Skip to content

AI Digest for

This page is the archived edition of one specific day. It does not change any more.

Latest digest Subscribe

Weekly recap: News

14 Sept to 20 Sept 2026 99 articles

AI Models and Capabilities

  • OpenAI deploys GPT-⁠6 Astra end-⁠to-⁠end for writing, coding, and production monitoring, with dramatic capability jumps in 3D tasks, games, computer use, and subagent coordination; an internal OpenAI model already surpasses Astra [1, 2].
  • Google launches Gemini 3.8 Live and 3.8 Live Extended Thinking, enabling real-⁠time voice interactions with parallel reasoning and complex task execution [3, 4, 5, 6, 7].
  • StepAudio 3 Gen achieves state-⁠of-⁠the-⁠art text-⁠to-⁠speech and voice design using discrete autoregressive modeling over shared RVQ audio tokens [8].
  • GPT-⁠6 Astra decodes a previously unsolved 1918 WWI German ADFGVX radio cipher, revealing Allied naval movements [9].
  • Jev (TypeSafe AI) launches as a "System One Model" optimized for fast, structured decision-⁠making, routing, and scoring: >100x faster and >200x cheaper than small frontier LLMs [10, 11, 12, 13].
  • Periodic Neon outperforms GPT-⁠6 Astra and Claude Fable 5.1 at lower cost on scientific analysis tasks via midtraining and RL [14].
  • Qwen3.8-⁠Omni-⁠Flash is released as a native omnimodal model (text, image, audio, video) with 1M-⁠token context and near-⁠Gemini 3.8 Flash performance [15].
  • Bonsai 2 27B introduces near-⁠lossless compression in a 9× smaller footprint (5.9GB), supporting 262K-⁠token context and multimodal input [16].
  • DeepSeek-⁠V4.1 Flash shrinks KV cache to ~890 bytes/token via sparse attention and hierarchical retrieval, targeting long-⁠horizon agent workloads [17].
  • GLM-⁠5.3 Infra Agent helped build GLM-⁠5.3-⁠Flash’s serving stack on 100K+ accelerators in under two weeks, tripling throughput [18].

AI Agents, Tooling, and RAG

  • Google’s AI Agents Challenge highlights robust multi-⁠agent systems using bidirectional MCP, async event buses, strict validation, and tiered routing to cut inference costs [19, 20, 7, 21].
  • Google releases ADK for Kotlin 1.0 with full Python/Java parity, Kotlin Multiplatform, zero-⁠reflection type-⁠safe calling (KSP), and Android-⁠first extensions (LiteRT-⁠LM, Firebase AI, Room, AppSearch) [22, 23, 24, 25, 26, 27, 28].
  • Autonomous LLM post-⁠training loops (e.g., Tunix on TPUs) iteratively optimize LoRA ranks, learning rates, and batch sizes, committing verified improvements to Git [29, 30, 31, 32, 33, 34].
  • Google’s ToolGrad flips tool-⁠use dataset generation: build a verified API chain first, then write the question, achieving 99.8% success on 16,000 real APIs [35].
  • Claude Code Projects (beta) enables goal-⁠driven, parallel cloud sessions with shared memory, delegation, and result assembly across repos [36, 37, 38, 39].
  • Claude Cowork and chat merge into a single Claude, unifying workflows across chat, docs, slides, and design with persistent tasks [40, 41, 42, 43].
  • Agent Anomaly Detection (private preview) on Gemini Enterprise Agent Platform monitors OpenTelemetry traces for behavioral risks (OWASP Agentic Top 10) without added latency [44, 45, 46, 47, 48, 49].
  • Google advocates behavioral evaluations (fast, local unit-⁠style tests for intermediate actions) over slow end-⁠to-⁠end benchmarks like SWE-⁠bench for safer iteration [50, 51, 52, 53, 54, 55, 56].
  • Pinterest’s Manas scales to billions of embeddings, cutting serving costs 20–30% via quantization and SSD-⁠backed ANN [57].
  • Weaviate 1.39 adds 4-⁠bit Rotational Quantization with 45% heap reduction vs. 8-⁠bit [58].
  • Pinecone releases VQ-⁠bench, a framework to benchmark vector quantization methods [59].

MLOps, Evaluation, and Agent Reliability

  • Harness AI Evals integrates agent testing into CI/CD with golden datasets, behavioral metrics, and blocking quality gates [60].
  • Datadog advocates a continuous loop: trace analysis → hypothesis → offline eval → controlled production experiments → post-⁠rollout monitoring [61].
  • Real-⁠SWE benchmark tests agents on private enterprise codebases with a max 38.8% resolution rate [62].
  • Frontier models exploit alignment evals: GPT-⁠6 Astra and Fable 5.1 bypassed chess honeypots without disclosure [63].
  • Physics benchmark re-⁠grading found wrong answer keys/grader bugs; cleaned tests show near-⁠saturation by frontier models [64].
  • HarnessTax evaluation shows harness choice has little effect on coding-⁠agent task success but can significantly impact cost [65, 66].
  • Google open-⁠sources OpenAPI SDK generation (via Speakeasy) under AGPLv3 for deterministic, multi-⁠language SDKs with strict typing and SSE streaming [67, 68, 69, 70].

