Reproduce “Uncertainty-Aware Clarification in LLM Agents with Information Gain” 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
Mengyi Deng, Zhiwei Li, Xin Li, Tingyu Zhu, Ying Zhao, Zhijiang Guo, Wei Wang
The authors present an amortized clarification policy for tool-using language-model agents. It asks targeted questions after tool feedback, rewarding a length-normalized prior-to-posterior shift in teacher-forced likelihood of the ground-truth goal. Across five task-agent backbones in controlled τ-Bench tasks, a learned 1.7B clarifier raises Pass@1 success from 23.6% to 27.3%, including 29.9% versus 24.5% on Airline. The evidence uses simulated users and task-defined goals, not real-user interactions or unsupervised goals.
Key insights
- An information-theoretic reward for clarification enables LLM agents to learn targeted questioning strategies that effectively reduce uncertainty about latent user goals.
- Specialized, smaller LLM clarifiers, trained with an information gain objective, can match or surpass larger general-purpose LLMs in ambiguity resolution while being significantly more interaction-efficient.
- There is an optimal clarification frequency; excessive clarification can lead to context pollution and redundant interactions, diminishing task success and efficiency gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.