RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning
What this work does
Trains multi-turn trajectories with state, reasoning, action and reward records; analyzes collapsed within-group reward variance and stabilizes updates with filtering and critic-related changes.
Connection to discovery
Useful for measuring whether exploration diversity and reward variance survive long discovery rollouts.
Evidence
Controlled bandit, Sokoban and Frozen Lake settings isolate stochasticity and interaction, with WebShop as a language-grounded task.
Scope and limitations
Stylized tasks and chosen reward structure bound the findings. Training collapse and absent learning signal should be diagnosed before attributing failure to missing discovery representations.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.