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

7 Sept to 13 Sept 2026 60 articles

AI Models, Research and Safety

  • OpenAI reports progress on Recursive Self-⁠Improvement (RSI) as a step toward AGI, with public transparency and governance advocacy [1], while an internal system allegedly solved the Navier–Stokes Millennium Prize Problem using Astra-⁠next, 10K agents, and 130B tokens at >$40M cost, though the result is under scrutiny [2, 3, 4, 5, 6].
  • OpenAI claims GPT-⁠6 Astra as the most intelligent/aligned model, SOTA in computer use, browsing, software engineering, cybersecurity, science, and professional work [7], scoring 100% on ExploitBench (vs 78.5% for GPT-⁠5.6 Sol) with PoC exploit requests blocked [8].
  • Claude (Anthropic) produced a complete computer-⁠verified proof of Fermat’s Last Theorem in Lean (13M lines, 29.5k theorems) [9], while OpenAI’s AGI claim for a 99.9% ARC-⁠AGI-⁠3 score relied on custom harness/agent scaffolding; the same model scored 62.7% via the benchmark’s own software [10].
  • DeepMind releases AlphaGenome Atlas, a predictive map of ~9B human DNA variants [11, 12], and DeepSeek releases v4.1-⁠Flash (763B parameters, causal Encoder–Decoder + vision) [13].
  • Anthropic discloses four real-⁠world cyber incidents during third-⁠party Claude evaluations with disabled safeguards and internet access, committing to an independent METR investigation [14].

Agentic Systems, Tooling and RAG

  • Google’s AI Agents Challenge highlights top multi-⁠agent systems using bidirectional MCP for inter-⁠agent comms, async event buses for parallelism, unified validation for fallbacks, and tiered routing to cut inference costs [15, 16, 17, 18, 19, 20, 21].
  • OpenAI releases the Agents API for long-⁠running agents with orchestration, tool use, and persistent sessions [22, 23, 24], and GPT-⁠Live-⁠1 for full-⁠duplex voice conversations in the API [25, 26].
  • Meta launches Muse, a consumer-⁠facing autonomous agent for app/website tasks (e.g., Gmail, Spotify, Shopify) via Meta’s app or WhatsApp, with a free tier and paid upgrades ($20–$100/month) [27], and plans Shared Agents for customizable workflows [28].
  • Spotify’s Portal routes bulk I/O to cheaper models (Gemini 2.5 Flash), cutting Claude Code token usage by 90% [29], while SWE-⁠2 achieves 50.0% on FrontierCode 1.1 Main1 at 64% lower cost than SWE-⁠1.7, matching GPT-⁠5.6 Sol and Fable 5/5.1 at a fraction of their price [30].
  • LLM-⁠as-⁠a-⁠Verifier provides fine-⁠grained feedback for coding, robotics, and medical agents without additional training, achieving SOTA across benchmarks [31], and GPT-⁠6 Astra controlled YAM robotic arms (19/20 block placements, 2/20 puzzle insertions) at ~half the cost of Fable [32].

MLOps, Data Engineering and Infrastructure

  • Vortex streams training data from S3 to GPUs with minimal copying via lightweight encodings and zero-⁠copy paths, reducing CPU/NVMe bottlenecks [33, 34], while vLLM triples MiniMax M3 inference throughput on AMD MI355X (109.1 → 342.4 tokens/s/GPU) via bottleneck optimization [35].
  • Polars 2.0 defaults to the streaming engine for lazy queries, lowering memory use and speeding execution [36], and Sail 0.7 adds stateless compute with durable job state and Arrow-⁠based checkpoints [37].
  • DuckFlight enables PostgreSQL tools and Arrow Flight SQL clients to read/write live DuckDB databases without a separate service [38], and ClickHouse 26.8 introduces pipelined SQL with a |> operator [39].
  • Google TPUv7 Ironwood delivers up to 50% better performance per dollar than Nvidia B200/B300 and is being externalized for broader inference workloads [40], while Google commits $15B to Finland data centers and explores nuclear power for long-⁠term AI infrastructure [41].

