Daily AI news, sourcedUpdates / Research
Updates Research · 222 items · newest issue first · page 5 of 12
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- SynAct helps LLM agents decide where scarce expert feedback will change the outcomeUnconfirmedarXiv cs.LGOct 12Research
- Paper treats multi-horizon event sequence forecasting as conditional generative modelingUnconfirmedarXiv cs.LGOct 12Research
- DiffusionART aims for efficient, stable RL fine-tuning of diffusion modelsUnconfirmedarXiv cs.LGOct 12Research
- Theory paper shows transformer masked diffusion models provably learn distributions with latent structureUnconfirmedarXiv cs.LGOct 12Research
- CacheScope speeds up diffusion transformers by choosing where to spend exact compute stepsUnconfirmedarXiv cs.LGOct 12Research
- GeopoliticalQA benchmarks RAG on 5,525 questions over 415 intelligence reportsUnconfirmedarXiv cs.CLOct 12Research
- Paper studies language models generating their own replay data to preserve capabilities during continual learningUnconfirmedarXiv cs.CLOct 12Research
- Study measures transferable gains versus prompt-level overfitting when optimizing a deep-research agent's promptsUnconfirmedarXiv cs.CLOct 12Research
- Elastic Latent Recurrence trains looped language models without the one-loop accuracy taxUnconfirmedarXiv cs.CLOct 12Research
- SWE-Duel proposes a self-scaling arena for coding agents that creates tasks at evaluation timeUnconfirmedarXiv cs.CLOct 12Research
- CLEAT trains interactive agents to ask clarifying questions guided by their decision valueUnconfirmedarXiv cs.AIOct 12Research
- Behavioral Self-Rewarding lets SWE agents judge their own debugging trajectories beyond pass/fail testsUnconfirmedarXiv cs.AIOct 12Research
- SpireBench tests 20 LLMs on full games of Slay the Spire for long-horizon strategyUnconfirmedarXiv cs.AIOct 12Research
- Agentic data synthesis evolves its generation space to close the loop between learner weaknesses and new training dataUnconfirmedarXiv cs.AIOct 12Research
- Position paper argues explainable-AI faithfulness metrics for feature importance are structurally misalignedUnconfirmedarXiv cs.AIOct 12Research
- Terence Tao's Math 2.0 Caltech talk argues mathematics is entering an era of proof abundanceStatementTerence Tao slidesOct 12Research
- Sakana AI publishes MASS, where one model organizes copies of itself into research teams and trains on what worked
- 414k-parameter transformer learns boid flocking; probes recover rules the model may not useUnconfirmedr/MachineLearningOct 12Research
- NeurIPS 2026 Paris complimentary Top Reviewer registration already full, per a Reddit discussionUnconfirmedr/MachineLearningOct 12Research
- arXiv caps submissions at two per person per month as AI papers overwhelm moderators, per The DecoderReportedThe DecoderOct 11Research