Local Causal Attribution of Chain-of-Thought Reasoning

The paper introduces AttriCoT, a black-box algorithm designed for local causal attribution of chain-of-thought reasoning in language models. It constructs a structural causal model to analyze the importance of individual components within a reasoning trace, requiring $O(U)$ forward passes, where $U$ is the number of units. The evaluation across five datasets and four reasoning models demonstrates that AttriCoT provides more accurate attributions compared to existing methods, highlighting significant variations in thought structures across different models and domains, which is crucial for enhancing transparency and safety in AI systems.