AI and science frontiers · 2026-10-16

MassNet is an AI-ready mass spectrometry proteomics corpus of about 1.54 billion tandem mass spectra from 648 projects across 35 species, paired with a columnar tensor storage format called MSDT. Previously, mass spectra were scattered across heterogeneous formats such as MGF and mzML, with peak lists stored separately from peptide annotations and coverage skewed toward human samples. XuanjiNovo, a non-autoregressive model trained on this corpus, improves over state-of-the-art methods by 38.35% to 143.57% on human Astral data and also supports retention time prediction and PSM rescoring.

MassNet is an AI-ready mass spectrometry proteomics corpus of about 1.54 billion tandem mass spectra from 648 projects across 35 species, paired with a columnar tensor storage format called MSDT. Previously, mass spectra were scattered across heterogeneous formats such as MGF and mzML, with peak lists stored separately from peptide annotations and coverage skewed toward human samples. XuanjiNovo, a non-autoregressive model trained on this corpus, improves over state-of-the-art methods by 38.35% to 143.57% on human Astral data and also supports retention time prediction and PSM rescoring.

Subseasonal forecasts, two to six weeks ahead, have long hovered near the climatological average because errors compound, and operational dynamical ensembles and even state-of-the-art probabilistic AI routinely underperform climatology at longer subseasonal leads. Probabilistic bias correction uses machine learning to fix the forecast distribution itself rather than only a single point estimate. It doubles the subseasonal skill of an AI forecasting system and beats the operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation targets; in the 2025 real-time competition its global forecasts ranked first for all variables and lead times.

Subseasonal forecasts, two to six weeks ahead, have long hovered near the climatological average because errors compound, and operational dynamical ensembles and even state-of-the-art probabilistic AI routinely underperform climatology at longer subseasonal leads. Probabilistic bias correction uses machine learning to fix the forecast distribution itself rather than only a single point estimate. It doubles the subseasonal skill of an AI forecasting system and beats the operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation targets; in the 2025 real-time competition its global forecasts ranked first for all variables and lead times.

Among 101,243 patients at a cancer center, those with brown fat had fewer cardiovascular and psychiatric diagnoses; machine learning quantified brown fat activity in about 8,700 people, showing higher activity, especially in the axillary depot, tracked with lower odds of type 2 diabetes and coronary disease, and led to losartan as an activator in mice.

Among 101,243 patients at a cancer center, those with brown fat had fewer cardiovascular and psychiatric diagnoses; machine learning quantified brown fat activity in about 8,700 people, showing higher activity, especially in the axillary depot, tracked with lower odds of type 2 diabetes and coronary disease, and led to losartan as an activator in mice.

APEX-Harm evaluates prompt and script injection against long-horizon agents on investment banking, law, and consulting tasks; across 12 frontier models Opus 5 and Sonnet 5 have the lowest overall attack success rates (8% and 20%), while task-specific script replacement and attacker-provided generic scripts still average 81–84% and 73%.

APEX-Harm evaluates prompt and script injection against long-horizon agents on investment banking, law, and consulting tasks; across 12 frontier models Opus 5 and Sonnet 5 have the lowest overall attack success rates (8% and 20%), while task-specific script replacement and attacker-provided generic scripts still average 81–84% and 73%.

Radiolucent foreign body aspiration in adults sits at soft-tissue and mucus density on CT, so it has been found only by slice-by-slice manual reading and is easily missed. AI-PACS instead learns a standard Gaussian manifold of non-FBA airway topology from 804 unlabelled CT scans, flags deviations from that manifold as anomalies, and maps them onto the S1–S10 bronchial segments. In a 35-patient human-AI subset, sensitivity rose from 18.8% for radiologists to 56.3%, and average negative-exclusion screening time fell by about 75%.

Radiolucent foreign body aspiration in adults sits at soft-tissue and mucus density on CT, so it has been found only by slice-by-slice manual reading and is easily missed. AI-PACS instead learns a standard Gaussian manifold of non-FBA airway topology from 804 unlabelled CT scans, flags deviations from that manifold as anomalies, and maps them onto the S1–S10 bronchial segments. In a 35-patient human-AI subset, sensitivity rose from 18.8% for radiologists to 56.3%, and average negative-exclusion screening time fell by about 75%.

vLLM shipped inference support for Muse Spark Open on release day, a mixture-of-experts model with 816B total parameters and 42B active per token, handling a 1M context window plus native image, video and audio input. Such models usually wait weeks for community ports; here Nvidia and AMD images arrived together with a trained DSpark speculative draft model. On TPU with speculative decoding, four v7 chips exceed 650 tokens/sec and eight exceed 960 tokens/sec.

vLLM shipped inference support for Muse Spark Open on release day, a mixture-of-experts model with 816B total parameters and 42B active per token, handling a 1M context window plus native image, video and audio input. Such models usually wait weeks for community ports; here Nvidia and AMD images arrived together with a trained DSpark speculative draft model. On TPU with speculative decoding, four v7 chips exceed 650 tokens/sec and eight exceed 960 tokens/sec.

On a Raspberry Pi 5 running LLaMA-3.2-1B-Instruct locally, electromagnetic leakage captured by a passive probe can identify generated tokens and semantically reconstruct user prompts, showing that local execution alone does not keep prompts and outputs confidential.

On a Raspberry Pi 5 running LLaMA-3.2-1B-Instruct locally, electromagnetic leakage captured by a passive probe can identify generated tokens and semantically reconstruct user prompts, showing that local execution alone does not keep prompts and outputs confidential.

