People thought big AI models were all learning the same overall picture of the world, but those measurements were secretly biased by model size and depth.
This report studies the biggest new dangers from super-capable AI and tests them in realistic, well-controlled labs so we can fix problems before they cause real harm.
This paper studies Moltbook, a giant social network made only of AI agents, to see if they start acting like a real society over time.
AutoWebWorld builds pretend websites with clear rules so AI can practice safely and be checked automatically.
Sparse autoencoders (SAEs) are popular for explaining what large language models are doing, but this paper shows they often don’t learn real, meaningful features.
Not all wrong answers from large language models (LLMs) mean they never learned the fact—many times the model knows it but can’t pull it out on demand.
This paper builds GUI-Owl-1.5, an AI that can use phones, computers, and web browsers like a careful human helper.
This paper teaches AI models to learn like good students: try, think about what went wrong, fix it, and remember the fix.
Video generators are slow because attention looks at everything, which takes a lot of time.
SLA2 is a new way for AI to pay attention faster by smartly splitting work between two helpers: a precise one (sparse attention) and a speedy one (linear attention).
SkillsBench is a big test playground that measures whether giving AI agents step-by-step 'Skills' actually helps them finish real tasks.
UniT teaches one multimodal model to think in steps with pictures and words, so it can check its own work and fix mistakes as it goes.