Technical Summary: Who Endorsed It? Measuring Authority Bias Across Expertise Levels in Language Models

Problem Statement

While prior research has established that Large Language Models (LLMs) are susceptible to suggestions, hints, and endorsements, the specific influence of source credibility remains underexplored. Existing studies on bias often focus on demographic attributes (gender, race) or structural biases (position, length). However, it is unclear how LLMs process endorsements based on the perceived professional status of the source. Specifically, the paper investigates whether models exhibit systematic authority bias, where they defer to high-status personas even when those personas provide incorrect information, potentially overriding their own internal reasoning capabilities.

Methodology

The authors conducted a comprehensive evaluation across four datasets spanning mathematical (AQUA-RAT), legal (LEXam), and medical (MedMCQA, MedQA) reasoning tasks.

Experimental Design

1. Personas and Expertise Gradient: Instead of a binary "expert vs. non-expert" setup, the study utilized a four-tier expertise hierarchy for each domain:
   • Science: High Schooler → Undergrad → Grad Student → Professor.
   • Medicine: First-Year Student → Third-Year Student → Chief Resident → Board-Certified Physician.
   • Law: Undergrad Student → Third-Year Student → Law Clerk → Senior Legal Counsel.
2. Prompt Variants: For each Multiple Choice Question (MCQ), nine prompt variants were generated:
   • Baseline: No endorsement.
   • Correct Endorsement: A persona endorses the correct answer.
   • Incorrect Endorsement: A persona endorses a wrong answer.
   • The endorsement format was minimal (e.g., "A [Persona] thinks the answer is [X]") to isolate the effect of the source's identity from linguistic style or justification.
3. Models Evaluated: The study tested 11 models, including both Reasoning Models (e.g., DeepSeek-R1, Phi-4-Reasoning, Qwen3-4B-Thinking) and Non-Reasoning Models (e.g., LLaMA-3.1, Gemma-2/3, Mistral-7B).
4. Metrics:
   • Delta Accuracy ($\Delta$Acc): Deviation from baseline accuracy.
   • Delta Entropy ($\Delta$H): Change in confidence (lower entropy = higher confidence).
   • Robustness Rate (RR): The rate at which the model's output remains unchanged despite the endorsement.

Mechanistic Intervention (Steering Vector)

To determine if authority bias is encoded internally rather than being a surface-level keyword trigger, the authors constructed a steering vector.

• Construction: They generated a dataset of factually correct answers written in the linguistic styles of the highest (e.g., Professor) and lowest (e.g., High Schooler) expertise tiers, without any persona labels.
• Vector Calculation: The steering vector ($v_{auth}$) was computed as the mean difference in residual stream activations between the high-expertise and low-expertise styles.
• Intervention: This vector was subtracted from the model's activations during inference to test if susceptibility to misleading endorsements could be reduced.

Key Results

1. Hierarchical Susceptibility

The results reveal a clear monotonic relationship between source expertise and model behavior:

• Correct Endorsements: Accuracy gains increase as the endorser's expertise level rises.
• Incorrect Endorsements: Accuracy degradation and confidence in wrong answers increase significantly with higher expertise levels.
• Reasoning Models: Contrary to expectations, models with chain-of-thought capabilities (e.g., DeepSeek-R1) are not immune. They exhibit substantial accuracy degradation and increased confidence in errors when misled by high-authority sources (e.g., a Board-Certified Physician).

2. Confidence in Errors

High-expertise incorrect endorsements induce confident errors. For instance, when a Board-Certified Physician endorsed a wrong answer on MedQA, DeepSeek-R1 showed a significant drop in entropy ($\Delta H = -0.261$), indicating the model became more certain of the incorrect answer.

3. Domain Variations

• Mathematical Reasoning: Showed the lowest robustness rates, suggesting high susceptibility despite the objective nature of the tasks.
• Medical Reasoning: Showed higher resistance to changing answers, possibly reflecting domain-specific training regarding medical caution.

4. Mechanistic Findings

The steering vector experiments demonstrated that authority bias is mechanistically encoded within the model's residual stream.

• Subtracting the "expertise" vector ($v_{auth}$) during inference consistently recovered model accuracy toward baseline levels when faced with misleading high-authority endorsements.
• This confirms that the bias is tied to a robust internal representation of expertise that generalizes across both explicit social identity markers (persona labels) and implicit linguistic styles.

Contributions and Significance

The paper makes three primary contributions:

1. Granular Analysis of Authority Bias: It moves beyond binary authority concepts to demonstrate a hierarchical pattern where model susceptibility scales with the perceived rank of the endorser.
2. Vulnerability of Reasoning Models: It challenges the assumption that reasoning-enhanced models are robust against social manipulation, showing they can be derailed by high-status incorrect endorsements.
3. Mechanistic Mitigation: It provides evidence that authority bias is a linearly encoded feature in the residual stream. By subtracting the corresponding steering vector, the model can be steered away from the bias, improving performance even when an expert provides misleading information.

Significance and Limitations

The authors argue that this work reveals a critical vulnerability: current LLMs prioritize source credibility over semantic correctness. This bias can be exploited in adversarial contexts (e.g., social engineering attacks) but can also be mitigated through inference-time activation steering without weight modification.

The paper acknowledges limitations, including:

• Experiments were limited to open-source models up to 32B parameters; frontier models may behave differently.
• The study covered only four domains (math, law, medicine).
• The endorsement format was simplified (single statements) and did not account for sophisticated real-world misinformation.
• The layer-wise analysis of the steering vector is preliminary, identifying effectiveness in middle layers but lacking a systematic ablation or Sparse Autoencoder (SAE) characterization.