Topic: Research
1414 stories · page 13 of 29
- EvolveTrade gives LLM trading agents a self-updating policy
- ComPO aligns LLMs via comparison oracles, not gradients
- Census Bureau finds AI-exposed grads facing recession-like job losses
- Andon Labs puts AI agents in charge of real businesses
- Vision-RL2 cuts MLLM visual tokens 4x without losing accuracy
- Video DeltaNet speeds MiniMax H3 video generation 14.5x
- Robotic hand teaches itself to walk on its own fingers
- Researchers find 'Value Flattening' flaw in PPO critics, propose fix
- PANORAMA grounds captions in pixels, tops new PanoCaps benchmark
- PACT finds enterprise AI assistants break rules under pressure
- OpenAI's Navier-Stokes proof sparks a credit fight with mathematicians
- LLM test-time scaling's GPU cost swings up to 6x with batching
- ICLR 2027 draws roughly 50,000 abstracts, up 2.5x from last year
- HypoEvolve evolves scientific hypotheses with genetic algorithms and LLM agents
- Grant Sanderson argues math needs credit for explanation, not proof alone
- GPT-Policy teaches robots new tasks from video, no retraining needed
- Fuse framework finds LLMs stumble at inferring hidden motives
- Fathom speeds long-context KV cache decoding 1.67x with per-query reads
- EvoSkill-GUI lets agents rewrite their own GUI skills without retraining
- CERA-MoA co-evolves agent routing and training in Mixture-of-Agents systems
- ActObs improves RL training for Qwen3 coding agents by also predicting observations
- XConf estimates LLM confidence from the model's own past episodes
- VC-Attention speeds up low-bit video attention by up to 3.6x
- UFO framework evaluates image generation across all conditions at once
- Terence Tao warns AI proof-solving threatens open math
- SoL-Pi cuts coding-agent token traffic by up to 49%
- RetireOPD teaches agentic RL models to fire their own teacher
- Researchers propose Infinite-Parameter LLM that learns weights live
- Reflective Recovery teaches LLMs to self-correct reasoning errors
- Qwen3.5-4B beats Claude Opus 4.8 on new UC-Bench conflict benchmark
- PCA finds stylistic axes in LLM activations, but only in some models
- ModularRSI splits agent harness self-improvement into five modules
- MIT Technology Review debates whether AI could destroy humanity
- JEPA-Anything extends one predictive architecture across seven domains
- Four schedules push MoE training to 1M-token context
- Edge0 runs 35B-parameter MoE models from SSD on a single 24GB GPU
- Context management pays off most as coding-agent budgets shrink, study finds
- Study finds in-context learning emerges across six modalities
- Study finds AI models vary widely in when they disengage from verbal abuse
- ScienceIDE trains AI agents on scientific code repositories
- RelateAnything predicts scene relations from any vocabulary in 20ms
- ProgramDistill: GPT-6 Astra beats Claude Opus 5 on undocumented-feature benchmark
- LLMs distort their own personality scores when told to fake good or bad
- LLMs commit to relation type before entity in factual recall
- ImpossibleRubrics benchmark: LLM-written rubrics reward fabricated answers up to 36% of the time
- HarnessVLN beats prior training-free robot navigation systems
- Google DeepMind launches DeepMind Institute for AGI safety and governance
- FLAT trains one encoder for both text-to-image and image-to-text tasks
- Dream-RSI lets AI agents replay past discoveries to self-improve
- BITCOS breaks the 1.585-bit barrier for ternary LLM weights