AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency

Max Henning Höth, Kristian Kersting, Björn Deiseroth, Letitia Parcalabescu

Track: tiny / short paper (up to 4 pages)

Keywords: reasoning, faithfulness, saliency, large language models, reasoning trace, chain-of-thought, attention manipulation, reinforcement learning, GRPO, interpretability, attention attribution

TL;DR: Using a differentiable attention mask, we reward reasoning tokens in RL training that causally influence the final answer, to generate more faithful chain-of-thought reasoning.

Abstract: Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks. Yet ensuring that the reasoning trace both contributes to and faithfully reflects the processes underlying the model’s final answer, rather than merely accompanying it, remains challenging. We introduce AtManRL, a method that leverages differentiable attention manipulation to learn more faithful reasoning through reinforcement learning. By training an additive attention mask that identifies tokens in the CoT crucial for producing correct answers, we derive a saliency reward signal that encourages the model to generate reasoning traces that genuinely influence its final predictions. We integrate this saliency reward with outcome-based rewards within the GRPO framework to jointly optimize for correctness and interpretability. Experiments on GSM8K with Llama-3.2-3B-Instruct demonstrate that our approach can identify influential reasoning tokens and enabletraining more transparent reasoning models.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 89