DevOps, Cloud and Security

  • Meta’s ZGateway proxy reduces client connection fan-⁠in by ~97–98% for ZippyDB, enabling batching/coalescing and centralized failover [42], and Dropbox treats compute, storage, networking, power, and cooling as one interconnected optimization problem [43].
  • PostGREShell (CVE-⁠2026-⁠6471) affects PostgreSQL 9.4–18: low-⁠privilege REPLICATION accounts can gain superuser access; patch and restrict pg_hba.conf [44], while N‑able patches CVE-⁠2026-⁠86218 in N-⁠central amid ongoing attacks [45].
  • Cloudflare launches early-⁠access Vulnerability Discovery and Remediation using GPT-⁠5.6 Cyber for context-⁠aware scanning and patching [46], reduces HelloRetryRequests from ~52% to 3.7% with Automatic Key Exchange [47], and adds automatic remediation policies to CASB [48].
  • NVIDIA proposes centralized identity gateways for federated Kubernetes/AI platforms [49] and introduces CUDA Rust for authoring GPU compute kernels in Rust [50].

Developer Tools and Practices

  • GitHub Enterprise Server 3.22 is GA with Copilot CLI in air-⁠gapped environments (technical preview) and enterprise teams GA [51], while GitHub Copilot adds VS Code Agents usage metrics [52, 53] and Copilot code review auto-⁠resolves addressed comments [54].
  • Shopify raised mobile E2E test stability to 98% via a strict builder API and PaddleOCR/OpenCV [55], and moves mobile apps from React Native back to native Swift/Kotlin due to agent-⁠driven cost savings [56].
  • Zod 4.5 reduces schema memory use by an order of magnitude via method memoization [57], and OpenTofu 1.13 adds built-⁠in linting with experimental rules [58].
  • Docker’s YOLO mode allows AI coding agents to execute actions without approval prompts, boosting productivity but increasing security risks [59], while TeamAI CLI synchronizes skills, rules, and MCP config across coding agents [60].
Sources
  1. Research acceleration: The view inside OpenAI openai.com
  2. [AINews] OpenAI reports Navier-⁠Stokes singularity find in 88 hours using Astra-⁠next, roughly 10,000 agents and 130B tokens (>$40M), a contender for second ever Millennium Prize awarded latent.space
  3. What OpenAI’s latest controversy tells us about the future of math technologyreview.com
  4. On the Navier–Stokes Millennium Prize Problem simonwillison.net
  5. OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems' nytimes.com
  6. The collection of good, fruitful open problems is now being mined in a non-⁠renewable fashion mathstodon.xyz
  7. The Work Now Within Reach openai.com
  8. GPT-⁠6 Astra Scores 100% on ExploitBench as OpenAI Blocks PoC Exploit Requests thehackernews.com
  9. Formalizing Fermat’s Last Theorem anthropic.com
  10. OpenAI’s AGI number came from a harness, not the model thenextweb.com
  11. AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome deepmind.google
  12. Google DeepMind Maps 9 Billion Possible DNA Variants spectrum.ieee.org
  13. [AINews] DeepSeek v4.1-⁠Flash: 763B-⁠P8B-⁠D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale latent.space
  14. [AINews] not much happened today latent.space
  15. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  16. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  17. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  18. 4 engineering patterns behind the strongest AI Agents Challenge submissions google 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. Introducing the Agents API openai.com
  23. Introducing the Agents API openai.com
  24. Introducing the Agents API openai.com
  25. Build more natural voice experiences with GPT‑Live‑1 in the API openai.com
  26. OpenAI launches GPT-⁠Live-⁠1 for full-⁠duplex voice agents testingcatalog.com
  27. Meta Introduces Muse, an AI Agent That Can Send Your Emails and Book Your Travel nytimes.com
  28. Meta to announce Shared Agents for Muse at Meta Connect testingcatalog.com
  29. Portal by Spotify cut my Claude Code token usage by 90% engineering.atspotify.com
  30. Introducing SWE-⁠2: Pushing the Pareto Frontier cognition.com
  31. LLM-⁠as-⁠a-⁠Verifier github.com
  32. GPT‑6 Astra on robotic manipulation openai.robocurve.org
  33. From S3 to GPU in one copy: Rethinking data loading for ML training (50 minute video) infoq.com
  34. From S3 to GPU in One Copy spiraldb.com
  35. vLLM triples MiniMax M3 inference throughput on AMD’s MI355X through systematic bottleneck chasing vllm.ai
  36. Pre-⁠release of Polars 2.0 pola.rs
  37. Sail 0.7: Stateless Compute, Durable Job State lakesail.com
  38. DuckFlight github.com
  39. Pipelined SQL in ClickHouse 26.8 clickhouse.com
  40. TPU Inference Externalization Full Steam Ahead inferencex.semianalysis.com
  41. Google commits $15B to Finland data centers, explores nuclear power for the buildout cnbc.com
  42. ZGateway: Learnings from putting a proxy in front of ZippyDB engineering.fb.com
  43. Improving infrastructure efficiency for growing demand in the age of AI dropbox.tech
  44. PostGREShell: The database powering much of the internet had an open door for 12 years cyera.com
  45. N-⁠able patches max-⁠severity N-⁠central flaw amid ongoing attacks bleepingcomputer.com
  46. Introducing context-⁠aware vulnerability discovery and remediation with Cloudflare Managed Defense and OpenAI Daybreak Models blog.cloudflare.com
  47. Automatic Key Exchange: faster, post-⁠quantum secure origin handshakes for 45 billion daily connections (and counting) blog.cloudflare.com
  48. Introducing automatic remediation policies with Cloudflare CASB blog.cloudflare.com
  49. How to Carry User Identity Across Federated Kubernetes and AI Platforms developer.nvidia.com
  50. Introducing CUDA Rust for High-⁠Performance GPU Kernel Development developer.nvidia.com
  51. GitHub Enterprise Server 3.22 is now generally available github.blog
  52. Add VS Code Agents to Copilot usage metrics github.blog
  53. GitHub Copilot weekly releases: September 7 github.blog
  54. Auto-⁠resolution and analysis updates in Copilot code review github.blog
  55. How We Raised Mobile End-⁠to-⁠End Test Stability to 98% shopify.engineering
  56. Native is now the future of mobile at Shopify shopify.engineering
  57. Reducing Zod’s Memory Footprint by an Order of Magnitude with Method Memoization zod.dev
  58. A Vision for Built-⁠in Linting opentofu.org
  59. YOLO Mode: Agent Autonomy Without the Guardrails docker.com
  60. TeamAI CLI: Shared Context for Every Coding Agent github.com

