OpenManus-RL

Why we included this project

People researching how reinforcement learning can make language models act as better agents will find OpenManus-RL one of the more transparent projects in the area. It is a collaborative research effort led by UIUC and MetaGPT that applies RL post-training techniques such as GRPO, PPO, DPO, and preference-based reward modeling to agentic models, borrowing from the reasoning-model work behind DeepSeek-R1 and QwQ-32B. The repository combines training code built on the verl framework with an open agent fine-tuning dataset hosted on Hugging Face, and it reports results on benchmarks like GAIA, AgentBench, WebShop, and OSWorld. Rather than a turnkey product, its value is visibility: the team publishes its roadmap, rollout strategies (ToT, MCTS, Graph-of-Thoughts), and reasoning formats as it updates the project in a live-streaming style. For engineers and researchers who want to reproduce experiments or track where agent RL tuning is heading, that openness matters more than polish.