rllm

Why we included this project

Most people try to improve LLM-based agents by writing better prompts. rLLM instead treats the agent as something you can train, applying reinforcement learning directly to the model that drives it. The framework handles the fiddly operational parts: you bring your own agent harness or wrap one, run rollouts in Docker, Daytona, Modal, or locally, and switch training backends with a single flag. Because the same agent code drives both evaluation and training, teams avoid the common mismatch between how a model is scored and how it is updated. With 60+ built-in benchmarks and support for GRPO, REINFORCE, and RLOO, it is a realistic starting point for labs and applied research teams that want modern RL training results without assembling the distributed-training stack themselves.