Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Trust Me, I'm an Expert: Decoding and Steering Authority Bias in Large Language Models

Priyanka Mary Mammen, Emil Joswin, Shankar Venkitachalam

TL;DR

This work addresses authority bias in LLM reasoning by systematically varying endorsement source expertise across four datasets (math, law, medicine) and testing 11 models with baseline, correct endorsement, and incorrect endorsement prompts. It finds a hierarchical pattern: higher-expertise endorsements improve accurate outcomes but increase the likelihood and confidence of incorrect decisions when endorsements are misleading, revealing a mechanistic bias encoded in the models' internal representations. The authors demonstrate mitigation by extracting a steering vector representing 'expertise' from the residual stream and subtracting it during inference, reducing bias across models. These findings illuminate a concrete vulnerability in current LLM reasoning and offer a practical, manipulable defense to improve reliability in critical domains.

Abstract

Prior research demonstrates that performance of language models on reasoning tasks can be influenced by suggestions, hints and endorsements. However, the influence of endorsement source credibility remains underexplored. We investigate whether language models exhibit systematic bias based on the perceived expertise of the provider of the endorsement. Across 4 datasets spanning mathematical, legal, and medical reasoning, we evaluate 11 models using personas representing four expertise levels per domain. Our results reveal that models are increasingly susceptible to incorrect/misleading endorsements as source expertise increases, with higher-authority sources inducing not only accuracy degradation but also increased confidence in wrong answers. We also show that this authority bias is mechanistically encoded within the model and a model can be steered away from the bias, thereby improving its performance even when an expert gives a misleading endorsement.

Trust Me, I'm an Expert: Decoding and Steering Authority Bias in Large Language Models

TL;DR

This work addresses authority bias in LLM reasoning by systematically varying endorsement source expertise across four datasets (math, law, medicine) and testing 11 models with baseline, correct endorsement, and incorrect endorsement prompts. It finds a hierarchical pattern: higher-expertise endorsements improve accurate outcomes but increase the likelihood and confidence of incorrect decisions when endorsements are misleading, revealing a mechanistic bias encoded in the models' internal representations. The authors demonstrate mitigation by extracting a steering vector representing 'expertise' from the residual stream and subtracting it during inference, reducing bias across models. These findings illuminate a concrete vulnerability in current LLM reasoning and offer a practical, manipulable defense to improve reliability in critical domains.

Abstract

Prior research demonstrates that performance of language models on reasoning tasks can be influenced by suggestions, hints and endorsements. However, the influence of endorsement source credibility remains underexplored. We investigate whether language models exhibit systematic bias based on the perceived expertise of the provider of the endorsement. Across 4 datasets spanning mathematical, legal, and medical reasoning, we evaluate 11 models using personas representing four expertise levels per domain. Our results reveal that models are increasingly susceptible to incorrect/misleading endorsements as source expertise increases, with higher-authority sources inducing not only accuracy degradation but also increased confidence in wrong answers. We also show that this authority bias is mechanistically encoded within the model and a model can be steered away from the bias, thereby improving its performance even when an expert gives a misleading endorsement.
Paper Structure (16 sections, 3 equations, 3 figures, 1 table)

This paper contains 16 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Experiments
  4. Datasets and prompts
  5. Models
  6. Evaluation Metrics
  7. Results and Discussion
  8. Measuring the impact of expertise levels
  9. Additional analysis using steering vectors
  10. Related Work
  11. Conclusions
  12. Limitations
  13. Ethical Considerations
  14. Example Appendix
  15. Finding steering vector
  16. ...and 1 more sections

Figures (3)

  • Figure 1: We design our experiment across various domains (math reasoning, medical, and legal MCQs) where multiple personas in increasing order of expertise in their respective domains provide correct and misleading endorsements.
  • Figure 2: Model accuracy for incorrect endorsement
  • Figure 3: Model accuracy for incorrect endorsement after steering away from expertise persona.