RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning

TL;DR

This paper introduces RAGEN, a modular framework for training large language model agents with multi-turn reinforcement learning, addressing challenges like reward variance and reasoning emergence, and providing insights into effective training strategies.

Contribution

The paper proposes StarPO, a novel trajectory-level RL framework, and RAGEN, a system for training and evaluating LLM agents, with new techniques for stabilizing training and enhancing reasoning.

Findings

Identification of the Echo Trap phenomenon in agent RL training.

Stabilization techniques like trajectory filtering and critic use improve training.

Diverse initial states and frequent sampling enhance RL rollout shaping.

Abstract

Training large language models (LLMs) as interactive agents presents unique challenges including long-horizon decision making and interacting with stochastic environment feedback. While reinforcement learning (RL) has enabled progress in static tasks, multi-turn agent RL training remains underexplored. We propose StarPO (State-Thinking-Actions-Reward Policy Optimization), a general framework for trajectory-level agent RL, and introduce RAGEN, a modular system for training and evaluating LLM agents. Our study on four stylized environments reveals three core findings. First, our agent RL training shows a recurring mode of Echo Trap where reward variance cliffs and gradient spikes; we address this with StarPO-S, a stabilized variant with trajectory filtering, critic incorporation, and gradient stabilization. Second, we find the shaping of RL rollouts would benefit from diverse initial states, medium interaction granularity and more frequent sampling. Third, we show that without fine-grained, reasoning-aware reward signals, agent reasoning hardly emerge through multi-turn RL and they may show shallow strategies or hallucinated thoughts. Code and environments are available at https://github.com/RAGEN-AI/RAGEN.