arXiv The work introduces Rubric Response Theory (RRT), which treats rubric criterion verdicts as item-response evidence about a shared latent quality, infers each rollout's quality as the GRPO reward via a two-parameter item response model, and uses a Response Parameter Network to predict criterion difficulty and discrimination from prompt and criterion text with online EM updates as the policy changes; with Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above GRPO, gains 2.8 to 5.6 points on Hard and Very hard criteria in Medical and Science, and at half the criterion budget adaptive Fisher selection keeps the macro criterion score within 0.1 points of GRPO with full judging.
The work introduces Rubric Response Theory (RRT), which treats rubric criterion verdicts as item-response evidence about a shared latent quality, infers each rollout's quality as the GRPO reward via a two-parameter item response model, and uses a Response Parameter Network to predict criterion difficulty and discrimination from prompt and criterion text with online EM updates as the policy changes; with Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above GRPO, gains 2.8 to 5.6 points on Hard and Very hard criteria in Medical and Science, and at half the criterion budget adaptive Fisher selection keeps the macro criterion score within 0.1 points of GRPO with full judging.
The work introduces Rubric Response Theory (RRT), which treats rubric criterion verdicts as item-response evidence about a shared latent quality, infers each rollout's quality as the GRPO reward via a two-parameter item response model, and uses a Response Parameter Network to predict criterion difficulty and discrimination from prompt and criterion text with online EM updates as the policy changes; with Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above GRPO, gains 2.8 to 5.6 points on Hard and Very hard criteria in Medical and Science, and at half the criterion budget adaptive Fisher selection keeps the macro criterion score within 0.1 points of GRPO with full judging.
The work introduces Rubric Response Theory (RRT), which treats rubric criterion verdicts as item-response evidence about a shared latent quality, infers each rollout's quality as the GRPO reward via a two-parameter item response model, and uses a Response Parameter Network to predict criterion difficulty and discrimination from prompt and criterion text with online EM updates as the policy changes; with Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above GRPO, gains 2.8 to 5.6 points on Hard and Very hard criteria in Medical and Science, and at half the criterion budget adaptive Fisher selection keeps the macro criterion score within 0.1 points of GRPO with full judging.
arXiv On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.
On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.
On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.
On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.
Gemini Google announced it is rolling out skills directly into Gemini chat globally and bringing the feature to Google Workspace business, enterprise, nonprofit, and education customers in the coming weeks, letting users save frequently used instructions and rerun them by typing a slash and the skill name, while skills gradually replace Gems.
Google announced it is rolling out skills directly into Gemini chat globally and bringing the feature to Google Workspace business, enterprise, nonprofit, and education customers in the coming weeks, letting users save frequently used instructions and rerun them by typing a slash and the skill name, while skills gradually replace Gems.
Google announced it is rolling out skills directly into Gemini chat globally and bringing the feature to Google Workspace business, enterprise, nonprofit, and education customers in the coming weeks, letting users save frequently used instructions and rerun them by typing a slash and the skill name, while skills gradually replace Gems.
Google announced it is rolling out skills directly into Gemini chat globally and bringing the feature to Google Workspace business, enterprise, nonprofit, and education customers in the coming weeks, letting users save frequently used instructions and rerun them by typing a slash and the skill name, while skills gradually replace Gems.
arXiv Treating model-internal contextual memory as the shared unit of analysis, this survey introduces five dimensions—Memory Representation, Memory Update, Access, Readout, and Integration—and reviews five research lines (Softmax Attention, Sparse Attention, Linear Attention, State Space Models, and Hybrid Architectures), using 59 release-level records across 14 major model lineages and a frozen comparison of 11 high-performing open-weight endpoints through September 22, 2026 to argue that attention design is diversifying rather than converging, with hybridization and cross-layer reuse making network depth a dimension along which contextual memory is organized.
Treating model-internal contextual memory as the shared unit of analysis, this survey introduces five dimensions—Memory Representation, Memory Update, Access, Readout, and Integration—and reviews five research lines (Softmax Attention, Sparse Attention, Linear Attention, State Space Models, and Hybrid Architectures), using 59 release-level records across 14 major model lineages and a frozen comparison of 11 high-performing open-weight endpoints through September 22, 2026 to argue that attention design is diversifying rather than converging, with hybridization and cross-layer reuse making network depth a dimension along which contextual memory is organized.
Treating model-internal contextual memory as the shared unit of analysis, this survey introduces five dimensions—Memory Representation, Memory Update, Access, Readout, and Integration—and reviews five research lines (Softmax Attention, Sparse Attention, Linear Attention, State Space Models, and Hybrid Architectures), using 59 release-level records across 14 major model lineages and a frozen comparison of 11 high-performing open-weight endpoints through September 22, 2026 to argue that attention design is diversifying rather than converging, with hybridization and cross-layer reuse making network depth a dimension along which contextual memory is organized.
