Answer-Centric or Reasoning-Driven? Uncovering the Latent Memory Anchor in LLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wu, Yang, Zhang, Yifan, Wang, Yiwei, Cai, Yujun, Wu, Yurong, Wang, Yuran, Xu, Ning, Cheng, Jian
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911018583588864
author Wu, Yang
Zhang, Yifan
Wang, Yiwei
Cai, Yujun
Wu, Yurong
Wang, Yuran
Xu, Ning
Cheng, Jian
author_facet Wu, Yang
Zhang, Yifan
Wang, Yiwei
Cai, Yujun
Wu, Yurong
Wang, Yuran
Xu, Ning
Cheng, Jian
contents While Large Language Models (LLMs) demonstrate impressive reasoning capabilities, growing evidence suggests much of their success stems from memorized answer-reasoning patterns rather than genuine inference. In this work, we investigate a central question: are LLMs primarily anchored to final answers or to the textual pattern of reasoning chains? We propose a five-level answer-visibility prompt framework that systematically manipulates answer cues and probes model behavior through indirect, behavioral analysis. Experiments across state-of-the-art LLMs reveal a strong and consistent reliance on explicit answers. The performance drops by 26.90\% when answer cues are masked, even with complete reasoning chains. These findings suggest that much of the reasoning exhibited by LLMs may reflect post-hoc rationalization rather than true inference, calling into question their inferential depth. Our study uncovers the answer-anchoring phenomenon with rigorous empirical validation and underscores the need for a more nuanced understanding of what constitutes reasoning in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2506_17630
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Answer-Centric or Reasoning-Driven? Uncovering the Latent Memory Anchor in LLMs
Wu, Yang
Zhang, Yifan
Wang, Yiwei
Cai, Yujun
Wu, Yurong
Wang, Yuran
Xu, Ning
Cheng, Jian
Computation and Language
While Large Language Models (LLMs) demonstrate impressive reasoning capabilities, growing evidence suggests much of their success stems from memorized answer-reasoning patterns rather than genuine inference. In this work, we investigate a central question: are LLMs primarily anchored to final answers or to the textual pattern of reasoning chains? We propose a five-level answer-visibility prompt framework that systematically manipulates answer cues and probes model behavior through indirect, behavioral analysis. Experiments across state-of-the-art LLMs reveal a strong and consistent reliance on explicit answers. The performance drops by 26.90\% when answer cues are masked, even with complete reasoning chains. These findings suggest that much of the reasoning exhibited by LLMs may reflect post-hoc rationalization rather than true inference, calling into question their inferential depth. Our study uncovers the answer-anchoring phenomenon with rigorous empirical validation and underscores the need for a more nuanced understanding of what constitutes reasoning in LLMs.
title Answer-Centric or Reasoning-Driven? Uncovering the Latent Memory Anchor in LLMs
topic Computation and Language
url https://arxiv.org/abs/2506.17630