Cumulative turn-level reasoning uncertainty as the source of trajectory-level uncertainty

Determine whether trajectory-level uncertainty in Reasoning-and-Acting processes frequently stems from cumulative turn-level reasoning uncertainty induced by large language models, including divergent reasoning chains and their resulting actions.

Background

The paper studies uncertainty in multi-turn Reasoning-and-Acting trajectories generated by LLM agents. Identical tasks can produce divergent trajectories, and the paper proposes that this trajectory-level uncertainty is accumulated from uncertainty at individual turns, particularly from branching reasoning chains and their associated textual actions.

The proposed GRUET method operationalizes this hypothesis by representing sampled reasoning branches as a directed acyclic graph and aggregating turn-level uncertainty scores. The experiments provide empirical evidence for the method’s usefulness, but the causal conjecture itself is not established as a general theoretical result.