Open-AgentRL
Open-AgentRL is an open-source collection of reinforcement-learning research for large language model agents, including a system that jointly trains policy and reward models, a study of what makes agentic RL training effective, and a method for dynamic tool selection.
Open-AgentRL packages three related research projects on reinforcement learning for LLM agents. RLAnything is a framework that optimizes the training environment, the policy and the reward model together in a closed loop, using consistency feedback to jointly improve the reward model and automatic environment adaptation informed by both models' feedback.
DemyAgent studies agentic reinforcement learning along three axes, data, algorithms and reasoning mode, and reports that real end-to-end trajectories and diverse datasets outperform synthetic data, that reward clipping and entropy maintenance speed up training, and that deliberate reasoning with selective tool calls beats either frequent tool use or lengthy self-reasoning; it also releases supervised and reinforcement-learning datasets that let small models outperform much larger ones on reasoning benchmarks.
AutoTool trains agents to dynamically select and integrate tools from large, evolving toolsets, backed by a 200,000-trajectory dataset covering over 1,300 tools across many task types.