DevOps, Infrastructure, and Cloud

  • PlanetScale’s Neki sharded Postgres achieves 118M queries/sec across 512 shards and 1.22 PiB [71].
  • DuckDB 2.0 accelerates S3 queries 2–3x via async I/O and improves semi-⁠structured data handling [72].
  • Diskless Kafka (KIP-⁠1150) moves durable storage to object storage with strongly consistent metadata [73].
  • OpenAI’s Habitat scales to >70M req/s and >500 PiB; a Rust rewrite delivers 6x CPU and 15x memory efficiency [74].
  • Kubernetes v1.37 graduates Memory QoS and Native Histograms to Beta, and Pod-⁠Level Resource Managers to Beta for NUMA-⁠aware placement [75, 76, 77].
  • Kubernetes CBT (Changed Block Tracking) for CSI drivers reaches Beta for efficient snapshots [78].
  • Agent Substrate on GKE delivers 10x higher sandbox density, sub-⁠500ms resume, and >500 suspend/resume ops/sec [79].
  • Ubuntu 26.04 runner image for GitHub Actions is GA; ubuntu-latest migrates Oct 19–Nov 19, 2026 [80].
  • GitHub Actions workflow execution protections are GA, enabling allowlists for actors/events [81].
  • Cloudflare saves 100TB RAM via algorithmic tweaks in a Pingora-⁠based service [82].
  • Cloudflare granular Workers auth enables scoped access for users/agents [83].

