RAGEN-2: Reasoning Collapse in Agentic RL | Multi-Turn Reinforcement Learning for LLM Agents

RAGEN-2 (ICML 2026 Oral) finds template collapse, reasoning collapse in agentic reinforcement learning invisible to entropy, fixed by SNR-aware filtering.

关键词：RAGEN-2, reasoning collapse, template collapse, agentic reinforcement learning, multi-turn reinforcement learning, LLM agents, agent RL, mutual information, SNR-aware filtering