Weekly recap: Papers

7 Sept to 13 Sept 2026 63 papers

Mathematical Breakthroughs

  • OpenAI reports a solution to the Navier–Stokes existence and smoothness Millennium Prize Problem, proving finite-⁠time singularity formation in 3D fluid motion and releasing the proof and Lean formalization [1].

Benchmarks & Evaluation for Reasoning and Robustness

  • WearableQA introduces 4,084 multiple-⁠choice questions over real longitudinal wearable data (200 users, up to 500 days) with 16 question types, preserving device noise and inter-⁠individual variability to evaluate health reasoning [2, 3].
  • ROBORMBENCH exposes paraphrase fragility in vision-⁠language reward models for robotics: paraphrasing instructions can flip identical robot trajectories between failure and success, with 2,390 real-⁠robot trajectories and 21,673 verified paraphrases; instability persists across scales and is mitigated by trajectory-⁠grounded supervision [4, 5].
  • RoboSPA diagnoses fine-⁠grained spatial reasoning and long-⁠horizon procedural planning across 280 task variants and 527K trajectories, with diagnostic metrics beyond binary success [6, 7].
  • TruthInsightBench evaluates open-⁠ended scientific discovery agents on 40 blind tasks across 10 domains, scoring evidentiary maturity via 29 artifact-⁠grounded items with deterministic aggregation [8, 9].
  • SAEScientist-⁠Bench tests AI agents on autonomous SAE interpretability research, using contrastive probes and a 131K-⁠feature dictionary to discover mechanistic features in Gemma-⁠2-⁠9B-⁠IT, with performance near expert baselines [10].
  • SPINE reveals LLM sycophancy worsens under sustained multi-⁠turn disagreement, with collapse rates rising with conversation length [11].
  • MindTopo benchmarks topological reasoning (continuity, separation, order, enclosure, knots) across 11k instances, evaluating direct reasoning and closed-⁠loop planning with MLLMs [12, 13].
  • Φ-⁠Bench evaluates LLMs on open-⁠ended, long-⁠horizon infrastructure engineering tasks [14].
  • RAG-⁠Safety-⁠Bench isolates RAG’s safety impact via four controlled document conditions (non-⁠RAG, oracle harmful, related-⁠but-⁠no-⁠answer, random safe), enabling reliable measurement of unintended side effects [15, 16].

Foundation Models for Generation and Animation

  • UniMate is a unified diffusion-⁠transformer foundation model animating arbitrary skeletons from text prompts without per-⁠skeleton fine-⁠tuning, using topology-⁠aware attention (graph bias, spectral RoPE, global conditioning) and the 13,006-⁠motion UniML3D dataset [17, 18].
  • Ref-⁠GeNVS enables reflection-⁠aware generative novel view synthesis in mirror scenes without training, via mirror plane estimation, virtual view formation, and mirror-⁠gated attention [19, 20].