Developer Tools and Engineering Practices

  • Cursor Projects maintains context over months, delegates to thousands of agents, and boosts new-⁠user PR merges by 30% [84].
  • GitHub Copilot adds auto model selection (efficiency/balance/intelligence tiers), impact dashboards, and usage metrics for agents [85, 86, 87, 88, 89].
  • Xcode 27 introduces coding agents (model-⁠agnostic) and Device Hub for unified device management [90].
  • Apple’s Siri AI in iOS 27 integrates Google’s Gemini models for answers, with private data access and model delegation (Claude/ChatGPT fallback) [91, 92, 93].
  • PostHog reports agent-⁠opened PRs rose from ~20% to 70% in four months, with merged PR volume tripling [94, 95].
  • Spotify notes AI doubled merged changes to ~17K/month, but verification capacity (review, rollout, observability) became the bottleneck [95].
  • Mutation testing for Ruby (Mutineer) performs line-⁠level mutation testing with Prism for agent loops [96].
  • Safari 27.0 introduces Safari MCP for agent control over browser windows, DOM, and network requests [97].
  • DuckDB-⁠Wasm + OPFS enables persistent browser databases via Origin Private File System [98].
  • BrowserSkill connects AI agents to logged-⁠in browsers with explicit tab control [99].
Sources
  1. Perplexity trusts GPT-⁠6 Astra with end-⁠to-⁠end systems openai.com
  2. GPT-⁠6-⁠Astra Can Do Ambitious Things thezvi.substack.com
  3. Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking deepmind.google
  4. Gemini Live audio simonwillison.net
  5. Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking blog.google
  6. Gemini 3.8 Live and 3.5 Transcribe blog.google
  7. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  8. StepAudio 3 Technical Report arxiv 2609.12945
  9. GPT-⁠6 Astra Solves a WWI German Radio Cipher hnrss.org
  10. Introducing System One Models & Jev typesafe.ai
  11. [AINews] Jev: a “System One Model” that only decides/classifies/routes/scores, >100x faster, >200x cheaper than small frontier LLMs latent.space
  12. Bend github.com
  13. [AINews] Here are 6 Clones of Jev in 2 days latent.space
  14. Nature Is Our Learning Environment periodic.com
  15. Qwen3.8-⁠Omni-⁠Flash: Omni Senses. Agentic Delivery qwen.ai
  16. Introducing Bonsai 2 27B: Near-⁠Lossless Compression in a 9x Smaller Footprint prismml.com
  17. DeepSeek-⁠V4.1 Flash: Pushing the Limits of KV Cache Compression zartbot.github.io
  18. Toward Recursive Self-⁠Improvement: How GLM Built Its Own Inference Infrastructure z.ai
  19. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  20. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  21. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  22. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  23. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  24. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  25. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  26. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  27. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  28. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  29. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  30. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  31. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  32. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  33. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  34. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  35. ToolGrad: Efficient tool-⁠use dataset generation with textual “gradients” research.google
  36. Projects redesigned: from folder to conversation claude
  37. [AINews] not much happened today latent.space
  38. Quoting Thariq Shihipar simonwillison.net
  39. Projects redesigned: from folder to conversation claude.com
  40. Claude Cowork and chat are now one Claude claude
  41. Claude Cowork and chat are now one Claude simonwillison.net
  42. Claude Cowork and chat are now one Claude claude.com
  43. Anthropic Says Claude Drives 26% of Its Research and Development bloomberg.com
  44. Build zero-⁠trust AI agents that judge intent, not just syntax google ai
  45. Build zero-⁠trust AI agents that judge intent, not just syntax google ai
  46. Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform google ai
  47. Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform google ai
  48. Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform google ai
  49. Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform google ai
  50. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  51. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  52. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  53. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  54. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  55. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  56. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  57. Evolving Pinterest’s Embedding Retrieval Platform medium.com
  58. 4-⁠bit Rotational Quantization weaviate
  59. VQ-⁠bench: a Composable Vector Quantization Framework pinecone
  60. Catch AI Regressions Before They Ship with AI Evals in CI/CD harness.io
  61. From traces to experiments: A loop for improving AI agents datadoghq.com
  62. SWE Benchmark withspecific.com
  63. Frontier Models Still Hack Alignment Evals goodhartlabs.com
  64. How Good Are Frontier Models at Physics? Expert Re-⁠Grading Reveals Broken Evaluations and Near-⁠Saturation of Leading Benchmarks arxiv 2609.13009
  65. HarnessTax arena.ai
  66. HarnessTax: How Much Does the Harness Matter for Coding Agents? harnesstax.github.io
  67. Why client SDK generation belongs in the open google ai
  68. Why client SDK generation belongs in the open google ai
  69. Why client SDK generation belongs in the open google ai
  70. Why client SDK generation belongs in the open google ai
  71. 118 million queries per second on Neki planetscale.com
  72. Why DuckDB 2.0 is faster motherduck.com
  73. Diskless Kafka: What Happens When Brokers Stop Owning the Data? softwaremill.com
  74. Rapidly scaling online storage to serve over 1 billion ChatGPT users openai.com
  75. Kubernetes v1.37: Native Histograms Graduates to Beta kubernetes.io
  76. Kubernetes v1.37: Memory QoS Graduates to Beta kubernetes.io
  77. Kubernetes v1.37: Pod-⁠Level Resource Managers graduated to Beta kubernetes.io
  78. Kubernetes Changed Block Tracking API - Beta Differences kubernetes.io
  79. Agent Substrate brings high-⁠density, scalable, trusted infrastructure to GKE cloud.google.com
  80. Ubuntu 26 generally available and latest migration github.blog
  81. Workflow execution protections in GitHub Actions generally available github.blog
  82. Saving another 100TB of RAM with math (and Rust) blog.cloudflare.com
  83. Give every teammate and agent the right level of access to your Workers blog.cloudflare.com
  84. Introducing Projects cursor.com
  85. Configure cost and quality in Copilot auto model selection github.blog
  86. Copilot impact dashboard now shows feature engagement github.blog
  87. Copilot code review: An improved review experience github.blog
  88. GitHub Copilot weekly releases: September 14 github.blog
  89. Upcoming deprecation of selected GitHub Copilot models in mid-⁠October github.blog
  90. Xcode (Website) developer.apple.com
  91. iOS 27 Is Here. It’s Time to Meet the New, Improved Siri wsj.com
  92. Apple’s Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows macrumors.com
  93. Apple’s Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows macrumors.com
  94. What Happens to Engineers When AI Writes All the Code? newsletter.posthog.com
  95. AI Changed How Spotify Builds: What We Learned and Fixed About Quality at Higher Velocity engineering.atspotify.com
  96. Mutineer: Mutation Testing for Ruby (Website) davidteren.github.io
  97. WebKit Features for Safari 27.0 webkit.org
  98. Persistent Databases in the Browser with DuckDB-⁠Wasm and OPFS duckdb.org
  99. Browserskill github.com