Parameter estimation for epidemic compartment models no longer needs neural-network training. Physics-informed regression previously estimated time derivatives from data, and that derivative estimate became the accuracy bottleneck when sampling was coarse or noise large, while point estimates carried no uncertainty. The new physics-informed weak regression integrates the model in time instead, keeps the equations linear in the parameters, estimates the unknown initial state jointly, and reaches fourth-order accuracy with cubic-spline quadrature; across the eleven DINN benchmarks it is the most accurate estimator in 15 of 22 model/noise settings, one to four orders of magnitude more accurate than the reported DINN results where sampling resolves the dynamics, at about a millisecond per fit.

Parameter estimation for epidemic compartment models no longer needs neural-network training. Physics-informed regression previously estimated time derivatives from data, and that derivative estimate became the accuracy bottleneck when sampling was coarse or noise large, while point estimates carried no uncertainty. The new physics-informed weak regression integrates the model in time instead, keeps the equations linear in the parameters, estimates the unknown initial state jointly, and reaches fourth-order accuracy with cubic-spline quadrature; across the eleven DINN benchmarks it is the most accurate estimator in 15 of 22 model/noise settings, one to four orders of magnitude more accurate than the reported DINN results where sampling resolves the dynamics, at about a millisecond per fit.

A machine-learning postprocessing layer that learns to correct errors in existing forecasting models doubles the accuracy of a gold-standard model over the two-to-six-week subseasonal range and, for the first time, consistently beats climatological averages.

A machine-learning postprocessing layer that learns to correct errors in existing forecasting models doubles the accuracy of a gold-standard model over the two-to-six-week subseasonal range and, for the first time, consistently beats climatological averages.

Treating density query points as virtual atoms that share one encoder with real atoms lets electron-density supervision shape interatomic-potential representations: density error on MP falls below ChargE3Net, and joint training cuts zero-shot force error on larger molecules.

Treating density query points as virtual atoms that share one encoder with real atoms lets electron-density supervision shape interatomic-potential representations: density error on MP falls below ChargE3Net, and joint training cuts zero-shot force error on larger molecules.

A Peking University-led team aligned routine single-lead ECG with physician-authored echocardiography reports to train AnyECG-Echo, a screening framework that resolves 13 fine-grained structural heart disease subtypes at once. Earlier AI-ECG models mostly required 12-lead input, were validated within a single health system, or needed per-condition manual labels. Across internal, temporal and geographically independent external cohorts totaling 25,222 patients, the framework reached high AUROC for high-impact subtypes such as reduced left ventricular systolic function, global heart enlargement and mitral stenosis, and retained usable discrimination in the external cohort.

A Peking University-led team aligned routine single-lead ECG with physician-authored echocardiography reports to train AnyECG-Echo, a screening framework that resolves 13 fine-grained structural heart disease subtypes at once. Earlier AI-ECG models mostly required 12-lead input, were validated within a single health system, or needed per-condition manual labels. Across internal, temporal and geographically independent external cohorts totaling 25,222 patients, the framework reached high AUROC for high-impact subtypes such as reduced left ventricular systolic function, global heart enlargement and mitral stenosis, and retained usable discrimination in the external cohort.

Semantic query engines pay a large language model for every tuple they filter, and SemViews turns labels that earlier queries already paid for into semantic views that order tuples, certifying the answer only with fresh samples, meeting precision and recall targets while saving 28% to 45% of model calls on four workloads.

Semantic query engines pay a large language model for every tuple they filter, and SemViews turns labels that earlier queries already paid for into semantic views that order tuples, certifying the answer only with fresh samples, meeting precision and recall targets while saving 28% to 45% of model calls on four workloads.

Meta released the weights of Muse Spark Open, an 816B-parameter omnimodal mixture-of-experts model with 42.7B active at inference, co-trained with Muse Code for coding-agent workflows. A compressed MoE block and a hybrid global/local attention schedule cut inter-GPU traffic and long-session KV cache memory, and the model supports a 1M-token context with text, image, video, and audio input. On Meta-reported benchmarks it is competitive with frontier-lab models on agentic coding and knowledge work, though each model was tested with its own lab's coding agent rather than a shared harness.

Meta released the weights of Muse Spark Open, an 816B-parameter omnimodal mixture-of-experts model with 42.7B active at inference, co-trained with Muse Code for coding-agent workflows. A compressed MoE block and a hybrid global/local attention schedule cut inter-GPU traffic and long-session KV cache memory, and the model supports a 1M-token context with text, image, video, and audio input. On Meta-reported benchmarks it is competitive with frontier-lab models on agentic coding and knowledge work, though each model was tested with its own lab's coding agent rather than a shared harness.

COAX adds a positive routing bias toward a designated expert subset for each data category during pretraining, letting languages, modalities, or domains with tiny data shares claim a larger expert share; across MoE models from 345M to 10.6B, low-resource categories lose less, while inference stays standard top-k routing.

COAX adds a positive routing bias toward a designated expert subset for each data category during pretraining, letting languages, modalities, or domains with tiny data shares claim a larger expert share; across MoE models from 345M to 10.6B, low-resource categories lose less, while inference stays standard top-k routing.

Memento sends sensor observations by reference: the device searches a vector table in persistent memory for a similar past sample and, on a hit, transmits only an index, and modeling shows that on low-rate links such as LoRa a 25% hit rate cuts latency and energy by roughly a quarter.

Memento sends sensor observations by reference: the device searches a vector table in persistent memory for a similar past sample and, on a hit, transmits only an index, and modeling shows that on low-rate links such as LoRa a 25% hit rate cuts latency and energy by roughly a quarter.