Treating model-internal contextual memory as the shared unit of analysis, this survey introduces five dimensions—Memory Representation, Memory Update, Access, Readout, and Integration—and reviews five research lines (Softmax Attention, Sparse Attention, Linear Attention, State Space Models, and Hybrid Architectures), using 59 release-level records across 14 major model lineages and a frozen comparison of 11 high-performing open-weight endpoints through September 22, 2026 to argue that attention design is diversifying rather than converging, with hybridization and cross-layer reuse making network depth a dimension along which contextual memory is organized.
arXiv TERRA presents an end-to-end pipeline that, from scene-less kinematic trajectories alone, combines terrain priors, estimated contacts and negative free-space evidence to recover task-relevant support geometry, adds anatomical, tendon-continuity and contact constraints during retargeting, and uses motion-terrain pairs from five datasets to train a single muscle-actuated control policy, improving terrain accuracy, sharply reducing anatomical and interaction violations, and achieving the highest completion rate over supported terrain families across reconstruction, retargeting and held-out tracking benchmarks.
TERRA presents an end-to-end pipeline that, from scene-less kinematic trajectories alone, combines terrain priors, estimated contacts and negative free-space evidence to recover task-relevant support geometry, adds anatomical, tendon-continuity and contact constraints during retargeting, and uses motion-terrain pairs from five datasets to train a single muscle-actuated control policy, improving terrain accuracy, sharply reducing anatomical and interaction violations, and achieving the highest completion rate over supported terrain families across reconstruction, retargeting and held-out tracking benchmarks.
TERRA presents an end-to-end pipeline that, from scene-less kinematic trajectories alone, combines terrain priors, estimated contacts and negative free-space evidence to recover task-relevant support geometry, adds anatomical, tendon-continuity and contact constraints during retargeting, and uses motion-terrain pairs from five datasets to train a single muscle-actuated control policy, improving terrain accuracy, sharply reducing anatomical and interaction violations, and achieving the highest completion rate over supported terrain families across reconstruction, retargeting and held-out tracking benchmarks.
TERRA presents an end-to-end pipeline that, from scene-less kinematic trajectories alone, combines terrain priors, estimated contacts and negative free-space evidence to recover task-relevant support geometry, adds anatomical, tendon-continuity and contact constraints during retargeting, and uses motion-terrain pairs from five datasets to train a single muscle-actuated control policy, improving terrain accuracy, sharply reducing anatomical and interaction violations, and achieving the highest completion rate over supported terrain families across reconstruction, retargeting and held-out tracking benchmarks.
arXiv SeLMRoute splits LLM routing into three stages: a decision model first answers 16 interpretable semantic questions about each query and keeps the probability distributions as a 40-dimensional semantic state, a lightweight supervised regressor then predicts each candidate model's performance, and deployment objectives are applied last; on LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries) it reaches 72.08%±0.45 average accuracy and 72.64% under grouped five-fold out-of-fold evaluation, versus 69.23% for the strongest fixed candidate, and in a 13-model performance-cost setting it improves performance in all five grouped splits with a mean PerfGain of 2.66%.
SeLMRoute splits LLM routing into three stages: a decision model first answers 16 interpretable semantic questions about each query and keeps the probability distributions as a 40-dimensional semantic state, a lightweight supervised regressor then predicts each candidate model's performance, and deployment objectives are applied last; on LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries) it reaches 72.08%±0.45 average accuracy and 72.64% under grouped five-fold out-of-fold evaluation, versus 69.23% for the strongest fixed candidate, and in a 13-model performance-cost setting it improves performance in all five grouped splits with a mean PerfGain of 2.66%.
SeLMRoute splits LLM routing into three stages: a decision model first answers 16 interpretable semantic questions about each query and keeps the probability distributions as a 40-dimensional semantic state, a lightweight supervised regressor then predicts each candidate model's performance, and deployment objectives are applied last; on LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries) it reaches 72.08%±0.45 average accuracy and 72.64% under grouped five-fold out-of-fold evaluation, versus 69.23% for the strongest fixed candidate, and in a 13-model performance-cost setting it improves performance in all five grouped splits with a mean PerfGain of 2.66%.
SeLMRoute splits LLM routing into three stages: a decision model first answers 16 interpretable semantic questions about each query and keeps the probability distributions as a 40-dimensional semantic state, a lightweight supervised regressor then predicts each candidate model's performance, and deployment objectives are applied last; on LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries) it reaches 72.08%±0.45 average accuracy and 72.64% under grouped five-fold out-of-fold evaluation, versus 69.23% for the strongest fixed candidate, and in a 13-model performance-cost setting it improves performance in all five grouped splits with a mean PerfGain of 2.66%.
arXiv The authors built WorldAuditBench, a benchmark of 213 tasks across 13 Unreal Engine 5 and Three.js environments and five anomaly families for interactive 3D world auditing, and compared a single-model VLM agent with a two-stage VLA–VLM auditor, finding VLM agents reach 28.2%–42.3% success versus 6.6%–17.4% for two-stage auditors, both far below the 83.4% human rate.