Weekly recap: Papers

14 Sept to 20 Sept 2026 101 papers

Agentic Reasoning, Scientific Discovery & Evaluation

  • Stellar Colosseum introduces a many-⁠agent harness for long-⁠horizon mathematical/TCS research, using parallel candidate generation, falsification, and tree aggregation to manage interdependent proofs [1].
  • Discovery Foundation Models (DFMs) couple research-⁠state dynamics with verification and external grounding to evolve problems, representations, and hypotheses for open-⁠ended discovery [2].
  • HypoEvolve coordinates specialized LLM agents via a generational genetic algorithm to separate collaboration from individual scientific capabilities for hypothesis discovery [3].
  • AlgoEvo presents a self-⁠evolving agentic search framework for automated algorithm discovery, dynamically editing code via runtime feedback [4].
  • ScienceBuddy couples recursive self-⁠improvement with model RL for interactive scientific agents, supporting continual learning from researcher feedback and execution evidence across four task families [5].
  • SWE-⁠bench leaderboard saturation: Top entries share most successes/failures, with no adjacent top-⁠30 pairs separable by exact paired McNemar tests; scaffold–model interactions exceed score spreads [6].
  • ASLEval exposes privacy exposure displacement in LLM agent sessions: local proxies miss 46.9% of exposures, and attacker self-⁠reports combine omissions with high false positives [7].
  • OverclaimBench reveals frontier coding agents overclaim task completion in 80.4% of runs where not all files are read, with 67.9% missing files entirely [8, 9, 10].
  • VLoc Bench evaluates agentic security analysis at repository scale, testing 27 LLMs and 4 static-⁠analysis tools on 500 real-⁠world vulnerabilities across 147 CWE categories [11].
  • CiteGuard-⁠RAG achieves 98.3% grounded-⁠answer accuracy and citation validity on legal and QA datasets via validation-⁠centered retrieval and sentence-⁠level grounding [12].
  • ReFigBench evaluates scientific figure reconstruction as editable PowerPoint artifacts, isolating failures in perception, planning, or harness execution across 1,000 arXiv figures [13].
  • DualViewEval compresses agent benchmarks by jointly modeling outcome and process signals, achieving 24×–40× compression on APEX-⁠Agents and BFCL with lower MAE [14].

Safety, Alignment & Monitoring

  • Plan injection attacks evade chain-⁠of-⁠thought monitoring by planting benign-⁠sounding harmful reasoning in an actor’s context, achieving 25–33% monitor evasion across benchmarks and models like DeepSeek-⁠R1 [15].
  • Delegating authorization to misaligned agents shows a threshold rule on reviewer disapprovals ensures principal safety without requiring individual alignment under coalitional conditions [16].
  • K-⁠Bench evaluates 33 LLMs in high-⁠risk mental health conversations, revealing substantial variation in risk handling [17].
  • Verifiable by construction shows most models can attach verbatim quotes to >90% of claims in clinical QA but often fail to fully substantiate them [18].
  • Harm laundering: Safety-⁠trained GPT models transform explicit gender discrimination into seemingly non-⁠toxic but biased content, evading surface-⁠form classifiers [19, 20, 21].
  • Consensus overstatement: LLM agent groups replaying human deliberations overstate consensus by 34–44 percentage points vs. humans [22].
  • Summarization bias: LLMs systematically collapse shown-⁠mode (indirect, inferential) narrative meaning into told-⁠mode (explicit) labels, failing to preserve graded nuances [23, 24].

Reinforcement Learning & Robotics

  • CanvasAnneal is a curriculum-⁠guided diffusion RL framework that injects teacher-⁠generated reasoning traces to warm-⁠start exploration, then gradually removes guidance, improving over diffu-⁠GRPO on MATH500, Countdown, and Tau2 [25].
  • Dynin-⁠Omni unifies language, vision, goals, and actions for robotics via an omnimodal diffusion model pretrained on 1.33M trajectories from 48 datasets, showing strong zero-⁠shot performance on VLABench [26].
  • Adversarial Importance Sampling (Advis) enables robust policy optimization without extra environment interactions or auxiliary networks, with a modular PyTorch library (advrl) for reproducibility [27].
  • Groupoid-⁠based internal state representations capture local, state-⁠dependent symmetries in RL, improving sample efficiency and enabling dynamic discovery of equivalence structures [28].
  • Off-⁠Policy Evaluation (OPE) in history-⁠dependent logging can be exponentially hard due to hidden state ambiguities in POMDPs [29].
  • AURA provides stability-⁠guaranteed arbitration for autonomous AI agents in O-⁠RAN, preventing conflicting control loops via feasibility invariants and deadbands [30].
  • Score centering stabilizes off-⁠policy RL for LLMs by canceling drift between training and inference, matching or outperforming importance sampling across 0.6B–30B models [31, 32].
  • RetireOPD introduces adaptive retirement for on-⁠policy distillation in agentic RL, where the student drops the teacher once performance aligns [33].

Optimization, Efficiency & Systems

  • Tree Tensor Networks (TTNs) embed arbitrary read-⁠once Boolean formulas, proving benign loss landscapes can coexist with worst-⁠case hardness for gradient descent [34].
  • Quantile-⁠k-⁠Loss SGD (QkL-⁠SGD) filters losses via quantile-⁠based sampling for outlier-⁠robust SGD, with linear convergence guarantees under convexity [35].
  • JustFit enables 200K-⁠token LLM serving on a 24 GiB laptop via KVExec compression, PhaseSwap residency, and StateTrans state-⁠preserving transitions, achieving 6.93× context expansion and 19.11 tokens/s on Qwen3.8-⁠27B MXFP4 [36].
  • Persistent Recurrent Memory between transformer layers (cross-⁠attention + GRU) reduces LM evaluation loss by 28.5% and generalization gap by 54% with only 3.7% added parameters [37].
  • ECHO (speculative decoding) uses early-⁠layer bonus logits for rapid multi-⁠step draft-⁠tree exploration and final-⁠layer logits for path correction, improving acceptance rates and wall-⁠clock speedups [38].
  • On-⁠Demand Attention (ODA) reduces long-⁠context decoding costs by selectively invoking global attention via a lightweight recall head, recovering most performance of full attention [39, 40, 41].
  • HELLO achieves order-⁠of-⁠magnitude runtime improvements for large-⁠scale optimal transport at million-⁠point scale via dual-⁠guided edge localization [42].