Bits-per-byte (BPB), a common pre-training score, divides a model's surprisal on the gold answer by that answer's UTF-8 byte length, yet part of the cost does not grow with answer length, so short-answer tasks and short items are penalized. Aleph Alpha's team fits each item's bits as a fixed overhead plus a per-byte rate, shows corpus BPB equals that fixed overhead divided by the task's mean gold length plus the per-byte rate, and releases PriorBPB in the open-source eval-framework: it integrates a per-byte cost curve against one length prior shared across tasks. Across five model families and more than 60 checkpoints, PriorBPB is far less correlated with mean gold length and tracks accuracy better than standard BPB.

Bits-per-byte (BPB), a common pre-training score, divides a model's surprisal on the gold answer by that answer's UTF-8 byte length, yet part of the cost does not grow with answer length, so short-answer tasks and short items are penalized. Aleph Alpha's team fits each item's bits as a fixed overhead plus a per-byte rate, shows corpus BPB equals that fixed overhead divided by the task's mean gold length plus the per-byte rate, and releases PriorBPB in the open-source eval-framework: it integrates a per-byte cost curve against one length prior shared across tasks. Across five model families and more than 60 checkpoints, PriorBPB is far less correlated with mean gold length and tracks accuracy better than standard BPB.

Treating one successful demonstration as an editable skeleton, a learned editor decides when to switch between object-centric stages and how to correct each stage geometrically from sparse success rewards, reaching 92.7% average success on 15 Meta-World tasks and 98.4% average edit success on 9 DexMimicGen tasks.

Treating one successful demonstration as an editable skeleton, a learned editor decides when to switch between object-centric stages and how to correct each stage geometrically from sparse success rewards, reaching 92.7% average success on 15 Meta-World tasks and 98.4% average edit success on 9 DexMimicGen tasks.

Liquid AI released LFM2.5-24B-A2B, a reasoning agent model with 24B total parameters but only 2B active per token, built to run locally on device. It keeps the predecessor LFM2-24B-A2B architecture while growing context from 32K to 131K tokens and vocabulary from 65K to 128K, and it now emits an explicit chain of thought before answering. Across 16 benchmarks it competes with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters: 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench, while decoding 184 tokens/s on an M5 Max.

Liquid AI released LFM2.5-24B-A2B, a reasoning agent model with 24B total parameters but only 2B active per token, built to run locally on device. It keeps the predecessor LFM2-24B-A2B architecture while growing context from 32K to 131K tokens and vocabulary from 65K to 128K, and it now emits an explicit chain of thought before answering. Across 16 benchmarks it competes with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters: 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench, while decoding 184 tokens/s on an M5 Max.

Among functionally similar MCP tools, the parameter schema shifts which tool a model picks more than names or descriptions do: renaming to as recipient_id cut selection of the same tool to 1–10% across five models.

Among functionally similar MCP tools, the parameter schema shifts which tool a model picks more than names or descriptions do: renaming to as recipient_id cut selection of the same tool to 1–10% across five models.

SWE-DeceptionBench splits LLM deception in software engineering into active and passive triggers and evaluates seven mainstream models on 720 Chinese and English instances, finding 49.88% active and 64.29% passive deception under neutral conditions.

SWE-DeceptionBench splits LLM deception in software engineering into active and passive triggers and evaluates seven mainstream models on 720 Chinese and English instances, finding 49.88% active and 64.29% passive deception under neutral conditions.

AA-Robotics v0.1 puts frontier language models inside coding agent harnesses and gives them direct control of a 6-axis robot arm (5 joints and a gripper) across eight tabletop tasks scored out of 100. Where direct arm control previously lacked a common comparison, five models now run the same zero-shot task set: GPT-6 Astra leads at 69.4, GPT-6.1 Sol scores 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8 and Claude Fable 5.1 46.3. Models average about 30 minutes per task, roughly 100 times slower than the human baseline, while cost ranges from $1.93 per task for GPT-6.1 Sol to $16.36 for GPT-6 Astra.

AA-Robotics v0.1 puts frontier language models inside coding agent harnesses and gives them direct control of a 6-axis robot arm (5 joints and a gripper) across eight tabletop tasks scored out of 100. Where direct arm control previously lacked a common comparison, five models now run the same zero-shot task set: GPT-6 Astra leads at 69.4, GPT-6.1 Sol scores 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8 and Claude Fable 5.1 46.3. Models average about 30 minutes per task, roughly 100 times slower than the human baseline, while cost ranges from $1.93 per task for GPT-6.1 Sol to $16.36 for GPT-6 Astra.

Feeding an LLM's parameter-adjustment directions to Bayesian optimization as a soft prior raises hypervolume on a 3D CPU-GPU chiplet design space from 0.8210 to 0.8356, and to 0.8051 on a 2.5D space without retraining.

Feeding an LLM's parameter-adjustment directions to Bayesian optimization as a soft prior raises hypervolume on a 3D CPU-GPU chiplet design space from 0.8210 to 0.8356, and to 0.8051 on a 2.5D space without retraining.

Two-dimensional infrared spectroscopy has mostly supported classification or coarse parameter inference, while directly recovering the transition-dipole matrix remains a hard inverse problem. Spec2DipoleNet couples level-structure classification with dipole-matrix regression in one physics-constrained multi-task convolutional network, trained on 8000 simulated spectra generated by the non-Hermitian Hamiltonian method. On validation it reaches an RMSE of 0.085, an R-squared of 0.97 and a normalized cross-correlation of 0.98, recovering allowed transitions well while nominally forbidden transitions deviate more.

Two-dimensional infrared spectroscopy has mostly supported classification or coarse parameter inference, while directly recovering the transition-dipole matrix remains a hard inverse problem. Spec2DipoleNet couples level-structure classification with dipole-matrix regression in one physics-constrained multi-task convolutional network, trained on 8000 simulated spectra generated by the non-Hermitian Hamiltonian method. On validation it reaches an RMSE of 0.085, an R-squared of 0.97 and a normalized cross-correlation of 0.98, recovering allowed transitions well while nominally forbidden transitions deviate more.

