Localizing Anchoring Pathways in Language Models

Researchers have mapped how irrelevant numbers slipped into a prompt can systematically skew the numerical judgments of large language models, identifying the internal circuits that carry this anchoring signal in 7B–8B parameter Qwen and Llama models, according to a paper submitted to arXiv on 11 June 2026 [1][2]. The study, posted on the open-access preprint repository arXiv, used a controlled multiple-choice setup with shared answer options to isolate the effect [2]. The authors defined a logit-difference metric that compares the correct answer option with the option corresponding to the anchor number, and they validated that this metric tracks behavioral anchoring [2]. arXiv, which was founded in 1991 and now receives roughly 24,000 submissions per month, hosts papers across physics, computer science, mathematics and related fields after moderation but before peer review [11]. Using attribution-based circuit localization, the team found that edge-level methods recovered the anchoring signal more faithfully than node-level methods [2]. Low- and high-anchor circuits transferred strongly within a given model, suggesting a shared pathway structure across anchor direction [2]. However, transfer between base and instruction-tuned variants of the same model was sparse and less reliable, indicating that post-training changes which pathways matter most [2]. The work provides a mechanistic account of how anchoring-related decision signals are carried inside language models [2]. The paper appeared with the standard arXivLabs framework on its abstract page, a system that allows community collaborators to develop and share experimental tools directly on the site [10]. arXivLabs features, which include bibliographic explorers, code-and-data linkers, and recommender systems, operate under guidelines that require partners to share arXiv’s values of openness, community, excellence and user-data privacy [10][9]. The repository itself passed the two-million-article milestone by the end of 2021 and continues to serve as the primary distribution channel for preprints in many quantitative disciplines [11].