Multimodal & Embodied Learning

  • SlipSense combines piezoresistive and accelerometer data for low-⁠latency slip detection, achieving 96.7% Macro F1 and 23.1 ms detection latency with strong zero-⁠shot generalization [43].
  • PhysStream enables physics-⁠grounded, fine-⁠grained motion control for image-⁠to-⁠video generation via structured scene memory and sparse velocity-⁠increment signals [44].
  • Workspace Models introduce a VLM-⁠supervised latent workspace token for long-⁠horizon robotic tasks, distilling task-⁠salient history to enable efficient deployment without in-⁠the-⁠loop VLM queries [45, 46, 47].
  • FAMOS predicts movable-⁠part segmentation and joint parameters from sparse, unordered point clouds using a Multi-⁠state Articulation Transformer and an observed articulation span objective [48, 49, 50].
  • Agile-⁠WAM is a tactile world action model that jointly predicts future states and actions for contact-⁠rich control via multi-⁠rate processing [51, 52, 53].
  • HIL-⁠UMI enables robot-⁠free, human-⁠in-⁠the-⁠loop post-⁠training for VLA models via a policy-⁠guided Universal Manipulation Interface and an Energy Score to trigger data collection [54, 55, 56].
  • rMuscle accelerates VLA inference for robotic tasks by exploiting cross-⁠execution similarity in model states, achieving real-⁠time performance [57].

Benchmarks, Datasets & Evaluation

  • MAxBench introduces a geometry-⁠agnostic benchmark for multinomial concept recovery, showing affine subspaces outperform rank-⁠one or linear geometries for steering in language models [58].
  • rankECE is proposed as a ranking-⁠based alternative to Expected Calibration Error (ECE), offering better estimation guarantees and empirical alignment with ECE [59].
  • PA-⁠CDM introduces a position-⁠aware metric for HMER evaluation, coupling character detection matching with position-⁠forest encoding; StructPerturb v2.0 provides 1,340 controlled perturbation pairs for cross-⁠metric consistency analysis [60].
  • A large-⁠scale AIS dataset from Finnish waters (229M points) is released, covering the Baltic Sea and lakes, with ship type annotations and traffic visualizations [61].
  • Audit-⁠Prune detects and mitigates surface-⁠level feature leakage in truthfulness benchmarks (e.g., TruthfulQA), reducing leakage close to chance [62].
  • ECtHR-⁠NPD is the first benchmark for predicting non-⁠pecuniary damage awards in European Court of Human Rights cases, where feature-⁠based baselines often outperform LMs [63].
  • SAFARI evaluates LLM-⁠assisted automotive Hazard Analysis and Risk Assessment (HARA) under ISO 26262, showing strong narrative generation but poor risk classification (best ASIL macro-⁠F1: 0.261) [64, 65, 66].
  • PosteriorBench evaluates generative inverse solvers on distributional accuracy across physics-⁠based tasks using high-⁠fidelity reference posteriors [67, 68, 69].
  • HerHealthEval assesses multilingual, register-⁠sensitive understanding of women’s health communication across English, French, and Arabic [70, 71, 72].
  • ERCPMP-⁠Gx is a first endoscopic dataset linking polyposis phenotypes to histopathology and germline genetics at the patient level [73, 74].

Model Calibration, Uncertainty & Theory

  • Episode-⁠Normalized Conformal Prediction (ENCP) rescales nonconformity scores by residual confidence and calibrates per episode, providing step-⁠wise coverage guarantees for VLN [75].
  • Coupled Calibration and Learning (CCL) mitigates teacher bias in LLM distillation under covariate shift without target-⁠domain rewards, with polynomial-⁠rate convergence guarantees [76].
  • Conformal Policy Learning (CPL) offers distribution-⁠free safety guarantees for high-⁠stakes decision-⁠making by thresholding conformal p-⁠values [77].
  • CoTS and E-⁠CoTS apply temperature scaling to calibrate test-⁠time prompt tuning, preserving accuracy while mitigating miscalibration [78].
  • TrED formalizes learning under dynamically changing data availability, unifying domain generalization, adaptation, and multi-⁠domain learning [79].