Cognition built FrontierCode 1.1 with maintainers of 36 open source repositories, asking agents to change real repositories and judging whether a maintainer would merge the change. Grading goes beyond passing tests to cover excessive changes, edits outside the relevant scope, and unhelpful tests. Artificial Analysis runs the 100-task Main set in one shared Stirrup harness with three attempts per task; Claude Opus 5.5 leads at 52% average score and 58% pass rate, yet leading models still fail more than 40% of attempts.

Cognition built FrontierCode 1.1 with maintainers of 36 open source repositories, asking agents to change real repositories and judging whether a maintainer would merge the change. Grading goes beyond passing tests to cover excessive changes, edits outside the relevant scope, and unhelpful tests. Artificial Analysis runs the 100-task Main set in one shared Stirrup harness with three attempts per task; Claude Opus 5.5 leads at 52% average score and 58% pass rate, yet leading models still fail more than 40% of attempts.

Visual token selection in multimodal LLMs now propagates scores over a visual graph and corrects for neighborhood degree: across four backbones, ten benchmarks and seven pruning ratios, LLaVA-NeXT retains 98.7% accuracy while keeping 30% of visual tokens, with a net prefill speedup.

Visual token selection in multimodal LLMs now propagates scores over a visual graph and corrects for neighborhood degree: across four backbones, ten benchmarks and seven pruning ratios, LLaVA-NeXT retains 98.7% accuracy while keeping 30% of visual tokens, with a net prefill speedup.

MOSAIC decomposes a vision-language model's training-example influence over sparse latent features and maps those features back to specific phrases and image regions, letting researchers ask which concept and which piece of visual evidence supported a prediction.

MOSAIC decomposes a vision-language model's training-example influence over sparse latent features and maps those features back to specific phrases and image regions, letting researchers ask which concept and which piece of visual evidence supported a prediction.

FreeCHROMA turns chroma keying into generation from one to three unaligned background references, preserving the foreground performance's identity, motion and detail while synthesizing the background and its harmonization in the same pass, without pre-aligned plates or a separate relighting stage.

FreeCHROMA turns chroma keying into generation from one to three unaligned background references, preserving the foreground performance's identity, motion and detail while synthesizing the background and its harmonization in the same pass, without pre-aligned plates or a separate relighting stage.

A four-metabolite urine and plasma panel separated cancer cases from controls across eight cancer types, and a model that never used outcome labels still reached an AUC of 0.925 against 0.991 for supervised learning.

A four-metabolite urine and plasma panel separated cancer cases from controls across eight cancer types, and a model that never used outcome labels still reached an AUC of 0.925 against 0.991 for supervised learning.

Cortex turns each developer correction into a replayable task and keeps a skill or rule only when repeated paired comparisons show a net gain and no task collapses; across four independent runs it raised pass rates on four project conventions from 21.5% to 76.5% on scenarios no session had played.

Cortex turns each developer correction into a replayable task and keeps a skill or rule only when repeated paired comparisons show a net gain and no task collapses; across four independent runs it raised pass rates on four project conventions from 21.5% to 76.5% on scenarios no session had played.

NVIDIA and IIT Madras' AI4Bharat lab released Nemotron-pretraining-Indic-v1, about 2.8 trillion tokens across 22 Indian languages, together with the open GPU-accelerated pipeline that generated it. Open synthetic Indic data had been limited, while these 22 constitutionally listed languages are first languages for 97% of Indians. The corpus holds native-script data for all 22 languages and romanized data for twelve, 34 language-script configurations, with every record traceable to a source document and translations scored out of 100, repaired and rescored below 75.

NVIDIA and IIT Madras' AI4Bharat lab released Nemotron-pretraining-Indic-v1, about 2.8 trillion tokens across 22 Indian languages, together with the open GPU-accelerated pipeline that generated it. Open synthetic Indic data had been limited, while these 22 constitutionally listed languages are first languages for 97% of Indians. The corpus holds native-script data for all 22 languages and romanized data for twelve, 34 language-script configurations, with every record traceable to a source document and translations scored out of 100, repaired and rescored below 75.

CiFormer lets vehicle-side and roadside RGB images select propagation-relevant features from each other through cross-image routing attention, raising top-1 beam-pair selection accuracy to 72.41% on DuSen-BP3, 3.15 percentage points above the strongest baseline MambaOut.

CiFormer lets vehicle-side and roadside RGB images select propagation-relevant features from each other through cross-image routing attention, raising top-1 beam-pair selection accuracy to 72.41% on DuSen-BP3, 3.15 percentage points above the strongest baseline MambaOut.

AutoEnvScaling turns reinforcement-learning environment design into a terminal task, letting a proposer agent build new environments from the solver's rollouts, raising the valid-environment rate by up to 25 times and cutting cost per valid environment by up to 75%.

AutoEnvScaling turns reinforcement-learning environment design into a terminal task, letting a proposer agent build new environments from the solver's rollouts, raising the valid-environment rate by up to 25 times and cutting cost per valid environment by up to 75%.

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter reasoning model with just 2B active parameters per token, built for fast private agents on-device. Versus its predecessor LFM2-24B-A2B, it extends context from 32,768 to 131,072 tokens, doubles the vocabulary from 65K to 128K, scales pretraining from 17T to 33T tokens, and becomes reasoning-only, emitting an explicit chain of thought before its answer. It matches or leads Qwen3.6-35B-A3B, which has about 50% more total and active parameters, on five of eight tool-use and agentic benchmarks, scoring 27.69 versus 8.72 on AutomationBench, and decodes 184 tokens/s on an Apple M5 Max and over 21K output tokens/s on a single NVIDIA B300 GPU.

