Unveiling Language Skills under Circuits
Keywords: Interpretability of Language Models
TL;DR: We extract the language skill paths and show them in a circuit graph based on the circuit analysis and causal effect estimation
Abstract: Circuit decomposition and counterfactual-based pruning have become the cornerstone framework for mechanism interpretability. However, the unfaithfulness to the output due to cumulative bias in the pruning process hinders more complex and detailed mechanism exploration. To address this, we propose a novel circuit discovery framework that faithfully identifies circuit graphs. This framework contains three steps: firstly, the language model is decomposed into a fully linear graph consisting of disentangled ``memory circuits"; secondly, greedy search is adopted to prune while ensuring output faithfulness; finally, we adopt causal analysis on the pruned circuit graph to identify salient circuit graph, estimated by counterfactuals and interventions. Our framework facilitates the discovery of complete circuit graphs and dissection of more complex mechanisms. To demonstrate this, we explored three generic language skills (Previous Token Skill, Induction Skill and In-Context Learning Skill). Using the circuit graphs discovered through our framework, we identify the complete skill paths of these skills.
Our experiments on various datasets confirm the correspondence between our identified skill paths and language skills, and validate three longstanding hypotheses: 1) Language skills are identifiable through circuit dissection; 2) Simple language skills reside in shallow layers, whereas complex language skills are found in deeper layers; 3) Complex language skills are formed on top of simpler language skills.
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: 2109
Loading