Agent uncertainty needs to track whole tool-use trajectories

A new arXiv paper argues that uncertainty estimates for language models need to change when the model is acting as an agent.

Most uncertainty quantification methods attach confidence to a single generated answer. Agents are different: they ask questions, call tools, update state, and make intermediate choices that can shape the final result. The paper evaluates whether three common uncertainty families transfer to this multi-turn setting.

Across five LLMs and four tool-use datasets from BFCL-v4 and tau-squared-bench, the authors test white-box scores based on action-token probabilities, black-box consistency scores from resampled trajectories, and reflexive scores based on model self-assessment. They find that transfer is useful but uneven. Token probability methods depend heavily on how scores are aggregated across turns, while self-consistency and trajectory-level comparisons often perform strongly.

The practical lesson is that agent confidence should not be treated as a property of the final sentence alone. Reliability depends on the path the system took to get there.