Localizing Anchoring Pathways in Language Models

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

Research investigating anchoring effects in language models reveals how irrelevant numerical prompts shift LLM judgments. Using a controlled multiple-choice task and a logit-difference metric, researchers applied attribution-based circuit localization, specifically EAP-IG, to 7B-8B Qwen and Llama base and instruction-tuned models. The study found that edge-level localization methods more faithfully recover anchor-sensitive signals than node-level methods. Within a model, low- and high-anchor circuits exhibit strong transferability and shared pathway structure, with approximately two-thirds of their top 5% edges overlapping. However, transferability across base and instruction-tuned variants is less consistent, suggesting that post-training significantly alters the most relevant pathways. Qwen models show attribution in mid-to-late layers, while Llama models concentrate it earlier, indicating family-specific differences in localization patterns.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on LLM robustness, understanding anchoring bias requires granular, edge-level circuit analysis. You should prioritize methods like EAP-IG to identify specific pathways, as low and high anchor effects share significant internal structure. Be aware that instruction tuning can alter these critical pathways, necessitating re-evaluation of bias mitigation strategies after post-training to ensure continued effectiveness.

Key insights

Anchoring bias in LLMs is localized to sparse, shared internal pathways, but these pathways shift with instruction tuning.

Principles

• Edge-level attribution is more faithful for localizing LLM behavior.
• Low and high anchor effects share substantial circuit structure.
• Instruction tuning alters critical sparse edges and pathway importance.

Method

Circuit localization uses attribution patching (EAP-IG) on a logit-difference metric comparing correct vs. anchor answer options in a multiple-choice task to identify anchor-sensitive pathways.

In practice

• Prioritize EAP-IG for precise mechanistic interpretability studies.
• Re-evaluate bias mitigation strategies post-instruction tuning.
• Consider model family for layer-wise attribution patterns.

Topics

• Language Models
• Anchoring Bias
• Mechanistic Interpretability
• Circuit Localization
• Attribution Patching
• Instruction Tuning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Localizing Anchoring Pathways in Language Models — Irrelevant numerical anchors in prompts shift LM judgments via specific, localizable internal pathways.
• Weight Patching: Toward Source-Level Mechanistic Localization in LLMs — Weight Patching localizes LLM behaviors to specific parameters by transferring weights between specialized and base models.
• Modeling Pathology-Like Behavioral Patterns in Language Models Through Behavioral Fine-Tuning — Consistent behavioral fine-tuning can induce stable, pathology-like biases in LLM generative distributions and action selection.
• More Thinking, More Bias: Length-Driven Position Bias in Reasoning Models — Longer reasoning trajectories in LLMs increase position bias, a critical factor for evaluation reliability.
• Routing-Aligned Fine-Tuning for Multilingual Downstream Tasks in Mixture-of-Experts Models — MoE model multilingual performance improves by aligning target-language routing with English expert activation in middle layers.
• Ful LLM / Deep Learning roadmap. — Transformers leverage attention mechanisms and positional encodings to process sequences bidirectionally or causally.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.