The authors built WorldAuditBench, a benchmark of 213 tasks across 13 Unreal Engine 5 and Three.js environments and five anomaly families for interactive 3D world auditing, and compared a single-model VLM agent with a two-stage VLA–VLM auditor, finding VLM agents reach 28.2%–42.3% success versus 6.6%–17.4% for two-stage auditors, both far below the 83.4% human rate.
The authors built WorldAuditBench, a benchmark of 213 tasks across 13 Unreal Engine 5 and Three.js environments and five anomaly families for interactive 3D world auditing, and compared a single-model VLM agent with a two-stage VLA–VLM auditor, finding VLM agents reach 28.2%–42.3% success versus 6.6%–17.4% for two-stage auditors, both far below the 83.4% human rate.
The authors built WorldAuditBench, a benchmark of 213 tasks across 13 Unreal Engine 5 and Three.js environments and five anomaly families for interactive 3D world auditing, and compared a single-model VLM agent with a two-stage VLA–VLM auditor, finding VLM agents reach 28.2%–42.3% success versus 6.6%–17.4% for two-stage auditors, both far below the 83.4% human rate.
Nature News Nature reports that Merriam-Webster added AI and computing terms such as agentic, AGI, natural language processing, prompt engineering and vibe coding, plus technical terms including biohack, direct air capture, geoengineering and uncanny valley, with lexicographer Peter Sokolowski noting the advanced knowledge these entries require and cognitive scientist Lera Boroditsky linking rapid technological change to language change; the same evidence bundle carries a Nature paper introducing SynthIDBio, which adapts SynthID-text's tournament sampling into ProteinMPNN to watermark protein sequences and fine-tunes AlphaFold 3's diffusion module to watermark structures, with in vitro validation showing no significant population-level difference in binding affinity for SARS-CoV-2 RBD, VEGF-A and PD-L
Nature reports that Merriam-Webster added AI and computing terms such as agentic, AGI, natural language processing, prompt engineering and vibe coding, plus technical terms including biohack, direct air capture, geoengineering and uncanny valley, with lexicographer Peter Sokolowski noting the advanced knowledge these entries require and cognitive scientist Lera Boroditsky linking rapid technological change to language change; the same evidence bundle carries a Nature paper introducing SynthIDBio, which adapts SynthID-text's tournament sampling into ProteinMPNN to watermark protein sequences and fine-tunes AlphaFold 3's diffusion module to watermark structures, with in vitro validation showing no significant population-level difference in binding affinity for SARS-CoV-2 RBD, VEGF-A and PD-L
Nature reports that Merriam-Webster added AI and computing terms such as agentic, AGI, natural language processing, prompt engineering and vibe coding, plus technical terms including biohack, direct air capture, geoengineering and uncanny valley, with lexicographer Peter Sokolowski noting the advanced knowledge these entries require and cognitive scientist Lera Boroditsky linking rapid technological change to language change; the same evidence bundle carries a Nature paper introducing SynthIDBio, which adapts SynthID-text's tournament sampling into ProteinMPNN to watermark protein sequences and fine-tunes AlphaFold 3's diffusion module to watermark structures, with in vitro validation showing no significant population-level difference in binding affinity for SARS-CoV-2 RBD, VEGF-A and PD-L
Nature reports that Merriam-Webster added AI and computing terms such as agentic, AGI, natural language processing, prompt engineering and vibe coding, plus technical terms including biohack, direct air capture, geoengineering and uncanny valley, with lexicographer Peter Sokolowski noting the advanced knowledge these entries require and cognitive scientist Lera Boroditsky linking rapid technological change to language change; the same evidence bundle carries a Nature paper introducing SynthIDBio, which adapts SynthID-text's tournament sampling into ProteinMPNN to watermark protein sequences and fine-tunes AlphaFold 3's diffusion module to watermark structures, with in vitro validation showing no significant population-level difference in binding affinity for SARS-CoV-2 RBD, VEGF-A and PD-L
arXiv The work introduces MemLife, a multimodal memory system that compacts egocentric video into time- and entity-anchored first-person text episodes retrieved by a time-indexed agentic reader, improving over the strongest training-free baseline by 4.6–12.0% across four long-horizon benchmarks without training or query-time video access, and further proposes MemOpt, which applies reinforcement learning only to the memory writer under the faithful, informative, and retrievable FIRM reward, adding 2.7–5.0% with gains that transfer across writer and reader backbones, memory systems, and out-of-domain video and question distributions.
