Daily AI news, sourcedUpdates / Research
Updates Research · 222 items · newest issue first · page 6 of 12
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- LocalLLaMA research post: a swarm of small Jev agents plays Sherlock Holmes murder casesUnconfirmedr/LocalLLaMAOct 11Research
- Sakana AI’s Multi-Layered Review catches 73% of core-claim errors in LLM peer review, per MarkTechPostReportedMarkTechPostOct 11Research
- Paper calculates the Lyapunov spectrum of random recurrent neural networks, showing their chaos is extensiveUnconfirmedarXiv cond-mat.dis-nnOct 10Research
- Association for Human Pharmacology criticizes OpenAI's October 6 release of disease-cure research papersUnconfirmedr/singularityOct 10Research
- Essay "What Are the Monkeys Typing?" argues we can see what capable AI systems do but not reliably whyUnconfirmedr/singularityOct 9Research
- Paper tests whether AI agents can do open-ended scientific discovery in the Station multi-agent environmentUnconfirmedr/MachineLearningOct 9Research
- Developer shows ALHR, a tree-based sparse attention method claiming 35x KV-cache read reductionUnconfirmedr/MachineLearningOct 9Research
- ByteDance Seed finds DeepSeek-V4's long-context recall swings with token position, per QbitAIReportedQbitAIOct 9Research
- DLoop paper proposes looped speculative decoding for faster LLM generationUnconfirmedr/LocalLLaMAOct 9Research
- Stepped MoE paper proposes segment-level routing with configurable inference costUnconfirmedr/LocalLLaMAOct 9Research
- Microsoft tests how reliably Foundry's multi-turn evaluators grade voice-agent calls
- Paper's EngramEdit aims to update facts in conditional-memory LLMs like DeepSeek Engram while keeping the Transformer fixedUnconfirmedr/LocalLLaMAOct 9Research
- Set theorist Asaf Karagila rejects OpenAI's Partition Principle preprint as unreadable and refuses the whisky betUnconfirmedAsaf KaragilaOct 9Research
- Association for Human Mathematics urges boycott of OpenAI after AI math manuscript dump, per The Decoder
- Nullify steers LLM activations in the null space for training-free unlearningUnconfirmedarXiv cs.LGOct 9Research
- Leakage-controlled multimodal learning targets Alzheimer's diagnosis and progressionUnconfirmedarXiv cs.LGOct 9Research
- Coverage-aware reasoning with medical tokens aims to improve diagnosis predictionUnconfirmedarXiv cs.LGOct 9Research
- Study finds visible chain-of-thought is not universally helpful across personas and thinking stylesUnconfirmedarXiv cs.LGOct 9Research
- Paper proposes dynamic cross-loop on-policy distillation for looped language modelsUnconfirmedarXiv cs.LGOct 9Research
- Study uses LLMs to help prepare transportation management plans for work zonesUnconfirmedarXiv cs.CLOct 9Research