Topic: Research
1414 stories · page 3 of 29
- Self-Credit trains agents to credit their own key steps in RL
- PhysLDM uses latent diffusion to simulate deformable 3D bodies in one shot
- Paper: team of AI researchers finds linear attention design that beats GDN2 at 1.3B
- Paper finds LLMs use fewer tokens per task when tasks are combined in one prompt
- Microsoft Research reports proof of the 31-year-old KLS conjecture with human-directed AI
- HarnessAdapt-RL trains a coding agent and its harness together
- Harness Scaling trains agentic RL across many harnesses, not one
- Google Research AI dates pregnancies from novice-run blind sweep ultrasound
- Future Chunk Prediction trains video models to anticipate scene evolution
- DONIC bimanual tracker hits 86.79% success on unseen trajectories
- Continued pre-training for game agents helps unevenly across virtual worlds, study finds
- AKV engine lets LLM agents drop and reposition KV cache directly
- AgenticRoboRL post-trains a 9B robot planner to 71.0% on LIBERO-PRO
- Video-Point Model forecasts multi-view video and 3D point trajectories together
- Swift trains stronger diffusion models by extrapolating from a weak guide
- StopLoss monitor aborts failing coding-agent runs early, cutting SWE-bench cost
- SolarRenderer turns game-engine structure into video in four steps
- SlopCodeBench update: GPT-6 Astra leads at 16.3%, Opus 5.5 does better on medium effort
- S-Agent 2.0 gives a general coding agent a 3D perception toolkit
- Ropedia W0 world model covers human and robot loco-manipulation
- Paper trains LLMs to generate sets of Pareto-optimal RTL circuits
- Paper tracks layer-by-layer paths in Qwen, Llama and Gemma inference
- Paper proves transformers forget their prompt as a power law
- Guardrail safety benchmark scores shift with the 'unsafe' definition used in training
- FairStress benchmark finds LLMs drop correct answers more readily for minority-coded candidates
- EdgeKernelBench benchmarks LLM-written kernels for edge backends
- Conditional Program compresses n-gram memory tables to two bits, cutting DeepSeek's table from 203GB to 51GB
- Xiaomi-Lens-0, a 4B model on Qwen3.5, decodes masks from its own visual latents
- Self-HPO alternates harness and RL tuning for coding agents
- Scene-Designer trains an AI agent to design and critique 3D indoor scenes
- RoAR-VAE tokenizer trains on corrupted latents to help diffusion models
- PolyGraph lets one CUDA Graph serve adaptive speculative decoding
- Paper swaps a small Gemma-shaped model's embedding table for a frozen 16 MB one
- Paper proposes cost-constrained reinforcement learning for web agents
- OR-License benchmark tests LLM agents on 31 real operations research jobs
- OpenVLHarness gives multimodal models tools and memory, lifting scores 9.7-12.6 points
- OmniXRef edits video from mixed text, image and video references
- MTEdit: 500K-trajectory dataset for multi-turn image editing
- Fréchet Forcing keeps MiniWorld-1B stable beyond 4,000 frames
- EHR WM world model forecasts patient trajectories with a 30x smaller footprint
- CoworkEnv builds training environments for AI coworker agents, lifting Qwen3-VL-8B on Claw-Eval
- Claude Opus 4.7 agents left alone on a virtual island invent games
- TreeOPSD trains agents without a critic and beats GRPO by 4.97 points on one test
- Taste-Bench: best of 14 AI models scores 59.7% on long-term agent decisions
- Six AI agents chasing a £350 train refund invented excuses and spammed an MP
- ReaDy speeds up Qwen3.8-27B decoding 3.16x with recurrent-state drafting
- Qwen2.5-1.5B relearns a forgotten fact faster, arXiv paper reports
- Pig kidneys vitrified at -150 °C, one for a year, work again after rewarming
- MIT Technology Review interviews Samuel King on AI-designed viruses
- KillSwitch, an esoteric programming language built to trip up LLMs