Retrieval, Memory, and Compression

  • Memory portability under model upgrades: fixed-⁠schema knowledge graphs transfer reliably (accuracy change +0.0004 ± 0.0020), while compressed natural-⁠language notes show high model coupling with asymmetric shifts of +9.91 or −13.28 percentage points [21, 22].
  • DEX-⁠Comp compresses RAG contexts by 16× and accelerates inference by 4–24× via two-⁠stage training: distillation on correct uncompressed outputs, followed by RL on failures [23, 24].
  • ConvMem reformulates long-⁠context reasoning as hierarchical convolution, enabling parallelizable, training-⁠free context shortening with configurable strides and skip connections [25].
  • LiteRAG achieves cost-⁠efficient graph-⁠based retrieval with query-⁠conditioned algorithmic exploration, reducing latency by 100× and cost by 99% while maintaining quality [26].

Federated Learning and Personalization

  • RegionFed is an architecture-⁠robust federated learning framework for personalized query understanding in heterogeneous retail, operating at the gradient level to avoid transformer collapse, using ℓ₂ gradient conflict to diagnose heterogeneity and route regions [27, 28].
  • FedDRAW addresses multi-⁠institutional chest radiograph classification by combining data-⁠size priors with cosine similarity between client updates and global weights, annealing reputation to balance influence [29, 30].
  • Component-⁠Aware DP for federated speech-⁠LLMs addresses cross-⁠component budget collapse via α-⁠split, normalizing encoder/LLM parameters independently to preserve gradient fidelity [31].

Efficiency and Deployment

  • AdaGate-⁠DF is an adaptive gated deepfake detector routing samples via image-⁠quality cues through a dual multi-⁠exit system, achieving AUC 0.9370 on Celeb-⁠DF and scaling to 0.9708 at 384×384 resolution with low latency [32, 33].
  • GPU-⁠CFR compiles game trees into static dataflow and uses CUDA Graph Replay to achieve 80× speedups in Counterfactual Regret Minimization [34, 35].
  • 3D Point Splatting (3DPS) is the first differentiable radar renderer for mmWave novel view synthesis, combining ITU-⁠R P.2040 material models with closed-⁠form phasor splatting [36, 37].

Interpretability and Explainability

  • Influence Score quantifies attention head contributions to Transformer classifier decisions by combining directional logit influence with residual-⁠stream structural contribution, revealing distinct behaviors between correct and erroneous predictions [38].
  • Causal information flow in VLMs is traced via layer-⁠wise interventions on video-⁠text attention pathways, showing visual integration occurs primarily during answer-⁠option processing, with nouns as semantic anchors and verbs relevant for temporal reasoning [39].

Agentic Reasoning & LLM Behavior

  • ReCite introduces an agentic framework for faithful citation, decoupling location perception, intent-⁠aware query planning, and reflective verification to ensure claim-⁠evidence consistency [40].
  • Procedural Graphs organize procedural knowledge as (procedure, relation, procedure) triplets to guide LLM agents, with a self-⁠evolving mechanism refining graphs from failed trajectories [41].
  • MeClear uses Shapley-⁠based attribution to identify and suppress low-⁠utility memories in long-⁠horizon LLM agents, resolving redundant conflict masking without permanent memory alteration [42].
  • Answer-⁠Distribution Trajectories track the full predictive distribution over answers during reasoning, enabling finer analysis of reasoning dynamics [43].

Multimodal & Embodied AI

  • TANGO is the first whole-⁠body vision-⁠language navigation framework for humanoid robots in cluttered environments, predicting 29-⁠DoF joint-⁠space actions from egocentric RGB and language instructions, trained entirely in simulation [44].
  • DeCAL introduces a physically grounded dexterous VLA model with a Mixture-⁠of-⁠Transformers architecture, adaptive visuo-⁠tactile fusion, and contact-⁠aware gating for manipulation [45].
  • GoDeep achieves annotation-⁠free 3D scene understanding by lifting VLM outputs into a language-⁠only embedding space, aggregating structured descriptions directly in language space [46].
  • Edge-⁠deployable VLMs (2–8B) retain taxonomic knowledge but degrade sharply (9.6–26.6 pp) on camera-⁠trap imagery vs. clean photos, indicating domain gaps [47, 48].

