Deciphering and Enhancing Commonsense Reasoning in LLMs from the Perspective of Intrinsic Factual Knowledge Retrieval

ICLR 2025 Conference Submission660 Authors

14 Sept 2024 (modified: 26 Nov 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Model interpretability, Chain-of-Thought, Large Language Models
Abstract: Commonsense reasoning in large language models (LLMs) bridges the gap to physical world, thus allowing them to think and behave more like humans. Previous research has shown that LLMs acquire the underlying factual knowledge from extensive training corpora and store it within their parameters. However, how LLMs apply this knowledge during the inference phase remains unclear. This lack of transparency makes it difficult to determine whether shortcomings in LLMs are due to a lack of factual knowledge or insufficient reasoning capabilities. In this work, we aim to decipher the commonsense reasoning process into human-understandable steps. By interpreting the hidden states in different transformer layers and token positions, we uncover a specific mechanism by which LLMs execute reasoning. Our extensive experiments indicate: 1) both attention head and multi-layer perceptron (MLP) contribute to the generation of factual knowledge from different perspective. 2) The process of commonsense reasoning in LLMs involves a clear sequence of knowledge augmentation, knowledge retrieval and answer generation, akin to retrieval-augmented generation. Building on these findings, we have discovered that LLMs often contain relevant facutal knowledge but fail to retrieve the correct knowledge at top. To address this issure, we selectively fine-tuned the key heads and MLPs, resulting in notably improvements in reasoning performance in both in-domain and out-of-domain settings.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 660
Loading