Topic: Research
1414 stories · page 4 of 29
- Into the Unknown benchmark: no AI agent meets scientists' bar
- HexaShell harness lets a coding agent reach 39.7% on RoboDojo
- GPT-6-Astra report: robot competence across 12 benchmarks, not yet practical embodied intelligence
- ECHOES trains GUI agents to replace screenshot history with text
- DLite speculative decoding drops history to speed up agentic LLMs
- Copy-Paste Language Models add a lightweight copy head to LLMs
- CarryOn organizes long-horizon agent context around task objects, not histories
- Waypoint-OPD gates distillation by early and late RL policy shifts
- Virtual Head Attention paper claims smaller KV cache and better quality than GQA
- Study finds GPT-5.5 and other AI models omit flaws from their own work reports
- SpecAgent has a weak model draft agent steps and a strong model review them
- RoTA-Bench: GPT-5.6-Sol solves only 17% of 3D rotation tasks
- OneWM paper reports world modeling and action learning help each other
- MRTPO paper assigns each reward's advantage to its relevant tokens
- Mathematicians dispute OpenAI's account of its Navier-Stokes proof
- LeanPRM uses Lean to auto-label proof steps for training verifiers
- Genomic AI like Evo 2 and AlphaGenome can't explain the genome, Quanta argues
- FlashMask-3 brings mask-aware attention to Hopper GPUs and context parallelism
- Data World synthesizes training environments for data agents
- CoDA trains robot policy and simulation data engine together, lifting OOD success to 27.6%
- CarryOn harness cuts a robot VLA's violation rate from 52.6% to 6.1%
- Anthropic's Claude traces NSA-supplied ECDSA curve seeds to lrand48 in under 30 minutes
- WorldCast benchmark: GPT-Image-2.5 System scores 72.8, best open model 20.1
- VLA-Δ trains robot models on physical state changes, beats π0.5 on real robots
- Sparse autoencoder features are only partly additive, paper finds
- Self-Evolving Agent Harness paper formalizes how agents revise their own scaffolding
- Mosaic trains agents across multiple sessions, beating broadcasting on BrowseComp
- MLExpert, a 35B agentic model, lifts MLE-Bench Lite medal rate from 37.9% to 72.7%
- MJev-Spatial answers spatial questions by classification, not text generation
- Mercor's Benchmark Map: about 90% of US work hours lack AI benchmark coverage
- Google Research ties AI agents to wet-lab robots, finds nanomolar protein binders
- Gemini 4 Argon places 3rd as GPT-6 Astra leads StarCraft bot benchmark
- FDE-Gym benchmark: best of 13 AI agent setups resolves only 14 of 95 delivery tasks
- Blog post derives memory capacity of linear layers, MLPs and attention
- Velox speeds up video world model attention 4.36 times at chunk 48
- Terminal-RSI trains a 27B terminal agent with recursive self-rewrite and policy-refine
- Speculative Forcing cuts denoising steps in autoregressive video generation
- Routing Policy Optimization uses attention routing for diffusion RL, claims nearly 4x training efficiency
- Qwen3-8B loses its step counter in 20 of 20 seeds when irrelevant text is added
- Orbia benchmark tests whether video world models keep a consistent 3D scene
- NVIDIA announces Nemotron-pretraining-Indic-v1, about 2.8T tokens in 22 Indian languages
- Microsoft Research's Magnetar proves Rust, Go and C programs with hash-only security
- JevAgentHarness hands agent tool choices to a 4B model
- Intern Lumina U2 uses multi-codebook visual tokens in a diffusion language model
- Google Research shows 10 updates at Research@ 2026, from Earth AI to AMIE
- FreeSolo sparse attention extrapolates to longer contexts without long-context fine-tuning
- Essay argues neural networks are already clean, superposition was misread
- EdgeWise lets a small local model hand hard agent steps to a cloud LLM
- Cheap RAG claim verifier ties a frontier LLM on score but admits more unsupported claims
- 3DCodeVerse introduces a 3D code dataset and multi-agent harness for coding agents