Daily AI news, sourcedUpdates / Research
Updates Research · 222 items · newest issue first · page 4 of 12
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- Loose Lips benchmark studies how LLM agents leak private information in agent-to-agent dialogueUnconfirmedarXiv cs.CLOct 14Research
- PsyBer-Agent simulates recommender-system users with fatigue, emotion and shifting preferencesUnconfirmedarXiv cs.CLOct 14Research
- Long-Horizon-Terminal-Bench 2.0 expands to 77 terminal tasks across 13 domains to measure sustained agent progressUnconfirmedarXiv cs.AIOct 14Research
- Correction Training targets exposure bias in diffusion modelsUnconfirmedarXiv cs.AIOct 14Research
- ClawLoop trains harness-based agents through a self-improving loop inside their own sandbox environmentUnconfirmedarXiv cs.AIOct 14Research
- The Silent Failures in Agentic Data Science looks past accuracy and code correctness when judging data science agentsUnconfirmedarXiv cs.AIOct 14Research
- Artificiality's First Exam proposes an LLM-referenced C-test for evaluating AI beyond human levelUnconfirmedarXiv cs.AIOct 14Research
- Google uses Gemini 4 Argon to turn videos of people playing cornhole into robot training simulationsOfficialX @GoogleAIOct 14Research
- OpenAI studies the internal signals behind metagaming in an o3 reinforcement learning runOfficialOpenAI Alignment Research BlogOct 14Research
- Ben Recht says OpenAI's claimed faster Fourier transform is a galactic algorithm with no practical speedup, per arg min
- Looped-DiT paper shows a 260M looped text-to-image model beating a 6.5× larger baseline with less inference compute
- Researchers release U-Space, a label-free way to read LLM uncertainty from hidden states in a single pass
- Google DeepMind's AlphaProof Nexus paper, which solved 9 open Erdos problems in Lean, is published in Science
- Microsoft researchers' Fork-and-Flush clears agent chat history but keeps the workspace, lifting long-run resultsOfficialarXiv (Microsoft Research)Oct 13Research
- Qwen introduces QwenGyre, an elastic RL framework for training agents on extremely long tasks
- BrokenML benchmark tests how much cluster access AI agents need to fix broken ML jobs
- Kaiming He's team releases VISTA, a visual harness that lets multimodal models re-inspect past frames, per QbitAI
- Huawei's Memento 3 agent clears all 25 ARC-AGI-3 public games with 56% fewer actions than humans, per QbitAIReportedQbitAIOct 12Research
- CL4SM opens a call for contributions on constrained learning for scientific modelling ahead of NeurIPS Paris
- Leading Chinese mathematicians unite behind use of AI in research, per South China Morning PostReportedSouth China Morning PostOct 12Research