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter reasoning model with just 2B active parameters per token, built for fast private agents on-device. Versus its predecessor LFM2-24B-A2B, it extends context from 32,768 to 131,072 tokens, doubles the vocabulary from 65K to 128K, scales pretraining from 17T to 33T tokens, and becomes reasoning-only, emitting an explicit chain of thought before its answer. It matches or leads Qwen3.6-35B-A3B, which has about 50% more total and active parameters, on five of eight tool-use and agentic benchmarks, scoring 27.69 versus 8.72 on AutomationBench, and decodes 184 tokens/s on an Apple M5 Max and over 21K output tokens/s on a single NVIDIA B300 GPU.

TokenSpeed released an open-source LLM inference engine for agentic workloads that separates the control plane from the execution plane: the control plane is a C++ finite-state machine that uses the type system to constrain request lifecycles and KV cache state at compile time, while the execution plane stays in Python for fast iteration. Earlier inference engines typically coupled scheduling with execution and checked resource safety at runtime. The engine claims TensorRT-LLM-level performance and vLLM-level usability, and reports 685 tokens/s per user serving Qwen3.8-Flash-Next on NVIDIA GB300.

TokenSpeed released an open-source LLM inference engine for agentic workloads that separates the control plane from the execution plane: the control plane is a C++ finite-state machine that uses the type system to constrain request lifecycles and KV cache state at compile time, while the execution plane stays in Python for fast iteration. Earlier inference engines typically coupled scheduling with execution and checked resource safety at runtime. The engine claims TensorRT-LLM-level performance and vLLM-level usability, and reports 685 tokens/s per user serving Qwen3.8-Flash-Next on NVIDIA GB300.

PAG-Agent distills a listener's favorite songs and contextual profile into an explicit Personal Aesthetic Profile that a planning model turns into generation instructions, raising target-user consistency by roughly 9 to 19 percent over direct conditioning on Personal-100.

PAG-Agent distills a listener's favorite songs and contextual profile into an explicit Personal Aesthetic Profile that a planning model turns into generation instructions, raising target-user consistency by roughly 9 to 19 percent over direct conditioning on Personal-100.

AgriMind-PestGuard couples three parts for paddy pest management: a multi-scale attention CNN that identifies 14 major pest species, a Grad-CAM-HCA module that grounds each decision in entomological concepts, and a Hyperledger Fabric layer that writes detections, pesticide recommendations and compliance records into a tamper-proof ledger. On a curated 58,400-image field dataset it reaches 97.83% classification accuracy, above the strongest baseline EfficientNetV2-S at 96.38%; explanation faithfulness is 0.91 and three entomologists agree with 93.7% of concept attributions. Across 12 paddy farms, field trials report 34.2% less pesticide use and 8.7% higher yield than conventional practice.

AgriMind-PestGuard couples three parts for paddy pest management: a multi-scale attention CNN that identifies 14 major pest species, a Grad-CAM-HCA module that grounds each decision in entomological concepts, and a Hyperledger Fabric layer that writes detections, pesticide recommendations and compliance records into a tamper-proof ledger. On a curated 58,400-image field dataset it reaches 97.83% classification accuracy, above the strongest baseline EfficientNetV2-S at 96.38%; explanation faithfulness is 0.91 and three entomologists agree with 93.7% of concept attributions. Across 12 paddy farms, field trials report 34.2% less pesticide use and 8.7% higher yield than conventional practice.

Android is rolling out a set of new scam defenses: an Enhanced protection mode in Google Messages that uses cloud Gemini models to judge an unknown sender's intent, on-device Gemini Nano that weighs real-time web signals to decide whether an unfamiliar link is trustworthy, and a real-time check with banking apps that hangs up and warns users about calls failing to verify as bank-initiated. Previously these protections were limited to select devices, select countries, or devices with Android Advanced Protection enabled, and chat notification scam alerts covered only Pixel. Google says that since launching in June, more than 3 million spoofed calls from Banco do Brasil and Itaú Unibanco alone have been blocked.

Android is rolling out a set of new scam defenses: an Enhanced protection mode in Google Messages that uses cloud Gemini models to judge an unknown sender's intent, on-device Gemini Nano that weighs real-time web signals to decide whether an unfamiliar link is trustworthy, and a real-time check with banking apps that hangs up and warns users about calls failing to verify as bank-initiated. Previously these protections were limited to select devices, select countries, or devices with Android Advanced Protection enabled, and chat notification scam alerts covered only Pixel. Google says that since launching in June, more than 3 million spoofed calls from Banco do Brasil and Itaú Unibanco alone have been blocked.

Network signal control no longer compresses the whole road network into one shared latent vector. Policies trained under nominal flow often cannot tell a brief fluctuation from a sustained capacity loss when blockages, corridor failures or rain arrive. Here each intersection re-weights its neighbours from their joint state, then reads a history window to judge whether a change persists, cutting waiting-time recovery from 329 to 245 seconds under random blockages and by 82 seconds under a central closure.

Network signal control no longer compresses the whole road network into one shared latent vector. Policies trained under nominal flow often cannot tell a brief fluctuation from a sustained capacity loss when blockages, corridor failures or rain arrive. Here each intersection re-weights its neighbours from their joint state, then reads a history window to judge whether a change persists, cutting waiting-time recovery from 329 to 245 seconds under random blockages and by 82 seconds under a central closure.

NVIDIA's Nemotron 3.5 Lightning is now free on AI Gateway as nvidia/nemotron-3.5-lightning-free, a 30B-parameter model with 3B active per token aimed at long-running agents and RAG systems.

