Researchers discover 'template collapse' in AI agents that traditional metrics miss, proposing mutual information as a better measure of reasoning quality.

RL training of multi-turn LLM agents exhibits inherent instability where reasoning quality directly impacts task performance

Models can develop input-agnostic fixed templates that appear diverse by entropy measures but fail to respond appropriately to different inputs

Traditional entropy metrics cannot detect template collapse, revealing a critical blind spot in existing stability measurement approaches

RAGEN-2 introduces mutual information (MI) proxies to diagnose reasoning quality by measuring both within-input diversity and cross-input distinguishability

Mutual information correlates with final task performance significantly more strongly than entropy across diverse tasks, making it a more reliable proxy for agentic reasoning quality