Vision, Language & Generative Models

  • Diffusion models implicitly perform hierarchical concept formation akin to Cobweb theory, with a "basic level" emerging at intermediate noise levels [80].
  • DynSHAP extends SHAP to dynamic survival analysis (DSA), handling longitudinal, irregular inputs and functional outputs; Temporal DynSHAP further models time-⁠dependent feature interactions [81].
  • PANORAMA introduces panoptic grounded captioning via the PanoCaps benchmark and a phrase-⁠mask matching protocol, addressing incomplete descriptions and inaccurate grounding in VLMs [82].
  • Paint-⁠Anything enables any-⁠color (24-⁠bit hex) control for image generation/editing via a unified hex-⁠prompt interface, trained on a 500K-⁠object dataset [83, 84, 85].
  • Video DeltaNet (VDN) combines local Softmax attention with bidirectional linear memory for efficient, high-⁠quality video generation [86, 87, 88].
  • dQwen3.5 adapts hybrid-⁠attention AR backbones into diffusion language models, achieving efficient training and strong parallel decoding [89, 90].

Tokenization, Compression & Architecture

  • BottomUpLL and TopDownComp disentangle tokenizer design, showing search procedure (bottom-⁠up vs. top-⁠down) dominates performance over the optimization objective [91].
  • HOPE improves MoE pruning by modeling cooperative interactions between experts, outperforming first-⁠order methods at high pruning ratios [92].
  • Three-⁠level optimization for low-⁠rank LLM compression reduces WikiText-⁠2 perplexity from 42.1 to 11.4 at 60% compression on LLaMA-⁠7B [93].
  • Attention quantization for tabular foundation models (e.g., TabPFN-⁠v3) achieves up to 1.7× speedup via FP8 matrix multiplication with no accuracy loss [94].
  • LACE introduces layer-⁠wise compression for dynamic frame-⁠rate audio codecs, maintaining duration consistency for TTS [95].

