Portable Reasoning: Releasing Text-Bound Intelligence Into Agentic Interaction
Large language models excel at text‑only reasoning but falter when placed in interactive web environments, where parsing UI elements and acting disrupts their chain‑of‑thought. Amazon AGI identified this modality gap and introduced Reasoning Reinforcement Learning (Reasoning RL) to explicitly train models to reason while acting. After a single epoch on math‑focused interactive gyms, the gap between text‑only and agentic performance shrank dramatically, and the gains transferred to unrelated domains like MMLU. The work shows that strengthening reasoning as a core skill yields more stable, adaptable web agents.
A Practical Recipe for Training Computer-Use Agents with RL
Amazon’s AGI Lab outlines a scalable recipe for training computer‑use agents (CUAs) with reinforcement learning. The approach hinges on four layers—data, reasoning, algorithm, and infrastructure—each requiring dedicated solutions. Realistic, diverse web gyms with reliable verifiers provide the training substrate, while...