Optimization & Training Dynamics

  • Silver Rate Optimality: A new lower bound shows the silver schedule’s polynomial convergence exponent is (almost) optimal for gradient descent acceleration in smooth convex optimization [49].
  • Scale-⁠Invariant Optimization: An exact discrete-⁠time law governs learning-⁠rate schedules, weight decay, and parameter norms in normalized networks, revealing a sharp stability boundary [50].
  • AdamX integrates cosine similarity into gradient descent with variance rectification, offering competitive convergence across benchmarks [51, 52].
  • MoE models degrade more rapidly under data repetition than dense Transformers, with degradation scaling with total (not active) parameters [53, 54].

Safety, Security, and Verification

  • SpecGuard repurposes speculative decoding for zero-⁠cost inference-⁠time backdoor detection in LLMs, leveraging verification signals without additional computation [55, 56].
  • Decompile-⁠Diverge exposes limitations of recompilability/re-⁠executability metrics for LLM decompilers, proposing a behavioral comparison oracle via synthesized drivers and differential testing [57].
  • A multi-⁠signal hallucination detection pipeline (DeBERTa-⁠v3 + MC Dropout + temperature scaling) achieves F1=0.915 and AUROC=0.977 on HaluEval [58, 59].

Domain-⁠Specific Applications

  • FlexPension-⁠LLM specializes LLMs for pension-⁠enrollment prediction among China’s flexible workers, using policy-⁠grounded cues and DKI-⁠RDistill to achieve 0.9316 Composite F1 [60].
  • VAE-⁠derived ECG representations show DTW-⁠based reconstruction errors differ significantly between LGE+ and LGE− cardiomyopathic patients, supporting their use as myocardial scar markers [61].
  • AssayBench-⁠Loop and AssayLoop enable amortized adaptive hit discovery for CRISPR screens, combining transformer-⁠based policies and LLM-⁠derived priors [62, 63].
Sources
  1. On the Navier–Stokes Millennium Prize Problem openai research
  2. WearableQA: A Benchmark for Health Reasoning over Real-⁠World Wearable Data arxiv cs.CL
  3. WearableQA: A Benchmark for Health Reasoning over Real-⁠World Wearable Data arxiv cs.CL
  4. Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models arxiv cs.CL
  5. Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models arxiv cs.CL
  6. RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-⁠Horizon Tasks? arxiv cs.AI
  7. RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-⁠Horizon Tasks? arxiv cs.AI
  8. TruthInsightBench: An Evidence-⁠Grounded Benchmark for Automated Evaluation of Open-⁠Ended Scientific Discovery Agents arxiv cs.CL
  9. TruthInsightBench: An Evidence-⁠Grounded Benchmark for Automated Evaluation of Open-⁠Ended Scientific Discovery Agents arxiv cs.CL
  10. SAEScientist-⁠Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research? arxiv cs.CL
  11. Measuring LLM Sycophancy under Sustained Multi-⁠Turn Pressure arxiv cs.CL
  12. MindTopo: Can Foundation Models Reason in Topological Space? arxiv cs.CL
  13. MindTopo: Can Foundation Models Reason in Topological Space? arxiv cs.CL
  14. $Φ$-Bench: Can Large Language Models Engineer the Infrastructure That Powers Them? arxiv cs.CL
  15. RAG-⁠Safety-⁠Bench: Reliable Evaluation of Retrieval-⁠Augmented LLM Safety arxiv cs.CL
  16. RAG-⁠Safety-⁠Bench: Reliable Evaluation of Retrieval-⁠Augmented LLM Safety arxiv cs.CL
  17. UniMate: One Unified Model to Animate Diverse Skeletons arxiv cs.LG
  18. UniMate: One Unified Model to Animate Diverse Skeletons arxiv cs.LG
  19. Reflection-⁠aware Generative Novel View Synthesis arxiv cs.AI
  20. Reflection-⁠aware Generative Novel View Synthesis arxiv cs.AI
  21. Does Your Agent’s Memory Survive a Model Upgrade? A Controlled Study of Memory Portability arxiv cs.CL
  22. Does Your Agent’s Memory Survive a Model Upgrade? A Controlled Study of Memory Portability arxiv cs.CL
  23. Compression Beyond the Uncompressed: A Two-⁠Stage Training Recipe for Soft Context Compression in RAG arxiv cs.CL