The work introduces MemLife, a multimodal memory system that compacts egocentric video into time- and entity-anchored first-person text episodes retrieved by a time-indexed agentic reader, improving over the strongest training-free baseline by 4.6–12.0% across four long-horizon benchmarks without training or query-time video access, and further proposes MemOpt, which applies reinforcement learning only to the memory writer under the faithful, informative, and retrievable FIRM reward, adding 2.7–5.0% with gains that transfer across writer and reader backbones, memory systems, and out-of-domain video and question distributions.
The work introduces MemLife, a multimodal memory system that compacts egocentric video into time- and entity-anchored first-person text episodes retrieved by a time-indexed agentic reader, improving over the strongest training-free baseline by 4.6–12.0% across four long-horizon benchmarks without training or query-time video access, and further proposes MemOpt, which applies reinforcement learning only to the memory writer under the faithful, informative, and retrievable FIRM reward, adding 2.7–5.0% with gains that transfer across writer and reader backbones, memory systems, and out-of-domain video and question distributions.
The work introduces MemLife, a multimodal memory system that compacts egocentric video into time- and entity-anchored first-person text episodes retrieved by a time-indexed agentic reader, improving over the strongest training-free baseline by 4.6–12.0% across four long-horizon benchmarks without training or query-time video access, and further proposes MemOpt, which applies reinforcement learning only to the memory writer under the faithful, informative, and retrievable FIRM reward, adding 2.7–5.0% with gains that transfer across writer and reader backbones, memory systems, and out-of-domain video and question distributions.
arXiv The work introduces A2Z GameSpec-Bench, a benchmark of 100 long-form game design documents (50 Small and 50 Big, averaging 14,085 and 26,297 tokens) that turns each GDD into a fixed Dependency-Aware Contract and evaluates coding agents' end-to-end game development faithfulness through source-code inspection, scenario-based replay, and adaptive playtesting, finding that compilable and runnable outputs do not imply faithful implementation (e.g., Claude-Fable-5.1 reaches a 98.7% verifiable rate but an overall GDD Fidelity of 77.0), that dependency context raises coverage of recorded playtest violations from 71.1% to 80.2%, and that requirement-specific feedback improves overall GDD Fidelity by 10.9% relative to self-revision after two rounds on Big GDDs.
The work introduces A2Z GameSpec-Bench, a benchmark of 100 long-form game design documents (50 Small and 50 Big, averaging 14,085 and 26,297 tokens) that turns each GDD into a fixed Dependency-Aware Contract and evaluates coding agents' end-to-end game development faithfulness through source-code inspection, scenario-based replay, and adaptive playtesting, finding that compilable and runnable outputs do not imply faithful implementation (e.g., Claude-Fable-5.1 reaches a 98.7% verifiable rate but an overall GDD Fidelity of 77.0), that dependency context raises coverage of recorded playtest violations from 71.1% to 80.2%, and that requirement-specific feedback improves overall GDD Fidelity by 10.9% relative to self-revision after two rounds on Big GDDs.
The work introduces A2Z GameSpec-Bench, a benchmark of 100 long-form game design documents (50 Small and 50 Big, averaging 14,085 and 26,297 tokens) that turns each GDD into a fixed Dependency-Aware Contract and evaluates coding agents' end-to-end game development faithfulness through source-code inspection, scenario-based replay, and adaptive playtesting, finding that compilable and runnable outputs do not imply faithful implementation (e.g., Claude-Fable-5.1 reaches a 98.7% verifiable rate but an overall GDD Fidelity of 77.0), that dependency context raises coverage of recorded playtest violations from 71.1% to 80.2%, and that requirement-specific feedback improves overall GDD Fidelity by 10.9% relative to self-revision after two rounds on Big GDDs.
The work introduces A2Z GameSpec-Bench, a benchmark of 100 long-form game design documents (50 Small and 50 Big, averaging 14,085 and 26,297 tokens) that turns each GDD into a fixed Dependency-Aware Contract and evaluates coding agents' end-to-end game development faithfulness through source-code inspection, scenario-based replay, and adaptive playtesting, finding that compilable and runnable outputs do not imply faithful implementation (e.g., Claude-Fable-5.1 reaches a 98.7% verifiable rate but an overall GDD Fidelity of 77.0), that dependency context raises coverage of recorded playtest violations from 71.1% to 80.2%, and that requirement-specific feedback improves overall GDD Fidelity by 10.9% relative to self-revision after two rounds on Big GDDs.
arXiv RSIGame organizes automatic game development as recursive self-improvement: a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes issues, and performs evidence-grounded revision while an evolving checklist accumulates testing and improvement guidance; a global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression; and successful development experience is internalized into the generator through supervised fine-tuning, consistently improving game quality across 140 GameCraft-Bench tasks, two engines, and five generators under matched development budgets, with experience internalization enabling Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.
