Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing
2026-04-09 • Artificial Intelligence
Artificial IntelligenceComputation and Language
AI summaryⓘ
The authors study how large language models can sometimes build on wrong or unsupported ideas while solving problems, which leads to mistakes over time. They propose a new method called SAVeR that checks and fixes these internal ideas before the model acts, making the reasoning more trustworthy. This method creates different possible ideas, finds errors through checks, and fixes them with minimal changes. Their tests on multiple datasets show SAVeR improves reasoning accuracy without hurting overall task performance.
large language modelsreasoning trajectoriesfaithfulnessinternal beliefsverificationadversarial auditingconstraint-guided interventionagentic systemsbenchmark datasets
Authors
Wenhao Yuan, Chenchen Lin, Jian Chen, Jinfeng Xu, Xuehe Wang, Edith Cheuk Han Ngai
Abstract
In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose \textbf{S}elf-\textbf{A}udited \textbf{Ve}rified \textbf{R}easoning (\textsc{SAVeR}), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.