Reproduce “UProp: Investigating the Uncertainty Propagation of LLMs in Multi-Step Agentic Decision-Making” 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

Jinhao Duan, James Diffenderfer, Sandeep Madireddy, Tianlong Chen, Bhavya Kailkhura, Kaidi Xu

UProp introduces an information-theoretic framework and estimator to quantify how uncertainty propagates in multi-step LLM agentic decision-making, decomposing total uncertainty into intrinsic and extrinsic components. The method consistently outperformed existing single-turn Uncertainty Quantification baselines on multi-step benchmarks like AgentBench-OS and StrategyQA, improving AUROC by up to 11% and enhancing selective prediction.

Key insights

  • Extrinsic uncertainty, which reflects the variability inherited from previous decisions, is a critical component for accurately quantifying total uncertainty in multi-step LLM agentic systems.
  • The contribution of extrinsic uncertainty to the total uncertainty increasingly dominates over intrinsic uncertainty as an LLM's decision process extends to later steps in a trajectory.
  • UProp demonstrates sampling efficiency, maintaining its performance advantage even with fewer trajectory and per-step samples, making multi-step uncertainty estimation practical.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.