Research

Post-training a small model to follow a live stream of typing and learn when to act or stay quiet.

An ultra long-horizon benchmark for coding agents doing engineering and research work.

Reinforcement-learning work for the browser-use model as a research intern.

An RL setup in which language models learn their own compact codebooks.

Hidden-information games for studying deceptive behavior learned from reward.

Cross-lingual alignment through encoder injection for low-resource languages.