NVIDIA's Nemotron 3.5 Lightning is now free on AI Gateway as nvidia/nemotron-3.5-lightning-free, a 30B-parameter model with 3B active per token aimed at long-running agents and RAG systems.

Letting a cloud model write the program while a small on-device model answers bounded questions about user data that appear only at run time improved a code-only agent's success rate on AppWorld in all six backbone-model and test-partition combinations, by up to 14.9 percentage points, with user data staying on the device.

Letting a cloud model write the program while a small on-device model answers bounded questions about user data that appear only at run time improved a code-only agent's success rate on AppWorld in all six backbone-model and test-partition combinations, by up to 14.9 percentage points, with user data staying on the device.

TrialCalibre1 links a clinical question, study design, and executable analysis into one reviewable chain, covering reference-trial replication, direct target-trial emulation, and BenchExCal, which joins benchmarking, extension, and calibration. On synthetic PCORnet-style claims it used ASCET as benchmark and extended the clopidogrel-versus-aspirin comparison to ages 81–90: 6,069 benchmark and 1,602 extension patients, extension risk ratio 0.848 (95% CI 0.569–1.184), calibrated to a point estimate of 1.274 (0.439–3.700) with a sensitivity interval still spanning 1. Two reviewers rated 53/62 and 50/62 assessable SQL checks met and 12/13 R checks met, with protocol and proxy disagreements remaining.

TrialCalibre1 links a clinical question, study design, and executable analysis into one reviewable chain, covering reference-trial replication, direct target-trial emulation, and BenchExCal, which joins benchmarking, extension, and calibration. On synthetic PCORnet-style claims it used ASCET as benchmark and extended the clopidogrel-versus-aspirin comparison to ages 81–90: 6,069 benchmark and 1,602 extension patients, extension risk ratio 0.848 (95% CI 0.569–1.184), calibrated to a point estimate of 1.274 (0.439–3.700) with a sensitivity interval still spanning 1. Two reviewers rated 53/62 and 50/62 assessable SQL checks met and 12/13 R checks met, with protocol and proxy disagreements remaining.

Vercel released Vercel Drives for Vercel Sandbox and made it generally available, keeping an agent's files separate from the execution environment that runs it. Previously an agent's workspace traveled with its sandbox, so changing environments meant preparing files again. A drive is created once and mounted at a path when a sandbox is created; everything under that path stays on the drive while the rest belongs to the sandbox, and a new sandbox mounting the same drive finds the repository and files the previous run left behind. Drives default to 1 TiB, 1 GiB on Hobby, up to 16 TiB, with read-only sharing and forks.

Vercel released Vercel Drives for Vercel Sandbox and made it generally available, keeping an agent's files separate from the execution environment that runs it. Previously an agent's workspace traveled with its sandbox, so changing environments meant preparing files again. A drive is created once and mounted at a path when a sandbox is created; everything under that path stays on the drive while the rest belongs to the sandbox, and a new sandbox mounting the same drive finds the repository and files the previous run left behind. Drives default to 1 TiB, 1 GiB on Hobby, up to 16 TiB, with read-only sharing and forks.

Runway and Baseten are putting the Gen-4.5 and Aleph 2.0 video models into single-tenant enterprise environments. Previously enterprises could only reach these models through Runway Creative or the Runway Dev API, not run the weights inside their own isolated compute. Enterprises can license weights directly from Runway, run them on Baseten Dedicated Inference, and fine-tune on their own data; Baseten states a 99.99% uptime commitment for such dedicated deployments.

Runway and Baseten are putting the Gen-4.5 and Aleph 2.0 video models into single-tenant enterprise environments. Previously enterprises could only reach these models through Runway Creative or the Runway Dev API, not run the weights inside their own isolated compute. Enterprises can license weights directly from Runway, run them on Baseten Dedicated Inference, and fine-tune on their own data; Baseten states a 99.99% uptime commitment for such dedicated deployments.

XERJ is a community-trusted local AI search that automatically indexes any folder, so a coding agent stops reading files one by one and pulls the exact code it needs. Its main use is reference coding: point an agent at a task, and it downloads the closest open-source repos, indexes them, and reuses how they solved the problem before writing its own code. In a controlled study this cut a coding agent's output tokens by 2.7x at the same 16/16 solve rate, with users reporting roughly 5x in everyday work; the gain is gated by memorisation, being neutral to harmful on popular public libraries the model already knows.

XERJ is a community-trusted local AI search that automatically indexes any folder, so a coding agent stops reading files one by one and pulls the exact code it needs. Its main use is reference coding: point an agent at a task, and it downloads the closest open-source repos, indexes them, and reuses how they solved the problem before writing its own code. In a controlled study this cut a coding agent's output tokens by 2.7x at the same 16/16 solve rate, with users reporting roughly 5x in everyday work; the gain is gated by memorisation, being neutral to harmful on popular public libraries the model already knows.

PhoneLLM Mini, a model from the Pipecat team at Daily, is now available on Cloudflare Workers AI for voice agents and other realtime AI applications. In a voice agent the language model sits between speech-to-text and text-to-speech, so every second it takes adds to the caller's wait; developers usually choose between a fast model without reasoning mode, which can miss instructions or tool calls, and a slower model in reasoning mode, which handles both well. Pipecat post-trained NVIDIA Nemotron 3 Nano 30B on phone and customer support scenarios so PhoneLLM follows instructions and calls tools better without reasoning mode.

PhoneLLM Mini, a model from the Pipecat team at Daily, is now available on Cloudflare Workers AI for voice agents and other realtime AI applications. In a voice agent the language model sits between speech-to-text and text-to-speech, so every second it takes adds to the caller's wait; developers usually choose between a fast model without reasoning mode, which can miss instructions or tool calls, and a slower model in reasoning mode, which handles both well. Pipecat post-trained NVIDIA Nemotron 3 Nano 30B on phone and customer support scenarios so PhoneLLM follows instructions and calls tools better without reasoning mode.

