Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Yuhui, Li, Changjiang, Chen, Guangke, Liang, Jiacheng, Wang, Ting
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914360463458304
author Wang, Yuhui
Li, Changjiang
Chen, Guangke
Liang, Jiacheng
Wang, Ting
author_facet Wang, Yuhui
Li, Changjiang
Chen, Guangke
Liang, Jiacheng
Wang, Ting
contents Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities. The code is available: https://github.com/ZJUWYH/FARL.
format Preprint
id arxiv_https___arxiv_org_abs_2509_24156
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models
Wang, Yuhui
Li, Changjiang
Chen, Guangke
Liang, Jiacheng
Wang, Ting
Artificial Intelligence
Computation and Language
Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities. The code is available: https://github.com/ZJUWYH/FARL.
title Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2509.24156