Measuring the Reliability of Causal Probing Methods: Tradeoffs, Limitations, and the Plight of Nullifying Interventions
Keywords: interpretability, probing, causal probing, interventions, mechanistic interpretability, language models
TL;DR: We introduce an empirical analysis framework to evaluate the reliability of causal probing interventions for interpreting foundation models.
Abstract: Causal probing aims to analyze large language models (or other foundation models) by examining how modifying their representation of various latent properties using interventions derived from probing classifiers impacts their outputs. Recent works have cast doubt on the theoretical basis of several leading causal probing intervention methods, but it has been unclear how to systematically evaluate the effectiveness of probing interventions in practice. To address this, we formally define and quantify two key causal probing desiderata: completeness (how thoroughly the representation of the target property has been transformed) and selectivity (how little other properties have been impacted). We introduce an empirical analysis framework to measure and evaluate completeness and selectivity, allowing us to make the first direct comparisons of the reliability of different families of causal probing methods (e.g., linear vs. nonlinear or counterfactual vs. nullifying interventions). Our experimental analysis shows that: (1) there is an inherent tradeoff between completeness and selectivity, (2) no leading probing method is able to consistently satisfy both criteria at once, and (3) across the board, nullifying interventions are far less complete than counterfactual interventions, which suggests that nullifying methods may not be an effective approach to causal probing.
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: 12864
Loading