Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·19h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Localizing Anchoring Pathways in Language Models
TL;DR
Language models can be influenced by irrelevant numerical anchors, affecting their reasoning. Researchers developed a logit-difference metric to analyze anchoring effects and used circuit localization techniques on Qwen and Llama models.
✦ Why It Matters
Engineers can leverage these insights to improve model robustness against anchoring biases in numerical tasks.
Key Takeaways
- Investigate how irrelevant numbers affect model judgments.
- Utilize logit-difference metrics for analyzing anchoring.
- Apply edge-level methods for better signal recovery.
- Refine prompts to reduce anchoring bias in outputs.
How It Works
The study employs a logit-difference metric to quantify the influence of anchors on model outputs. By comparing the logits of correct answers to those influenced by irrelevant numbers, researchers can track anchoring effects.
Circuit localization techniques reveal that specific pathways within the model are responsible for carrying these signals, with edge-level methods proving more effective than node-level methods in capturing the nuances of anchoring.