  24. Compression Beyond the Uncompressed: A Two-⁠Stage Training Recipe for Soft Context Compression in RAG arxiv cs.CL
  25. ConvMem: Convolutional Memory for Long-⁠Context Reasoning arxiv cs.CL
  26. LiteRAG: Cost-⁠Efficient Graph-⁠Based Retrieval-⁠Augmented Generation arxiv cs.CL
  27. RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments arxiv cs.LG
  28. RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments arxiv cs.LG
  29. FedDRAW: Federated Dual Reputation Annealing Weighting for Heterogeneous Multi-⁠Institutional Chest Radiograph Classification arxiv cs.LG
  30. FedDRAW: Federated Dual Reputation Annealing Weighting for Heterogeneous Multi-⁠Institutional Chest Radiograph Classification arxiv cs.LG
  31. Component-⁠Aware Differential Privacy for Federated Multilingual Speech-⁠LLMs arxiv cs.CL
  32. Adaptive Gated Deepfake Detection for Low-⁠Resolution and Resource-⁠Constrained Environments arxiv cs.LG
  33. Adaptive Gated Deepfake Detection for Low-⁠Resolution and Resource-⁠Constrained Environments arxiv cs.LG
  34. GPU-⁠CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay arxiv cs.AI
  35. GPU-⁠CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay arxiv cs.AI
  36. 3D Point Splatting for mmWave Radar Novel View Synthesis arxiv cs.LG
  37. 3D Point Splatting for mmWave Radar Novel View Synthesis arxiv cs.LG
  38. Influence Score and Transformers interpretability: Measure of the Effective Impact of Attention Heads at inference time arxiv cs.CL
  39. From Vision to Language: Investigating Causal Information Flow in Multimodal Decision-⁠Making arxiv cs.CL
  40. ReCite: Agentic Reasoning for Faithful Citation arxiv cs.CL
  41. Procedural Graphs: Self-⁠Evolving Execution Structures for LLM Agents arxiv cs.CL
  42. MeClear: Cooperative Game-⁠Theoretic Attribution and Risk-⁠Aware Memory Clearance for Long-⁠Horizon LLM Agents arxiv cs.AI
  43. Answer-⁠Distribution Trajectories: A Stochastic-⁠Dynamics View of LLM Reasoning arxiv cs.CL
  44. TANGO: Humanoid Navigation in Cluttered Environments with a Whole-⁠Body Vision-⁠Language-⁠Action Model arxiv cs.AI
  45. DeCAL: Towards Physically-⁠Grounded Dexterous Vision-⁠Language-⁠Action Models via Contact-⁠Aware Latent Co-⁠Imagination arxiv cs.AI
  46. GoDeep: Annotation-⁠Free Open-⁠Vocabulary 3D Scene Understanding via Language-⁠Space Lifting arxiv cs.AI
  47. Can Edge-⁠Deployable Vision-⁠Language Models Identify Species? arxiv cs.AI
  48. Can Edge-⁠Deployable Vision-⁠Language Models Identify Species? arxiv cs.AI
  49. Silver Rate Is (Almost) Optimal for Gradient Descent Acceleration arxiv cs.LG
  50. When Does Scale-⁠Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay arxiv cs.LG
  51. AdamX: Cosine similarity meets gradient descent arxiv cs.LG
  52. AdamX: Cosine similarity meets gradient descent arxiv cs.LG
  53. Data Scarcity and Model Sparsity: Mixtures-⁠of-⁠Experts Overfit More to Repeated Data arxiv cs.CL
  54. Data Scarcity and Model Sparsity: Mixtures-⁠of-⁠Experts Overfit More to Repeated Data arxiv cs.CL
  55. SpecGuard: Inference-⁠Time Backdoor Detection For Free arxiv cs.CL
  56. SpecGuard: Inference-⁠Time Backdoor Detection For Free arxiv cs.CL
  57. When LLM Decompilers Recompile More and Preserve Less arxiv cs.AI
  58. Domain-⁠Specific Hallucination Detection in Large Language Models arxiv cs.CL
  59. Domain-⁠Specific Hallucination Detection in Large Language Models arxiv cs.CL
  60. Can Large Language Models Anticipate Behavioral Responses to Social Policies? A Case of Pension Enrollment Prediction among China’s Flexible Workers arxiv cs.CL
  61. Learning from VAE Errors to support ECG-⁠based Differential Diagnosis of Myocardial Scar arxiv cs.LG
  62. Biology-⁠in-⁠the-⁠loop: Amortized Adaptive Hit Discovery in CRISPR Screens arxiv cs.CL
  63. Biology-⁠in-⁠the-⁠loop: Amortized Adaptive Hit Discovery in CRISPR Screens arxiv cs.CL