AReaL

Why we included this project

Making an LLM reason more reliably, or behave like an agent inside a real system, usually means standing up a reinforcement learning loop and getting rewards, trajectories, and scale to play nicely together. AReaL is training infrastructure built for that specific job. It runs on a fully asynchronous RL paradigm, so it avoids the synchronization bottlenecks that slow down large reasoning and agentic models, and it works with black-box agent stacks: point it at a base_url and run online RL against whatever you already have. The project ships worked examples for math, coding, and end-to-end software-engineering tasks, plus a lite variant whose algorithm-first API suits researchers prototyping new RL methods. If your team has been putting off agentic RL because the setup looks expensive, this is the tool designed to bring that cost down.