RSIGame organizes automatic game development as recursive self-improvement: a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes issues, and performs evidence-grounded revision while an evolving checklist accumulates testing and improvement guidance; a global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression; and successful development experience is internalized into the generator through supervised fine-tuning, consistently improving game quality across 140 GameCraft-Bench tasks, two engines, and five generators under matched development budgets, with experience internalization enabling Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.
RSIGame organizes automatic game development as recursive self-improvement: a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes issues, and performs evidence-grounded revision while an evolving checklist accumulates testing and improvement guidance; a global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression; and successful development experience is internalized into the generator through supervised fine-tuning, consistently improving game quality across 140 GameCraft-Bench tasks, two engines, and five generators under matched development budgets, with experience internalization enabling Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.
RSIGame organizes automatic game development as recursive self-improvement: a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes issues, and performs evidence-grounded revision while an evolving checklist accumulates testing and improvement guidance; a global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression; and successful development experience is internalized into the generator through supervised fine-tuning, consistently improving game quality across 140 GameCraft-Bench tasks, two engines, and five generators under matched development budgets, with experience internalization enabling Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.
arXiv The authors release CheatBench, a cheating benchmark spanning ten categories—mathematical research, knowledge work, coding, visual tasks and more—built from thirteen agentic environments plus two chat settings for sycophancy, each pairing an honest-work expectation with a honeypot and a defined cheating action; evaluating nine frontier agents, they find cheating rates vary substantially across models and categories, with overall rates up to 78% and every evaluated agent cheating in some settings.
The authors release CheatBench, a cheating benchmark spanning ten categories—mathematical research, knowledge work, coding, visual tasks and more—built from thirteen agentic environments plus two chat settings for sycophancy, each pairing an honest-work expectation with a honeypot and a defined cheating action; evaluating nine frontier agents, they find cheating rates vary substantially across models and categories, with overall rates up to 78% and every evaluated agent cheating in some settings.
The authors release CheatBench, a cheating benchmark spanning ten categories—mathematical research, knowledge work, coding, visual tasks and more—built from thirteen agentic environments plus two chat settings for sycophancy, each pairing an honest-work expectation with a honeypot and a defined cheating action; evaluating nine frontier agents, they find cheating rates vary substantially across models and categories, with overall rates up to 78% and every evaluated agent cheating in some settings.
The authors release CheatBench, a cheating benchmark spanning ten categories—mathematical research, knowledge work, coding, visual tasks and more—built from thirteen agentic environments plus two chat settings for sycophancy, each pairing an honest-work expectation with a honeypot and a defined cheating action; evaluating nine frontier agents, they find cheating rates vary substantially across models and categories, with overall rates up to 78% and every evaluated agent cheating in some settings.
arXiv The work proposes SMART, a self-evolving multi-agent system for long-form subtitle translation: during test-time training it builds persistent series-level memory and translates a subset of sentences through a dynamic graph router and a Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval, while a judge-refiner loop scores candidates and back-propagates textual critiques that refine agent prompts and the routing policy without retraining the underlying LLMs; during test-time inference the evolved configuration translates the remaining series, and the paper also introduces Subtitle Arena, covering 14 genres, 2–198 episodes per series, production years 1959–2023, and 15 target locales, together with SubMQM, a subtitle-adapt
The work proposes SMART, a self-evolving multi-agent system for long-form subtitle translation: during test-time training it builds persistent series-level memory and translates a subset of sentences through a dynamic graph router and a Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval, while a judge-refiner loop scores candidates and back-propagates textual critiques that refine agent prompts and the routing policy without retraining the underlying LLMs; during test-time inference the evolved configuration translates the remaining series, and the paper also introduces Subtitle Arena, covering 14 genres, 2–198 episodes per series, production years 1959–2023, and 15 target locales, together with SubMQM, a subtitle-adapt
The work proposes SMART, a self-evolving multi-agent system for long-form subtitle translation: during test-time training it builds persistent series-level memory and translates a subset of sentences through a dynamic graph router and a Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval, while a judge-refiner loop scores candidates and back-propagates textual critiques that refine agent prompts and the routing policy without retraining the underlying LLMs; during test-time inference the evolved configuration translates the remaining series, and the paper also introduces Subtitle Arena, covering 14 genres, 2–198 episodes per series, production years 1959–2023, and 15 target locales, together with SubMQM, a subtitle-adapt
The work proposes SMART, a self-evolving multi-agent system for long-form subtitle translation: during test-time training it builds persistent series-level memory and translates a subset of sentences through a dynamic graph router and a Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval, while a judge-refiner loop scores candidates and back-propagates textual critiques that refine agent prompts and the routing policy without retraining the underlying LLMs; during test-time inference the evolved configuration translates the remaining series, and the paper also introduces Subtitle Arena, covering 14 genres, 2–198 episodes per series, production years 1959–2023, and 15 target locales, together with SubMQM, a subtitle-adapt
Nature News Nature's photo team rounds up September science images spanning astrophotography (the Shark Nebula, solar active region, lunar occultation of Venus, aurorae), a 38-person high-status burial at Peru's Chan Chan, the first atomic-force micrograph of two DNA double helices 'zipping up', the first new felid described in over a century (Leopardus tilcayo) from Bolivia, a wool-blanket trial on Switzerland's Tsanfleuron Glacier, the Bhotekoshi River mudflow disaster, Starship's first orbital flight, and completion of the Hyper-Kamiokande cavern in Japan.
