Mitigating Conversational Inertia in Multi-Turn Agents

In three linesLLMs exhibit 'conversational inertia' in multi-turn agent scenarios: they over-imitate their previous responses instead of exploring. Authors identify this bias through attention analysis and propose Context Preference Learning to favor low-inertia responses. Validated across 8 agent environments.

Summary generated by Claude — human-verified