Localizing Anchoring Pathways in Language Models

In three linesStudy of internal anchoring mechanisms in language models. Researchers localize circuits responsible for anchoring bias (where irrelevant numbers influence answers) in Qwen and Llama 7B-8B models. Edge-level attribution methods recover this signal more faithfully than node-level methods.

Summary generated by Claude — human-verified