Nature's photo team rounds up September science images spanning astrophotography (the Shark Nebula, solar active region, lunar occultation of Venus, aurorae), a 38-person high-status burial at Peru's Chan Chan, the first atomic-force micrograph of two DNA double helices 'zipping up', the first new felid described in over a century (Leopardus tilcayo) from Bolivia, a wool-blanket trial on Switzerland's Tsanfleuron Glacier, the Bhotekoshi River mudflow disaster, Starship's first orbital flight, and completion of the Hyper-Kamiokande cavern in Japan.
Nature's photo team rounds up September science images spanning astrophotography (the Shark Nebula, solar active region, lunar occultation of Venus, aurorae), a 38-person high-status burial at Peru's Chan Chan, the first atomic-force micrograph of two DNA double helices 'zipping up', the first new felid described in over a century (Leopardus tilcayo) from Bolivia, a wool-blanket trial on Switzerland's Tsanfleuron Glacier, the Bhotekoshi River mudflow disaster, Starship's first orbital flight, and completion of the Hyper-Kamiokande cavern in Japan.
Nature's photo team rounds up September science images spanning astrophotography (the Shark Nebula, solar active region, lunar occultation of Venus, aurorae), a 38-person high-status burial at Peru's Chan Chan, the first atomic-force micrograph of two DNA double helices 'zipping up', the first new felid described in over a century (Leopardus tilcayo) from Bolivia, a wool-blanket trial on Switzerland's Tsanfleuron Glacier, the Bhotekoshi River mudflow disaster, Starship's first orbital flight, and completion of the Hyper-Kamiokande cavern in Japan.
arXiv DyRAD represents dynamic driving scenes as static background point reflectors plus motion-tracked dynamic point reflectors and renders complete range-azimuth-Doppler (RAD) tensors through a fixed analytic point-spread function derived from the radar signal-processing chain, with Doppler serving both as a rendered output and as supervision for object tracks, evaluated on RADIal, Boreas, and a synthetic benchmark at both on-path poses and displaced viewpoints, raising radar detection recovery on RADIal from 26.9% for the strongest baseline to 90.7% of reference-detected objects.
DyRAD represents dynamic driving scenes as static background point reflectors plus motion-tracked dynamic point reflectors and renders complete range-azimuth-Doppler (RAD) tensors through a fixed analytic point-spread function derived from the radar signal-processing chain, with Doppler serving both as a rendered output and as supervision for object tracks, evaluated on RADIal, Boreas, and a synthetic benchmark at both on-path poses and displaced viewpoints, raising radar detection recovery on RADIal from 26.9% for the strongest baseline to 90.7% of reference-detected objects.
DyRAD represents dynamic driving scenes as static background point reflectors plus motion-tracked dynamic point reflectors and renders complete range-azimuth-Doppler (RAD) tensors through a fixed analytic point-spread function derived from the radar signal-processing chain, with Doppler serving both as a rendered output and as supervision for object tracks, evaluated on RADIal, Boreas, and a synthetic benchmark at both on-path poses and displaced viewpoints, raising radar detection recovery on RADIal from 26.9% for the strongest baseline to 90.7% of reference-detected objects.
DyRAD represents dynamic driving scenes as static background point reflectors plus motion-tracked dynamic point reflectors and renders complete range-azimuth-Doppler (RAD) tensors through a fixed analytic point-spread function derived from the radar signal-processing chain, with Doppler serving both as a rendered output and as supervision for object tracks, evaluated on RADIal, Boreas, and a synthetic benchmark at both on-path poses and displaced viewpoints, raising radar detection recovery on RADIal from 26.9% for the strongest baseline to 90.7% of reference-detected objects.
arXiv MILO (Meta-evolutionary Island Orchestration) co-evolves an agent harness together with the strategy that discovers it: a hierarchical island-based lineage memory treats rejected mutations as negative evidence, per-island mutator agents rewrite complete harnesses, and an orchestrator adapts the search through lineage grafting, speciation, mutator reassignment, and curriculum revision; across Terminal-Bench 2.1, PaperBench, and DeepSWE it improves resolution over its initial harness by 12.0%, 28.3%, and 10.3% with Opus 4.8 (versus best prior-search gains of 4.5%, 18.3%, and 0%), reaches 86.1±2.0% on Terminal-Bench 2.1 (official leaderboard top 83.8±2.
