Agent-R1

Why we included this project

Agent-R1 is a modular framework for training multi-step LLM agents with end-to-end reinforcement learning. Instead of treating each interaction as one growing prompt-response sequence, it models every turn as a step-level MDP transition, so the model observes an environment, takes an action, and reacts to tool or environment feedback before the next turn. That keeps reward assignment and policy optimization explicit across the whole trajectory, rather than squeezing everything into a single context window. The layered abstractions let you reuse the same trainer across quite different tasks, and the repo ships documented recipes for benchmarks like GSM8K, HotpotQA, ALFWorld, and WebShop, plus processed datasets on ModelScope, so a training run is reproducible without wiring up every component yourself. The technical report and docs explain the step-level representation and context management, which matters if you plan to adapt the framework to your own environments or swap in training strategies such as StepPO or GRPO without rewriting the entire stack.