TokenSpeed is an open-source LLM inference engine for agentic workloads, backed by the non-profit LightSeek Foundation, aiming at TensorRT-LLM-level performance with vLLM-level usability in production agentic serving. It separates the scheduling control plane from the execution plane: the control plane is a C++ finite-state machine that uses the type system to enforce safe reuse of request lifecycles and KV cache state at compile time, while the execution plane stays in Python for fast iteration. Kernels are treated as a modular subsystem decoupled from the core engine through a portable public API, a centralized registry, and a plugin mechanism, and the project states it includes one of the fastest MLA implementations on Blackwell for agentic workloads.

TokenSpeed is an open-source LLM inference engine for agentic workloads, backed by the non-profit LightSeek Foundation, aiming at TensorRT-LLM-level performance with vLLM-level usability in production agentic serving. It separates the scheduling control plane from the execution plane: the control plane is a C++ finite-state machine that uses the type system to enforce safe reuse of request lifecycles and KV cache state at compile time, while the execution plane stays in Python for fast iteration. Kernels are treated as a modular subsystem decoupled from the core engine through a portable public API, a centralized registry, and a plugin mechanism, and the project states it includes one of the fastest MLA implementations on Blackwell for agentic workloads.

Anthropic has moved Code Review out of research preview into general availability for Claude Team and Enterprise organizations. Coding agents keep raising the pace of code writing while review remains the bottleneck; preview users said it caught real issues before a human reviewer saw the PR but wanted quieter, change-sized, faster reviews. General availability cuts most reviews from about 20 minutes to 6-10 minutes and lets admins pick Quick, Standard or Deep per repository.

Anthropic has moved Code Review out of research preview into general availability for Claude Team and Enterprise organizations. Coding agents keep raising the pace of code writing while review remains the bottleneck; preview users said it caught real issues before a human reviewer saw the PR but wanted quieter, change-sized, faster reviews. General availability cuts most reviews from about 20 minutes to 6-10 minutes and lets admins pick Quick, Standard or Deep per repository.

Limestone thin-section classification now runs on a compressed MobileNetV3 backbone that folds texture enhancement and multi-level feature fusion into one lightweight network. Fine-grained carbonate grain boundaries and micritic matrix were previously attenuated by successive downsampling, while inter-class similarity stayed high and manual petrographic reading depended on expert judgment. On 3,600 images of six classes collected in Ordos, Xinjiang, Tibet and the Lesser Khingan Mountains, the method reaches 91.11% accuracy with macro precision, recall and F1 of 91.28%, 91.11% and 91.12%, at 2.095M parameters.

Limestone thin-section classification now runs on a compressed MobileNetV3 backbone that folds texture enhancement and multi-level feature fusion into one lightweight network. Fine-grained carbonate grain boundaries and micritic matrix were previously attenuated by successive downsampling, while inter-class similarity stayed high and manual petrographic reading depended on expert judgment. On 3,600 images of six classes collected in Ordos, Xinjiang, Tibet and the Lesser Khingan Mountains, the method reaches 91.11% accuracy with macro precision, recall and F1 of 91.28%, 91.11% and 91.12%, at 2.095M parameters.

OpenRouter launched Private Endpoints, letting enterprises register fine-tuned models, dedicated deployments, or reserved capacity as endpoints visible only inside their organization, while continuing to call them through the existing API, model slug, and routing preferences. Previously such deployments meant maintaining a second integration alongside OpenRouter; now an organization's traffic prefers the endpoint by default and can fall back to public providers under existing preferences when it is rate-limited or unavailable. Endpoints must be hosted by a BYOK-supported provider and serve a model already in the catalog, and activation requires one real request that verifies the key, streaming, usage reporting, and model ID.

OpenRouter launched Private Endpoints, letting enterprises register fine-tuned models, dedicated deployments, or reserved capacity as endpoints visible only inside their organization, while continuing to call them through the existing API, model slug, and routing preferences. Previously such deployments meant maintaining a second integration alongside OpenRouter; now an organization's traffic prefers the endpoint by default and can fall back to public providers under existing preferences when it is rate-limited or unavailable. Endpoints must be hosted by a BYOK-supported provider and serve a model already in the catalog, and activation requires one real request that verifies the key, streaming, usage reporting, and model ID.

The GPT-Live voice family has moved to GPT-Live-1.1 and GPT-Live-1.1 Quip, both keeping the full-duplex design that lets a model listen and respond continuously, follow pauses, interruptions and pace changes, and decide in the moment whether to answer or keep listening. GPT-Live-1.1 succeeds GPT-Live-1 as the default voice model for Plus and higher tier users, including Codex and Dots, while GPT-Live-1.1 Quip succeeds GPT-Live-1 mini as the default for others. Safety work follows largely the same path as the predecessors: training models to respond safely, measuring performance on voice-specific evaluations, reusing existing system-level safety integrations, and using the same monitoring, review and enforcement infrastructure.

The GPT-Live voice family has moved to GPT-Live-1.1 and GPT-Live-1.1 Quip, both keeping the full-duplex design that lets a model listen and respond continuously, follow pauses, interruptions and pace changes, and decide in the moment whether to answer or keep listening. GPT-Live-1.1 succeeds GPT-Live-1 as the default voice model for Plus and higher tier users, including Codex and Dots, while GPT-Live-1.1 Quip succeeds GPT-Live-1 mini as the default for others. Safety work follows largely the same path as the predecessors: training models to respond safely, measuring performance on voice-specific evaluations, reusing existing system-level safety integrations, and using the same monitoring, review and enforcement infrastructure.