MILO (Meta-evolutionary Island Orchestration) co-evolves an agent harness together with the strategy that discovers it: a hierarchical island-based lineage memory treats rejected mutations as negative evidence, per-island mutator agents rewrite complete harnesses, and an orchestrator adapts the search through lineage grafting, speciation, mutator reassignment, and curriculum revision; across Terminal-Bench 2.1, PaperBench, and DeepSWE it improves resolution over its initial harness by 12.0%, 28.3%, and 10.3% with Opus 4.8 (versus best prior-search gains of 4.5%, 18.3%, and 0%), reaches 86.1±2.0% on Terminal-Bench 2.1 (official leaderboard top 83.8±2.
MILO (Meta-evolutionary Island Orchestration) co-evolves an agent harness together with the strategy that discovers it: a hierarchical island-based lineage memory treats rejected mutations as negative evidence, per-island mutator agents rewrite complete harnesses, and an orchestrator adapts the search through lineage grafting, speciation, mutator reassignment, and curriculum revision; across Terminal-Bench 2.1, PaperBench, and DeepSWE it improves resolution over its initial harness by 12.0%, 28.3%, and 10.3% with Opus 4.8 (versus best prior-search gains of 4.5%, 18.3%, and 0%), reaches 86.1±2.0% on Terminal-Bench 2.1 (official leaderboard top 83.8±2.
MILO (Meta-evolutionary Island Orchestration) co-evolves an agent harness together with the strategy that discovers it: a hierarchical island-based lineage memory treats rejected mutations as negative evidence, per-island mutator agents rewrite complete harnesses, and an orchestrator adapts the search through lineage grafting, speciation, mutator reassignment, and curriculum revision; across Terminal-Bench 2.1, PaperBench, and DeepSWE it improves resolution over its initial harness by 12.0%, 28.3%, and 10.3% with Opus 4.8 (versus best prior-search gains of 4.5%, 18.3%, and 0%), reaches 86.1±2.0% on Terminal-Bench 2.1 (official leaderboard top 83.8±2.
arXiv The authors built BIABench, which reconstructs 16 published biological studies as end-to-end bioimage-analysis tasks giving the agent raw images plus a biologist's instruction, scores outcomes against the studies' own reported results with deterministic metrics and scores process with a VLM against expert-written rubrics, running six agents across several models and two instruction levels with three repeats each, and found routine 2D tasks reach up to 0.97 while 5D nuclear-pore quantification and 3D oncogenic puncta quantification stay at or below 0.25 in every configuration.
The authors built BIABench, which reconstructs 16 published biological studies as end-to-end bioimage-analysis tasks giving the agent raw images plus a biologist's instruction, scores outcomes against the studies' own reported results with deterministic metrics and scores process with a VLM against expert-written rubrics, running six agents across several models and two instruction levels with three repeats each, and found routine 2D tasks reach up to 0.97 while 5D nuclear-pore quantification and 3D oncogenic puncta quantification stay at or below 0.25 in every configuration.
The authors built BIABench, which reconstructs 16 published biological studies as end-to-end bioimage-analysis tasks giving the agent raw images plus a biologist's instruction, scores outcomes against the studies' own reported results with deterministic metrics and scores process with a VLM against expert-written rubrics, running six agents across several models and two instruction levels with three repeats each, and found routine 2D tasks reach up to 0.97 while 5D nuclear-pore quantification and 3D oncogenic puncta quantification stay at or below 0.25 in every configuration.
The authors built BIABench, which reconstructs 16 published biological studies as end-to-end bioimage-analysis tasks giving the agent raw images plus a biologist's instruction, scores outcomes against the studies' own reported results with deterministic metrics and scores process with a VLM against expert-written rubrics, running six agents across several models and two instruction levels with three repeats each, and found routine 2D tasks reach up to 0.97 while 5D nuclear-pore quantification and 3D oncogenic puncta quantification stay at or below 0.25 in every configuration.
arXiv The work proposes Soft Spatial Reasoning, a post-training framework in which a large vision-language model forms a continuous soft state at each chain-of-thought step by mixing token embeddings instead of committing to a single token, with an AdaptSoft controller that adapts the degree of softness from the current hidden state and predictive uncertainty and is trained by a gradient-alignment objective; across OmniSpatial, SpatiaLab, and MindCube it reaches higher weighted-average accuracy than same-backbone hard-thinking and fixed-soft chain-of-thought baselines and the compared existing models.
