Removing Spurious Correlation from Neural Network Interpretations
Not on the map. Why?
Abstract (arXiv)
The existing algorithms for identification of neurons responsible for undesired and harmful behaviors do not consider the effects of confounders such as topic of the conversation. In this work, we show that confounders can create spurious correlations and propose a new causal mediation approach that controls the impact of the topic. In experiments with two large language models, we study the localization hypothesis and show that adjusting for the effect of conversation topic, toxicity becomes less localized.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 12 of its 37 references that are in the map. First posted on arXiv 3 Dec 2024.
Builds on
12 of 37 references.
References · 12
- 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,372
- 2022Locating and Editing Factual Associations in GPTcited by 2,691
- 2022BLOOM: A 176B-Parameter Open-Access Multilingual Language Modelcited by 1,984
- 2024Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phonecited by 1,917
- 2022In-context Learning and Induction Headscited by 964
- 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
- 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
- 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
- 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 359
- 2024Applied Causal Inference Powered by ML and AIcited by 30
- 2021End-to-End Balancing for Causal Continuous Treatment-Effect Estimationcited by 9
- 2024Probabilities of Causation for Continuous and Vector Variablescited by 8
Built on by
Not cited yet.