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

News from Saturday, 29 August 2026

2 articles

AI/ML Ecosystem & Tooling

  • OpenAI terminates Cursor’s access to its models, citing a history of contract violations by Elon Musk’s companies; the move follows SpaceX’s acquisition of Cursor and escalates tensions between the organizations [1].
  • GPT-⁠5.6 is now a competitive coding alternative to Claude 5, shifting the landscape for AI-⁠powered coding assistants [1].

Security & Developer Tooling

  • Security exploits are increasingly being discovered within minutes of bug discussions, as automated agents (e.g., DeepSeek V4 Pro) rapidly identify and probe vulnerabilities in public repositories [2].
Sources
  1. [AINews] OpenAI shuts off Cursor latent.space
  2. Just a rumour of a bug is enough to find a security exploit these days simonwillison.net

Papers from Saturday, 29 August 2026

20 papers

LLM Reasoning & Post-⁠Training

  • CritICL leverages structured failure modes from weaker models as critique-⁠based in-⁠context guidance to improve inference-⁠time reasoning efficiency, offering dynamic and static variants with consistent gains over baselines [1].
  • TTPO introduces an asymmetric test-⁠time policy optimization objective that distills agreeing rollouts and penalizes disagreeing ones, enabling robust updates without ground-⁠truth labels [2].
  • SWE-⁠Prime filters SFT data at trajectory and segment levels to remove noisy supervision (ineffective/redundant steps), improving code-⁠resolution performance [3].
  • Evolution Strategies (ES) for LLM reasoning show broader reasoning coverage than GRPO, with theoretical and empirical evidence; a sequential GRPO-⁠ES strategy combines their strengths [4].
  • Boosting LLM Exploration in RLVR uses weak-⁠model-⁠generated partial reasoning trajectories to disrupt overconfidence and improve pass@k via distributional diversity [5].

Agentic Systems & Safety

  • RedEvoAgent is a black-⁠box red-⁠teaming agent that distills cross-⁠case attack trajectories into human-⁠readable skills, evolving via tool-⁠effectiveness profiling and validation ratchets [6].
  • WikiSkill co-⁠evolves agent skills with a persistent knowledge base (wiki) to systematically consolidate experience, outperforming prior skill-⁠evolution methods [7].
  • Persona-⁠Execution Separation (PES) decouples persona (drift-⁠prone) and execution (audited) into separate trust domains, enabling free persona evolution while maintaining traceable execution [8].
  • INTENT-⁠AS-⁠A-⁠TOOL adds intent-⁠targeted tools to monitor agentic misalignment, providing fine-⁠grained signals of harmful intent during generation [9].
  • Beyond F1 for AI Model Security Scanners evaluates ModelScan, ModelAudit, and Fickling on a synthetic benchmark, revealing gaps in coverage and failure recovery beyond precision/recall [10].

Benchmarks & Evaluation

  • MCR-⁠Bench introduces a defect-⁠state-⁠aware benchmark for realistic multi-⁠round code review, capturing iterative interactions and dynamic state annotations across five languages [11].
  • CorporateBench is a large-⁠scale, human-⁠validated Q&A benchmark with temporally evolving knowledge bases, exposing LLM performance degradation at realistic document scales [12].
  • RATIO benchmarks retrieval across typed ideation operations (Address, Broaden, Specify) in scientific literature, with operation-⁠specific fine-⁠tuning improving performance [13].
  • PAWBench evaluates probabilistic alignment in world models by testing whether repeated video rollouts recover the correct distribution of physical behaviors [14].

Efficiency & Scalability

  • LeVJEPA trains a collapse-⁠free video encoder with a single hyperparameter and uniform token dropping, reducing pretraining cost while maintaining performance [15].
  • Successive Capacity Growth (SCG) dynamically expands Vision Transformer encoders in width/depth for JEPA world models, driven by task-⁠agnostic validation and SIGReg to prevent collapse [16].
  • Puro-⁠2B demonstrates cost-⁠efficient pretraining of a Qwen2-⁠1.5B-⁠scale model on consumer RTX 5090 GPUs for under \$6.9K, using FP8 and recipe optimizations [17].

Domain-⁠Specific Applications

  • CLAP enables cross-⁠embodiment video world models by reconciling disparate action spaces (end-⁠effector poses, language, latent actions) and shows zero-⁠shot physical simulation capabilities [18].
  • MM-⁠Spectrum introduces a sparse MoE framework for multimodal multispectral molecular structural elucidation, addressing modality imbalance with explicit routing [19].
  • CAST uses sparse autoencoders to expose auditable clinical concepts in language models, suppressing artifacts and improving robustness in mortality prediction [20].
Sources
  1. CritICL: Inference-⁠Time Weak-⁠to-⁠Strong Generalization from Small Language Model Failure Modes arxiv cs.CL
  2. TTPO: Test-⁠Time Policy Optimization arxiv cs.CL
  3. SWE-⁠Prime: Fewer Trajectories, Better Performance arxiv cs.CL
  4. Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO arxiv cs.LG
  5. Boosting LLM Exploration via Weak-⁠Model Guidance in RLVR arxiv cs.CL
  6. RedEvoAgent: Automatic Red-⁠Teaming Agent with Experience-⁠Driven Skill Evolution arxiv cs.AI
  7. WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution arxiv cs.CL
  8. Persona-⁠Execution Separation: An Architecture Pattern for Evolving LLM Agents under Execution Audit arxiv cs.AI
  9. INTENT-⁠AS-⁠A-⁠TOOL Makes it Easy to Track Agentic Misalignment arxiv cs.CL
  10. Beyond F1: Evaluating Coverage and Failure Recovery in AI Model Security Scanners arxiv cs.AI
  11. From Static to Dynamic: Benchmarking Real-⁠World Code Review with MCR-⁠Bench arxiv cs.CL
  12. CorporateBench: Large-⁠Scale Q&A Benchmarking with Temporal Knowledge Bases arxiv cs.CL
  13. RATIO: A Benchmark for Retrieval Across Typed Ideation Operations in Scientific Literature arxiv cs.CL
  14. PAWBench: How Far Are We from Probabilistically Aligned World Modeling? arxiv cs.AI
  15. LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics arxiv cs.AI
  16. Successive Capacity Growth: Task-⁠Complexity-⁠Driven Width and Depth Expansion for Vision Transformer Encoders in JEPA World Models arxiv cs.AI
  17. Puro-⁠2B: Poor Lab’s Qwen2-⁠1.5B Trained on RTX 5090 within $5090 arxiv cs.CL
  18. CLAP: Cross-⁠Embodiment Video World Models are Zero-⁠Shot Physical Simulators arxiv cs.AI
  19. MM-⁠Spectrum: Multimodal Multi-⁠spectral Molecular Structural Elucidation with a Stable MoE Framework arxiv cs.LG
  20. Making Clinical Language Models Auditable: Concept-⁠Guided Fine-⁠Tuning for Robust Prediction arxiv cs.CL