A Consistent Pattern for Identifying Decisive Code Snippets for LLM-Based Code Inference
Keywords: Programming Languages; interpretability; Knowledge editing
Abstract: Which parts of pre-target input are most influential for next-token prediction in the context of programming languages? In this paper, we present evidence that code snippets at specific locations in pre-target inputs play a decisive role in large language model (LLM) inference, and these snippets exhibit a consistent pattern. Firstly, we introduce a novel causal tracing method to identify tokens, so-called high-information tokens, that significantly contribute to next-token prediction. Building on this, we propose a multi-phase causal tracing process to analyze the importance distribution of high-information tokens, revealing a consistent pattern, named the Important Position Rule (IPR). To further validate this hypothesis, we assess the role of IPR across various LLMs, languages, and tasks. Our extensive evaluations for code translation, code correction and code completion tasks (Java, Python, C++) on models CodeLlama-7b/13b/34b-Instruct and GPT-3.5/4-turbo, confirm this hypothesis. Furthermore, we observe that IPR exhibits structural and semantic properties similar to the $\langle \text{subject}, \text{relation}, \text{object} \rangle$ paradigm in natural language. Leveraging this insight, we successfully combine IPR with the knowledge editing method ROME in order to repair translation errors, achieving a correction rate of 62.73% to 75.31%. To our knowledge, this is the first application of knowledge editing in the context of programming languages.
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: 3819
Loading