Topic: Research
1414 stories · page 1 of 29
AI papers: arXiv preprints, lab publications, new training and evaluation methods. Each story says what the authors did, on what data, and what came out of it. Every story links to the paper.
- ThinkWorld trains robot VLA models on action-free human video
- Steering vectors transfer between text and image generation in unified multimodal models
- SPOT drafter speeds up diffusion language model decoding by up to 3.4x
- RSI Arena: ten AI agents post-train models, then learn from human votes
- Rollback trains robot policies by rewinding world models to past decisions
- ResCom adds a low-rank fix for 4-bit LLM quantization error
- ParetoRSI-Gym benchmark tests AI agents improving LLMs across five domains
- Paper replicates METR developer study, says AI time savings must be counted with errors
- Paper finds top reasoning effort costs up to 42x more on easy tasks
- iPrefBench tests whether iOS agents know when to infer and when to ask
- IEEE Access catches two AI-faked papers in a war-game test
- Google Research proposes a sleep phase for LLMs to retain new knowledge
- Classification of finite simple groups formalized in Lean
- ChatGPT gets OpenAI partner content as pre-made summaries, other web pages as raw snippets
- Anthropic reports early results from Claude drug-development work
- Ai2 study finds LLM novelty judges swing wildly on small prompt changes
- WorldCast generates consistent character video under large-angle camera moves
- World-in-Context turns view, action and camera controls into tokens for video world models
- Trajectory2Gym trains tool-use agents on grounded decisions, lifting Tau1, Tau2 and BFCL
- TraceQuery cuts fresh computation in diffusion LLM decoding by recomputing only trajectory-relevant tokens
- TenetBench: 11 of 14 LLM setups write text that conflicts with ethics codes
- StreamVibe: a benchmark and streaming model for video agents that infer implicit user intent
- Qwen-Image-PE teaches prompt enhancement visual design, 9B model released
- OpenESM paper claims energy-steered models beat GPTs on most of 26 tasks
- OpenAI says its models refuted the ETH and proved local smoothing
- NYU and UT Austin study finds six patterns that make AI text hard to read
- MMVistaReason recipe trains a 4B multimodal reasoner on open data
- InteractiveBench tests 12 MLLMs as egocentric AI assistants
- DynaFlow trains an agent to build its own multi-agent workflows
- DF claims all-diamond MOSFETs with 50x faster switching
- Continual Learning Exam benchmark: best LLM agent clears 47.7% of levels
- Δτ-Bench tests AI agents as their tools change; best result is 42.3%
- UserArena benchmark offers 64,729 real shopping sessions to test user simulators
- TWIG paper represents every neural network weight as a shared path of discrete choices
- SplatLab runs 3D Gaussian Splatting research end to end, reports up to +3.78 dB PSNR
- RynnUnify: 16B embodied model joins scene understanding, video planning and robot control
- Qwen3.5-27B activation edits fix target picks but not condition judgments
- Paper uses sparse autoencoders to make AI-text detectors auditable
- Paper proposes on-policy self-training that lets LLMs improve with no human data
- Paper proposes attention-free Causal Grassmann layer, within about 10-15% of Transformers on Wikitext-2
- Paper finds .916 overall F1 hid scorer errors in LLM follow-up prompt tests
- MisconceptDialogue: scoring method moves LLM tutor success by 39 points
- HyperDrone quadrotor simulator claims 20x speedup and millions of FPS
- FilmFlow uses a Director agent and three helpers to plan 3D camera moves
- Darmok AI solves Times cryptic crossword final faster than human winner
- BrickWright lets coding agents build LEGO models that are physically valid
- Anthropic's Jack Clark moves self-improving AI forecast to 2027 after TasteVal benchmark
- Agentic framework traces emotion-reasoning errors and patches prompts, lifting MER2025 F-score 0.6 to 2.2 points
- US creative industries lost 200k jobs in four years: how much is AI?
- TTR-Compressor uses test-time regression to compress token sequences