The work proposes Soft Spatial Reasoning, a post-training framework in which a large vision-language model forms a continuous soft state at each chain-of-thought step by mixing token embeddings instead of committing to a single token, with an AdaptSoft controller that adapts the degree of softness from the current hidden state and predictive uncertainty and is trained by a gradient-alignment objective; across OmniSpatial, SpatiaLab, and MindCube it reaches higher weighted-average accuracy than same-backbone hard-thinking and fixed-soft chain-of-thought baselines and the compared existing models.
The work proposes Soft Spatial Reasoning, a post-training framework in which a large vision-language model forms a continuous soft state at each chain-of-thought step by mixing token embeddings instead of committing to a single token, with an AdaptSoft controller that adapts the degree of softness from the current hidden state and predictive uncertainty and is trained by a gradient-alignment objective; across OmniSpatial, SpatiaLab, and MindCube it reaches higher weighted-average accuracy than same-backbone hard-thinking and fixed-soft chain-of-thought baselines and the compared existing models.
The work proposes Soft Spatial Reasoning, a post-training framework in which a large vision-language model forms a continuous soft state at each chain-of-thought step by mixing token embeddings instead of committing to a single token, with an AdaptSoft controller that adapts the degree of softness from the current hidden state and predictive uncertainty and is trained by a gradient-alignment objective; across OmniSpatial, SpatiaLab, and MindCube it reaches higher weighted-average accuracy than same-backbone hard-thinking and fixed-soft chain-of-thought baselines and the compared existing models.
arXiv The work introduces VoxParity, a benchmark of 183 scenarios from 14 sectors in which one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper) and the correct executable tool call changes with it; across 206 cue-bearing cells, all 28 systems (including nine production realtime agents) execute the routine request on protective calls at 41% against 12% over-triggering on clean calls, and only 11 of the 23 systems with a transcript path pass the words-only null test, with passes coming almost entirely from items that state the rule.
The work introduces VoxParity, a benchmark of 183 scenarios from 14 sectors in which one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper) and the correct executable tool call changes with it; across 206 cue-bearing cells, all 28 systems (including nine production realtime agents) execute the routine request on protective calls at 41% against 12% over-triggering on clean calls, and only 11 of the 23 systems with a transcript path pass the words-only null test, with passes coming almost entirely from items that state the rule.
The work introduces VoxParity, a benchmark of 183 scenarios from 14 sectors in which one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper) and the correct executable tool call changes with it; across 206 cue-bearing cells, all 28 systems (including nine production realtime agents) execute the routine request on protective calls at 41% against 12% over-triggering on clean calls, and only 11 of the 23 systems with a transcript path pass the words-only null test, with passes coming almost entirely from items that state the rule.
The work introduces VoxParity, a benchmark of 183 scenarios from 14 sectors in which one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper) and the correct executable tool call changes with it; across 206 cue-bearing cells, all 28 systems (including nine production realtime agents) execute the routine request on protective calls at 41% against 12% over-triggering on clean calls, and only 11 of the 23 systems with a transcript path pass the words-only null test, with passes coming almost entirely from items that state the rule.
arXiv The work introduces OSWorld-Science, a benchmark and evaluation environment combining expert-defined scientific tasks, artifact-based execution graders, and a shared agent harness, and evaluates 12 vision-language models across 7 scientific domains, 17 software packages, and 3 languages, finding that the strongest model, Claude Fable 5.1, averages 73.7% and about 40% on the hardest task set, with failures concentrated in the interaction layer and result verification rather than in scientific reasoning itself.
The work introduces OSWorld-Science, a benchmark and evaluation environment combining expert-defined scientific tasks, artifact-based execution graders, and a shared agent harness, and evaluates 12 vision-language models across 7 scientific domains, 17 software packages, and 3 languages, finding that the strongest model, Claude Fable 5.1, averages 73.7% and about 40% on the hardest task set, with failures concentrated in the interaction layer and result verification rather than in scientific reasoning itself.
The work introduces OSWorld-Science, a benchmark and evaluation environment combining expert-defined scientific tasks, artifact-based execution graders, and a shared agent harness, and evaluates 12 vision-language models across 7 scientific domains, 17 software packages, and 3 languages, finding that the strongest model, Claude Fable 5.1, averages 73.7% and about 40% on the hardest task set, with failures concentrated in the interaction layer and result verification rather than in scientific reasoning itself.
The work introduces OSWorld-Science, a benchmark and evaluation environment combining expert-defined scientific tasks, artifact-based execution graders, and a shared agent harness, and evaluates 12 vision-language models across 7 scientific domains, 17 software packages, and 3 languages, finding that the strongest model, Claude Fable 5.1, averages 73.7% and about 40% on the hardest task set, with failures concentrated in the interaction layer and result verification rather than in scientific reasoning itself.