Learning to Reason Faithfully through Step-Level Faithfulness Maximization

ABSTRACT

Large Language Models frequently produce correct final answers based on flawed or unfaithful intermediate reasoning steps. This paper proposes Step-Level Faithfulness Maximization, a training paradigm that enforces strict logical alignment at every node of the reasoning chain. By penalizing correct answers derived from hallucinated logic, the framework ensures the model's explanations genuinely reflect its internal decision-making process.