Factor-Graph Based Calibration For LLM Agents In Complex Reasoning

ACL ARR 2026 January Submission1491 Authors

Keywords: Calibration, LLM Agents, Reinforcement Learning

Abstract: LLM Agents achieve strong performance on complex reasoning tasks but remain poorly calibrated, often assigning unjustified confidence to incorrect intermediate steps. Most existing confidence estimators operate primarily at the answer level, and therefore fail to model how uncertainty evolves and propagates across intermediate reasoning steps. Factor graphs provide a natural fit for this setting: they decompose global trajectory uncertainty into local dependency factors and use message passing to propagate and fuse evidence across steps, yielding principled step-wise belief updates. Motivated by this alignment, we propose the first framework that converts an agent’s multi-step trajectory into a factor graph for step-wise confidence modeling, explicitly capturing step-to-step dependencies so that uncertainty can be propagated and calibrated throughout the entire reasoning process. Building on this factor-graph view, we further leverage reinforcement learning to align the model’s verbal confidence with calibrated step-level uncertainty estimates. Experiments on six QA benchmarks demonstrate improved calibration and stronger overall performance.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: Interpretability and Analysis of Models for NLP, Special Theme Track,AI / LLM Agents

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1491