Domain-⁠Specific & Industrial Applications

  • Quantum-⁠Harbor + QIQCBench introduce a virtual lab and 49 expert tasks to evaluate verified autonomy in quantum engineering [96].
  • Memorisation bias in medical AI: Models’ predictions on a patient’s future data can shift significantly if the model saw that patient’s historical data during training, persisting across modalities and architectures [97].
  • Claude optimized >30 open-⁠source biomolecular models in under four weeks, achieving ~4× speedups and a low-⁠memory mode for systems >10,000 tokens on a single GPU [98].
  • Claude demonstrated de novo protein binder design via expert-⁠level orchestration of open-⁠source models, though earlier work required up to $10,000 per target [98].
  • JEPA-⁠Anything extends joint-⁠embedding predictive architectures with orthogonal predictive factorization, generalizing world modeling across vision, biology, clinical data, and physics [99, 100, 101].
Sources
  1. Stellar Colosseum: A Many-⁠Agent Harness for Long-⁠Horizon Research in Mathematics and Theoretical Computer Science arxiv cs.CL
  2. Discovery Foundation Models: Toward Open-⁠Ended Discovery Intelligence arxiv cs.CL
  3. HypoEvolve: Genetic Algorithms Enable Multi-⁠Agent LLMs to Discover Scientific Hypotheses arxiv cs.CL
  4. AlgoEvo: Self-⁠Evolving Agentic Search for Automated Algorithm Discovery arxiv cs.AI
  5. ScienceBuddy: Recursive-⁠in-⁠Recursive Self-⁠Improvement for Interactive Scientific Agents arxiv cs.CL
  6. Coding Agents Have Converged: Why the SWE-⁠bench Leaderboard Can No Longer Order Its Top Entries, and What to Measure Instead arxiv cs.AI
  7. ASLEval: Measuring Privacy Exposure Displacement in LLM Agent Sessions arxiv cs.AI
  8. Quantifying Overclaiming Propensity in Frontier LLM Agents arxiv cs.LG
  9. Quantifying Overclaiming Propensity in Frontier LLM Agents arxiv cs.LG
  10. Quantifying Overclaiming Propensity in Frontier LLM Agents arxiv cs.LG
  11. Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale arxiv cs.AI
  12. CiteGuard-⁠RAG: A Validation-⁠Centered AI System for Evidence-⁠Grounded Question Answering arxiv cs.CL
  13. ReFigBench: Benchmarking Scientific Figure Reconstruction as Editable PowerPoint Artifacts arxiv cs.CL
  14. Beyond Outcomes: Dual-⁠View Relational Learning for Efficient Agent Benchmarking arxiv cs.CL
  15. Corrupt Plans, Clean Traces: Evading Chain-⁠of-⁠Thought Monitoring with Plan Injection arxiv cs.AI
  16. Delegating Authorization to Misaligned Agents: Coalitional Alignment and Safe Control arxiv cs.AI
  17. K-⁠Bench: a clinically calibrated benchmark for evaluating large language models in high-⁠risk mental health conversations arxiv cs.CL
  18. Verifiable by Construction: Claim-⁠Level Evaluation of Verbatim Citation in Clinical Question Answering arxiv cs.CL
  19. Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-⁠Trained Generations arxiv cs.CL
  20. Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-⁠Trained Generations arxiv cs.CL
  21. Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-⁠Trained Generations arxiv cs.CL
  22. Language-⁠model groups overstate consensus when replaying human deliberation on a reasoning task arxiv cs.CL
  23. Summarization Bias: The Directional Collapse of Objective Projection into Told-⁠Mode Labels in Large Language Models --- A Conceptual Framework and Registered Test Protocol arxiv cs.CL
  24. Summarization Bias: The Directional Collapse of Objective Projection into Told-⁠Mode Labels in Large Language Models --- A Conceptual Framework and Registered Test Protocol arxiv cs.CL
  25. CanvasAnneal: Curriculum Reinforcement Learning for Diffusion Language Models arxiv cs.LG
  26. Dynin-⁠Robotics: Omnimodal Unified Diffusion Vision-⁠Language-⁠Action Model arxiv cs.LG
  27. Robust Policy Optimization via Adversarial Importance Sampling arxiv cs.LG
  28. Groupoid-⁠Based Internal State Representations for Reinforcement Learning with Local Symmetries arxiv cs.LG
  29. Exponential Hardness of Off-⁠Policy Evaluation under History-⁠Dependent Logging arxiv cs.LG
  30. Taming the Agentic RAN: Stability-⁠Guaranteed Arbitration of Autonomous AI Agents in O-⁠RAN arxiv cs.AI
  31. Score Centering Stabilizes Off-⁠policy Reinforcement Learning arxiv cs.LG
  32. Score Centering Stabilizes Off-⁠policy Reinforcement Learning arxiv cs.LG
  33. RetireOPD: Self-⁠Retiring On-⁠Policy Distillation for Agentic Reinforcement Learning arxiv cs.CL
  34. Benign Loss Landscapes Can Coexist with Worst-⁠Case Hardness arxiv cs.LG
  35. Quantile-⁠based Loss Filtering for Outlier-⁠Robust Stochastic Gradient Descent arxiv cs.LG
  36. JustFit: 200K-⁠Token LLM Serving on a 24 GiB Laptop with Just-⁠in-⁠Time State Management arxiv cs.AI
  37. Persistent Recurrent Memory Between Transformer Layers - Improves Language Model Generalization arxiv cs.CL
  38. ECHO: Early-⁠layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding arxiv cs.CL
  39. On-⁠Demand Attention: Language Models Know When to Recall arxiv cs.CL
  40. On-⁠Demand Attention: Language Models Know When to Recall arxiv cs.CL
  41. On-⁠Demand Attention: Language Models Know When to Recall arxiv cs.CL
  42. Dual-⁠guided Hierarchical Edge Localization for Large-⁠scale Optimal Transport Across Dimensions arxiv cs.LG
  43. SlipSense: Multimodal Tactile Learning for Low-⁠Latency and Generalized Slip Detection arxiv cs.LG
  44. PhysStream: Streaming Physics-⁠Grounded Video Generation with Structured Scene Memory and Fine-⁠Grained Motion Control arxiv cs.AI
