Topic: Research
1580 stories · page 1 of 32
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.
- Wild Dolphin Project uses AI with DeepMind to find patterns in dolphin sounds
- VisSearch-Bench: best of 16 multimodal models fully solves 26.47% of tasks
- Thinking Machines' Inkling 2 Small swears 316 times in benchmark runs, Vals AI finds
- SwarmEvolve uses latent-space agent swarms to speed up program search
- Study finds AI auditors catch most agent misbehavior but raise many false alarms
- Sparse autoencoders show Qwen3-30B-A3B MoE experts specialize finely
- Soft-MTP: a one-line fix for multi-token prediction training
- Paper tests rewriting past reasoning into compact records to cut future tokens
- Open-HumanScale releases 50,000-hour egocentric video recipe for robot models
- Muir linear attention reports 1.8-2.0% lower perplexity than GDN-2
- Microsoft Research's AI Frontiers group is over, says a founding leader
- MaxxBench rewrites 89 DeepSWE tasks; equal-scoring coding models split apart
- Into the Unknown benchmark: no AI agent passes experts' bar on open research
- Harness-S trains agents with sibling harnesses on AppWorld, SWE-Gym, DBBench
- DreamFree robot model skips video generation, runs a 14B backbone at 11Hz
- VFlash predicts whole visual plans in one pass, with a claimed 25x speedup
- ToyBlocks introduces a controlled simulated world for robot learning
- SLA adds 3D spatial alignment to pi05 robot policy during continual pre-training
- SIMPLE law predicts LLM inference energy with four parameters
- PLatour paper finds some cross-model transfer gains need no real sender signal
- OpenAI introduces BioBasicsBench, a 100-task wet-lab benchmark for AI agents
- Octopus trains a 4B model as many agents sharing one KV cache
- METR report: Anthropic's capabilities progress ran about 2x faster in a year
- Memorilla packs document collections into 16 memory tokens for frozen LLMs
- Hightouch: Jev plus Opus hybrid hits frontier accuracy at over 4x lower cost
- Frontier Radar #5 maps 45 AI neolabs founded by big-lab exits
- DPDM paper uses a frozen LimiX-2 predictor for agent decisions
- D3D diffusion drafter reaches 5.00x speedup over DFlash's 4.03x on Qwen3-8B
- BlockLock makes LLM inference reproducible without custom kernels
- Ampere open-sources Panthera, its cycle-accurate CPU performance model
- 2011 essay answers Chomsky: statistical language models are real science
- Village's LLM Atlas: open-weight models win only 3% of budgets, down from 41%
- VideoMemBench: Claude Opus 5 tops video-memory test at only 45.7%
- TurnWise brings turn-level credit to streaming video RL
- StreamDuo gives a live game commentator a wiki research agent
- SpecOPD proposes branch-parallel distillation for faster speculative decoding
- SemiAnalysis releases GDPx work benchmark; top agent, Claude Opus 5.5, scores 12 percent
- Russian machine-text detectors disagree when only task wording changes
- PyRUA-DataFarm: coding agent builds robot training data without teleoperation
- Paper: multi-turn agentic RL gains shrink when turn budgets tighten
- OpenAI finds GPT-6 Astra covertly underperforms on safety research tasks
- NEO-VLA drops the vision encoder for robot action policies
- H2RB co-trains robot VLA policies on unpaired human demos
- Google Research AI detects signs of hypertension from a smartwatch
- FrontierPhysics benchmark: AI agents struggle with end-to-end physics research
- Axiom releases Lean formalization of the entire Stacks Project
- YANAMI introduces 400K multimodal dialogues to test stateful interaction
- WebHarbor releases 153 dockerized website environments for training GUI agents
- ToolSpace builds MCP agent training environments from real tool-call traces
- Stress test finds LLMs overestimate causal links in causal discovery