Topic: Research
1414 stories · page 12 of 29
- GameHorizon Suite benchmarks 47 AI models on AAA gameplay
- Flash-dLLM speeds up diffusion LLMs by up to 11x
- AIDE² rewrites its own code and generalizes to new tasks
- Runtime fix restores safety in 73% of ACC perception failures
- RoPE flaw explains why video diffusion models break physics
- RoboDawn transfers vision-language model intelligence to robotic control
- RAND urges the US to keep its options open on superintelligence, as scientists grow human brain tissue in mice
- PsyAgentBench finds LLM psychological effects come from several different mechanisms
- PARTS retrains pretrained robot policies only on the subtasks they keep failing
- OpenAI forms math advisory group as its AI claims 100+ solved problems
- onPanda cuts LLM annotation time by 52% in small study
- OmniVChat-RL training improves Qwen3-Omni-Instruct's dialogue skills
- Multiverse Computing prunes LLMs using an Ising spin-glass optimization
- Irrelevant context degrades LLM retrieval accuracy, study finds
- Harness-Zero folds specialized agent-harness behavior into model weights
- Gebru and Bender debunk Anthropic and OpenAI's summer of AI hype
- Designer-RSI lifts AI graphic-design agent success to 99.3%
- CLIC tool maps how 10 LLMs differ in coding style, not just pass rate
- BI-Agent lifts LLM accuracy on real business intelligence tasks
- Apollo, the first LLM for Ancient Greek, aims to fill gaps in torn papyrus
- Anthropic builds a biology lab where Claude guides robots through drug experiments
- A single model claims to fuse any modality without task-specific adaptation
- VākQA benchmark tests AI on Telugu spoken question answering
- Runway demos real-time AI video streaming with GWM-1
- Reviser generates text via cursor actions, not left to right
- Recursive language models beat chain-of-thought out of domain
- Po-Shen Loh argues AI advances will create more jobs than it cuts
- Paint-Anything adds exact hex-color control to image models
- OpenAI finds ChatGPT use outside a worker's own job keeps recurring
- OmniVBench benchmarks omni reference-to-video generation with 12,172 checklist items
- MIT Technology Review editors answer readers' questions on AI extinction risk
- Microsoft's StudentSim trains AI tutors on simulated students
- GraphSkillEvo evolves graph-structured skills for LLM agents
- GPT-6 Astra leads new computer-use benchmark but fully passes just 2.8%
- Epoch AI survey finds daily AI use among US adults doubled in six months
- CodeMidas builds RL environments for coding agents from raw source code
- Cal-OPD strips teacher noise out of on-policy distillation
- AMPLE-Math benchmark shows reference solutions add little to self-distillation
- AI trained on AI-written peer reviews loses judgment diversity, study finds
- When2Think cuts reasoning tokens 28% by skipping thought on easy problems
- VA-Bench finds MLLMs struggle at embodied spatial reasoning
- TACLS pairs GNSS satellite data with machine learning to catch flash floods sooner
- SELF-INDEX lets retrieval indexes optimize themselves without humans
- RiskChainBench evaluates AI on disguised links and website risk calls
- Register tokens lift diffusion LLM scores on math and code
- Qwen3, Llama and Gemma disagree on stop tokens, inflating distillation length
- MiniMax-H3 scores 41.97% on new physical-world reasoning benchmark
- LimiX-2 debuts contextual mechanism networks for tabular data
- Google Research open-sources MilleMiglia, a middle-mile logistics benchmark generator
- FAMOS predicts 3D object articulation from sparse views