StepER: Step-wise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented LM

ACL ARR 2025 February Submission5210 Authors

16 Feb 2025 (modified: 09 May 2025)ACL ARR 2025 February SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Abstract: To answer complex real-world questions, it is crucial to retrieve and integrate relevant information step-by-step to generate well-grounded responses. However, existing methods struggle to effectively distill step-specific reasoning abilities, as they do not account for the varying amount of information accessed at each reasoning step. To address this limitation, we propose Step-wise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented LM (StepER). StepER leverages step-wise datasets and reasoning difficulty-aware training to enhance reasoning abilities essential for multi-step retrieval-augmented LM. Moreover, StepER is adaptable to various multi-step retrieval-augmented LM frameworks, including reasoning path-based retrieval and question decomposition-based approaches. Extensive experiments demonstrate that StepER outperforms existing methods on multi-hop QA datasets, with an 8B model achieving performance on par with a 70B teacher model.
Paper Type: Long
Research Area: Question Answering
Research Area Keywords: retrieval-augmented generation, retrieval-augmented models, multihop QA, distillation
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 5210
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview