Toward Mechanistic Localization of Behavioral Effects in Language Models: Anchoring as a Case Study

Mechanistic interpretability aims to understand how model behavior arises from internal computation. We study this question through anchoring, a cognitive bias in which irrelevant numerical context shifts subsequent judgments. Using Qwen2.5-7B and Llama-3.1-8B base and instruction-tuned models, we localize the internal pathways associated with anchoring and compare component-level and connection-level methods. We find that circuits based on connections between components capture the effect more effectively than circuits based on individual components, while low- and high-anchor conditions rely on substantially shared internal structure. We also find that model family and instruction tuning affect which pathways are most important. These results motivate broader questions about whether behavioral effects in language models can be localized and selectively modified.