 °C and 30–1089 s, all inside the conventional torrefaction window.

Anthropic has open-sourced 11 knowledge-work plugins that let Claude act as a specialist for roles such as sales, finance and data by bundling skills, connectors, slash commands and sub-agents. Previously, fitting Claude to a role meant users organizing their own instructions, tool access and workflows. Each plugin is markdown and JSON with no code or build steps, and Anthropic presents them as generic starting points that become more useful once customized to a company's tools, terminology and processes.

Anthropic has open-sourced 11 knowledge-work plugins that let Claude act as a specialist for roles such as sales, finance and data by bundling skills, connectors, slash commands and sub-agents. Previously, fitting Claude to a role meant users organizing their own instructions, tool access and workflows. Each plugin is markdown and JSON with no code or build steps, and Anthropic presents them as generic starting points that become more useful once customized to a company's tools, terminology and processes.

Many injuries to young children come from relations between objects, such as a chair beside a sofa under a window. AuroraNet turns a room into an age-scaled climbing graph, sums severity-weighted expected harm along sequences a child can climb, and learns 15 edge parameters from scene-level labels. On simulated rooms it reaches a 0.986 rank correlation with gradient boosting, 0.878 versus 0.798 within strata; all evidence is simulation-based.

Many injuries to young children come from relations between objects, such as a chair beside a sofa under a window. AuroraNet turns a room into an age-scaled climbing graph, sums severity-weighted expected harm along sequences a child can climb, and learns 15 edge parameters from scene-level labels. On simulated rooms it reaches a 0.986 rank correlation with gradient boosting, 0.878 versus 0.798 within strata; all evidence is simulation-based.

AnyLanguageModel 1.0 turns Apple's Foundation Models interface into a replaceable layer: an app changes its import from FoundationModels to AnyLanguageModel, and the same session code can call the system model, open-weight models run through llama.cpp, MLX and Core ML, and several hosted providers. Before this, developers curious about local models were unsure how to start and wanted to avoid the added complexity of supporting multiple inference providers. Over the year the library aligned tool calling and guided generation across providers and added image input and Linux support.

AnyLanguageModel 1.0 turns Apple's Foundation Models interface into a replaceable layer: an app changes its import from FoundationModels to AnyLanguageModel, and the same session code can call the system model, open-weight models run through llama.cpp, MLX and Core ML, and several hosted providers. Before this, developers curious about local models were unsure how to start and wanted to avoid the added complexity of supporting multiple inference providers. Over the year the library aligned tool calling and guided generation across providers and added image input and Linux support.

In dermoscopic images, hair occlusion, uneven illumination and weak local contrast fragment lesion boundaries, while self-attention tends to emphasize lower-frequency content and leaves those weak borders underrepresented in skip connections. Swin-WaveNet decomposes each encoder skip feature with a wavelet transform, applies selective state-space scanning only to the high-frequency subbands to reconnect interrupted contours, gates shallow high-frequency responses with deeper semantic guidance to suppress hair and artifact structures, and sharpens boundaries during upsampling with high-frequency attention. On ISIC 2018 it reached a mean Dice of 89.69% and HD95 of 23.60 pixels over three runs, close to Swin-Unet, and 91.69% Dice on ISIC 2017 without fine-tuning.

In dermoscopic images, hair occlusion, uneven illumination and weak local contrast fragment lesion boundaries, while self-attention tends to emphasize lower-frequency content and leaves those weak borders underrepresented in skip connections. Swin-WaveNet decomposes each encoder skip feature with a wavelet transform, applies selective state-space scanning only to the high-frequency subbands to reconnect interrupted contours, gates shallow high-frequency responses with deeper semantic guidance to suppress hair and artifact structures, and sharpens boundaries during upsampling with high-frequency attention. On ISIC 2018 it reached a mean Dice of 89.69% and HD95 of 23.60 pixels over three runs, close to Swin-Unet, and 91.69% Dice on ISIC 2017 without fine-tuning.

Google has added a business landing page inside Gemini that gathers branding, financial analysis and customer insight into one entry point, connectable to Google Business Profile, Google Ads, Google Workspace and third-party apps such as Adobe, Canva and Shopify. Businesses define a brand style by uploading their website, brand kit, fonts or color palettes, and with the image model Nano Banana 2.1 Gemini produces end-to-end visuals while retaining that style, so guidelines need not be re-uploaded. Gemini also turns messy spreadsheets, receipts and purchase records into financial graphics and interactive models, and these features begin rolling out globally to all Gemini app users in the coming days.

Google has added a business landing page inside Gemini that gathers branding, financial analysis and customer insight into one entry point, connectable to Google Business Profile, Google Ads, Google Workspace and third-party apps such as Adobe, Canva and Shopify. Businesses define a brand style by uploading their website, brand kit, fonts or color palettes, and with the image model Nano Banana 2.1 Gemini produces end-to-end visuals while retaining that style, so guidelines need not be re-uploaded. Gemini also turns messy spreadsheets, receipts and purchase records into financial graphics and interactive models, and these features begin rolling out globally to all Gemini app users in the coming days.

Fading, stains and water marks on Dunhuang murals are semi-transparent damage that still leaves pigment and contours visible, yet a binary mask marks only where damage lies and cannot describe how severity varies inside that region. SC-PAD estimates a continuous terminal severity field from the degraded image and a coarse mask, reparameterizes a shared diffusion progress into pixel-wise local restoration progress, and drives image-space residual activation through a dynamic Alpha state with consistent endpoints and monotonic evolution. On a fixed 100-image test split, replacing the binary mask with predicted soft Alpha raised PSNR from 27.3421 dB to 30.6122 dB, while ground-truth Alpha added a further 3.2539 dB, marking severity estimation as the remaining bottleneck.

Fading, stains and water marks on Dunhuang murals are semi-transparent damage that still leaves pigment and contours visible, yet a binary mask marks only where damage lies and cannot describe how severity varies inside that region. SC-PAD estimates a continuous terminal severity field from the degraded image and a coarse mask, reparameterizes a shared diffusion progress into pixel-wise local restoration progress, and drives image-space residual activation through a dynamic Alpha state with consistent endpoints and monotonic evolution. On a fixed 100-image test split, replacing the binary mask with predicted soft Alpha raised PSNR from 27.3421 dB to 30.6122 dB, while ground-truth Alpha added a further 3.2539 dB, marking severity estimation as the remaining bottleneck.

Pollen microscopy recognition must separate subtle differences between species while absorbing large variation within a species, and the heat-conduction vision backbone vHeat runs a complete two-dimensional heat-diffusion transform on every channel. To cut that redundancy, part of the channels are routed through a lightweight local branch, a controlled frequency-retention constraint is applied to mid-to-late DCT responses, and both ratios are set per network stage. On the in-house Pollen149 set, parameters fall from 30.20M to 25.69M and FLOPs from 4.94G to 3.97G with a 0.19 percentage-point Top-1 drop, while throughput rises from 924.15 to 1015.27 images/s.

Pollen microscopy recognition must separate subtle differences between species while absorbing large variation within a species, and the heat-conduction vision backbone vHeat runs a complete two-dimensional heat-diffusion transform on every channel. To cut that redundancy, part of the channels are routed through a lightweight local branch, a controlled frequency-retention constraint is applied to mid-to-late DCT responses, and both ratios are set per network stage. On the in-house Pollen149 set, parameters fall from 30.20M to 25.69M and FLOPs from 4.94G to 3.97G with a 0.19 percentage-point Top-1 drop, while throughput rises from 924.15 to 1015.27 images/s.