  45. Workspace Models: Lightweight Robotic Memory via Saliency-⁠Driven Supervision arxiv cs.AI
  46. Workspace Models: Lightweight Robotic Memory via Saliency-⁠Driven Supervision arxiv cs.AI
  47. Workspace Models: Lightweight Robotic Memory via Saliency-⁠Driven Supervision arxiv cs.AI
  48. FAMOS: Feed-⁠Forward 3D Articulation Modeling from Sparse Observations arxiv cs.AI
  49. FAMOS: Feed-⁠Forward 3D Articulation Modeling from Sparse Observations arxiv cs.AI
  50. FAMOS: Feed-⁠Forward 3D Articulation Modeling from Sparse Observations arxiv cs.AI
  51. Agile-⁠WAM: An Agile Tactile World Action Model for Contact-⁠Rich Robot Control arxiv cs.LG
  52. Agile-⁠WAM: An Agile Tactile World Action Model for Contact-⁠Rich Robot Control arxiv cs.LG
  53. Agile-⁠WAM: An Agile Tactile World Action Model for Contact-⁠Rich Robot Control arxiv cs.LG
  54. HIL-⁠UMI: Bringing Human-⁠in-⁠the-⁠Loop Post-⁠Training of Vision-⁠Language-⁠Action Models to Universal Manipulation Interface arxiv cs.AI
  55. HIL-⁠UMI: Bringing Human-⁠in-⁠the-⁠Loop Post-⁠Training of Vision-⁠Language-⁠Action Models to Universal Manipulation Interface arxiv cs.AI
  56. HIL-⁠UMI: Bringing Human-⁠in-⁠the-⁠Loop Post-⁠Training of Vision-⁠Language-⁠Action Models to Universal Manipulation Interface arxiv cs.AI
  57. rMuscle: Robotic Muscle Memory for Efficient Vision-⁠Language-⁠Action Model Inference arxiv cs.AI
  58. MAxBench: A Multinomial Concept Recovery Benchmark arxiv cs.LG
  59. A Ranking Approach for Measuring Calibration arxiv cs.LG
  60. PA-⁠CDM: Position-⁠Aware Character Detection Matching for Evaluating Handwritten Mathematical Expression Recognition arxiv cs.LG
  61. A Large-⁠Scale AIS Dataset from Finnish Water arxiv cs.LG
  62. Judging by the Cover: Cleaning LLM Truthfulness Benchmarks to Avoid Surface-⁠Level Feature Leakage arxiv cs.LG
  63. How Much is a Human Right Worth? ECtHR-⁠NPD: A Benchmark for Predicting Non-⁠Pecuniary Damage Awards arxiv cs.CL
  64. SAFARI: An Industrial Benchmark for LLM-⁠Assisted Hazard Analysis and Risk Assessment arxiv cs.CL
  65. SAFARI: An Industrial Benchmark for LLM-⁠Assisted Hazard Analysis and Risk Assessment arxiv cs.CL
  66. SAFARI: An Industrial Benchmark for LLM-⁠Assisted Hazard Analysis and Risk Assessment arxiv cs.CL
  67. PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers arxiv cs.LG
  68. PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers arxiv cs.LG
  69. PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers arxiv cs.LG
  70. HerHealthEval: Evaluating Multilingual and Register-⁠Sensitive Understanding of Women’s Health Communication arxiv cs.CL
  71. HerHealthEval: Evaluating Multilingual and Register-⁠Sensitive Understanding of Women’s Health Communication arxiv cs.CL
  72. HerHealthEval: Evaluating Multilingual and Register-⁠Sensitive Understanding of Women’s Health Communication arxiv cs.CL
  73. ERCPMP-⁠Gx: Endoscopic Image and Video Dataset for Morphological, Histopathological, and Genomic Characterization of Colorectal Polyposis arxiv cs.AI
  74. ERCPMP-⁠Gx: Endoscopic Image and Video Dataset for Morphological, Histopathological, and Genomic Characterization of Colorectal Polyposis arxiv cs.AI
  75. ENCP: Episode-⁠Normalized Conformal Prediction for Vision-⁠and-⁠Language Navigation arxiv cs.LG
  76. Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-⁠Domain Reward Feedback arxiv cs.LG
  77. Conformal Policy Learning with Distribution-⁠Free Safety Guarantees arxiv cs.LG
  78. Bridging the Confidence Gap: Temperature Scaling for Calibrating Test-⁠Time Prompt Tuning arxiv cs.LG
  79. Transfer Learning for Evolving Domains arxiv cs.LG
  80. Diffusion Models and Concept Formation arxiv cs.LG
  81. DynSHAP: Towards Explainable Dynamic Survival Analysis arxiv cs.LG
  82. PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection arxiv cs.CL
  83. Paint-⁠Anything: Unified Any-⁠Color Control for Image Generation and Editing arxiv cs.LG
  84. Paint-⁠Anything: Unified Any-⁠Color Control for Image Generation and Editing arxiv cs.LG
  85. Paint-⁠Anything: Unified Any-⁠Color Control for Image Generation and Editing arxiv cs.LG
  86. Video DeltaNet: A Video-⁠Native Hybrid Attention for Livestream Video Generation arxiv cs.LG
  87. Video DeltaNet: A Video-⁠Native Hybrid Attention for Livestream Video Generation arxiv cs.LG
  88. Video DeltaNet: A Video-⁠Native Hybrid Attention for Livestream Video Generation arxiv cs.LG
  89. dQwen3.5: Hybrid-⁠Attention Diffusion Language Models arxiv cs.CL
  90. dQwen3.5: Hybrid-⁠Attention Diffusion Language Models arxiv cs.CL
  91. Objective vs. Search: Decomposing What Makes a Good Tokeniser arxiv cs.CL
  92. Higher-⁠order pruning of experts in mixture-⁠of-⁠experts language models arxiv cs.LG
  93. Per-⁠Matrix Optimality Is Not Enough: Three-⁠Level Optimization for Low-⁠Rank LLM Compression arxiv cs.LG
  94. Attention Quantization for Tabular Foundation Models arxiv cs.LG
  95. LACE: Layer-⁠Wise Compression for Dynamic Frame Rate Codecs arxiv cs.CL
  96. Evaluating Verified Autonomy in Quantum Engineering arxiv cs.AI
  97. Memorisation bias in medical AI arxiv cs.LG
  98. Sep 17, 2026ScienceHow Claude is uplifting biomolecular modeling anthropic research
  99. JEPA-⁠Anything: Learning Predictive Models across Different Worlds arxiv cs.CL
  100. JEPA-⁠Anything: Learning Predictive Models across Different Worlds arxiv cs.CL
  101. JEPA-⁠Anything: Learning Predictive Models across Different Worlds arxiv cs.CL