Reproduce “Mitigating Conversational Inertia in Multi-Turn Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Wan, Zheng Cao, Zhenhao Zhang, Zhengwen Zeng, Shuheng Shen, Changhua Meng, Linchao Zhu
Researchers from Zhejiang University and Ant Group identified 'conversational inertia' in multi-turn LLM agents, characterized by a self-imitation bias from diagonal attention patterns to past outputs. They introduce Context Preference Learning (CPL), a reward-free calibration method, and Clip Context, an efficient inference-time context management strategy, which together improved agent success rates by up to 7.6% relative to baselines and achieved a 2-7x prefill speedup.
Key insights
- Conversational inertia is a distinct failure mode in multi-turn LLM agents, mechanistically characterized by strong diagonal attention patterns where LLMs disproportionately attend to corresponding positional tokens in their own prior outputs.
- A fundamental tension exists where longer contexts, while providing richer environmental feedback for exploitation, simultaneously amplify conversational inertia, undermining exploration.
- It is possible to calibrate model preferences to resist inertia without explicit environment rewards by exploiting the intrinsic observation that shorter contexts inherently exhibit weaker self-imitation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.