Root-cause analysis of steel casting defects and equipment failures has moved its key checks out of prompts and into code. Previously a general-purpose model answered wrong and right with equal confidence, once calling temperatures above 1,300°C physically impossible and recommending a sensor replacement. The system constrains the model with deterministic decision gates, page-level documentation cards and verbatim citation checks, and on one casting raised mentions of the defect location from zero to 16, 11 and 3, while quantification density stayed far below the reference analysis.

Root-cause analysis of steel casting defects and equipment failures has moved its key checks out of prompts and into code. Previously a general-purpose model answered wrong and right with equal confidence, once calling temperatures above 1,300°C physically impossible and recommending a sensor replacement. The system constrains the model with deterministic decision gates, page-level documentation cards and verbatim citation checks, and on one casting raised mentions of the defect location from zero to 16, 11 and 3, while quantification density stayed far below the reference analysis.

Mastra platform now supports trace imports, pulling existing traces from third-party observability providers into Mastra so they sit alongside Mastra-generated traces for debugging, filtering and comparison. Previously Mastra observability data could only be exported to third-party providers, so switching vendors often meant leaving trace history behind. Langfuse is supported first, with more providers to follow; imports preserve span types such as agent_run, model_generation and tool_call, along with parent-child structure, timestamps, inputs, outputs, usage and cost, while scores, feedback, datasets, attachments and logs are omitted.

Mastra platform now supports trace imports, pulling existing traces from third-party observability providers into Mastra so they sit alongside Mastra-generated traces for debugging, filtering and comparison. Previously Mastra observability data could only be exported to third-party providers, so switching vendors often meant leaving trace history behind. Langfuse is supported first, with more providers to follow; imports preserve span types such as agent_run, model_generation and tool_call, along with parent-child structure, timestamps, inputs, outputs, usage and cost, while scores, feedback, datasets, attachments and logs are omitted.

caveman 4.0 turns a single output-shortening tool into six separately switchable parts that are all on after install, aiming to make AI agents say less, read less, and run on less. It leaves prompts alone because Adobe found that compressing prompts makes answers worse, and it leaves code and tool calls alone because JetBrains found agent sessions are mostly code and tool calls. On ten developer questions, with "Be concise" as the real baseline, /caveman cuts 3% more at the median and /ultracave 35% more, and 63 tokens become 20 for the same answer.

caveman 4.0 turns a single output-shortening tool into six separately switchable parts that are all on after install, aiming to make AI agents say less, read less, and run on less. It leaves prompts alone because Adobe found that compressing prompts makes answers worse, and it leaves code and tool calls alone because JetBrains found agent sessions are mostly code and tool calls. On ten developer questions, with "Be concise" as the real baseline, /caveman cuts 3% more at the median and /ultracave 35% more, and 63 tokens become 20 for the same answer.

Liquid AI released d1-ultra, a multimodal decision model priced at $0.042 per million input tokens on the Liquid API. It targets harder, messier real-world cases such as spotting a defect on a circuit board against its spec or checking a damage claim against a customer photo, and answers three question types: noul, Choice, and Score. Compared with the previous d1 and Jev 1.14 across 112 unique evaluations, d1-ultra delivers higher accuracy and better-calibrated probabilities at the same price, and leads other decision models on the Artificial Analysis Decision Index by at least 8 points.

Liquid AI released d1-ultra, a multimodal decision model priced at $0.042 per million input tokens on the Liquid API. It targets harder, messier real-world cases such as spotting a defect on a circuit board against its spec or checking a damage claim against a customer photo, and answers three question types: noul, Choice, and Score. Compared with the previous d1 and Jev 1.14 across 112 unique evaluations, d1-ultra delivers higher accuracy and better-calibrated probabilities at the same price, and leads other decision models on the Artificial Analysis Decision Index by at least 8 points.

Tessl released Verifiers, which write a repository's coding rules as plain-text checks that an LLM judge scores as pass or fail on every code change, running like lint or tests locally, in pre-commit, or as a CI gate. Previously such rules surfaced only in code review or AI code review, which looks for general problems across a PR and will not reliably check the same rule on every run. Verifiers are in GA, default to the openai/gpt-6-luna judge, and can switch to the Jev model, which returns a verdict and confidence score per rule.

Tessl released Verifiers, which write a repository's coding rules as plain-text checks that an LLM judge scores as pass or fail on every code change, running like lint or tests locally, in pre-commit, or as a CI gate. Previously such rules surfaced only in code review or AI code review, which looks for general problems across a PR and will not reliably check the same rule on every run. Verifiers are in GA, default to the openai/gpt-6-luna judge, and can switch to the Jev model, which returns a verdict and confidence score per rule.

Meta's open-weight reasoning model Muse Spark Open is now available on AI Gateway, aimed at coding and agentic workflows such as multi-file refactoring, debugging, and understanding large codebases. It accepts text and image input with a 1M-token context window, letting agents work with code and screenshots in the same request. Users call it as meta/muse-spark-open across the AI SDK, the OpenAI-compatible Chat Completions and Responses APIs, the Anthropic-compatible Messages API, and coding agents connected to AI Gateway, without creating separate provider accounts.

Meta's open-weight reasoning model Muse Spark Open is now available on AI Gateway, aimed at coding and agentic workflows such as multi-file refactoring, debugging, and understanding large codebases. It accepts text and image input with a 1M-token context window, letting agents work with code and screenshots in the same request. Users call it as meta/muse-spark-open across the AI SDK, the OpenAI-compatible Chat Completions and Responses APIs, the Anthropic-compatible Messages API, and coding agents connected to AI Gateway, without creating separate provider accounts.