Steering Vectors Reduce AI Evaluators' Bias Toward Their Own Outputs

When large language models serve as automated evaluators, they frequently show favoritism toward outputs from their own kind—a phenomenon that undermines fair assessment in AI comparison workflows and model selection pipelines. Researchers have now developed a method to suppress this "self-preference bias" using steering vectors applied at inference time, without requiring model retraining.

The Problem: Self-Preference Bias in AI Judges#

LLMs increasingly evaluate and rank outputs from other models, a task critical for preference fine-tuning and routing systems. However, they exhibit a clear tendency to rate their own responses more favorably, even when objectively less capable alternatives exist. This bias threatens the integrity of any evaluation system that relies on LLM judgments.

The Solution: Activation Steering#

Researchers created a curated dataset distinguishing between justified instances of self-preference (where a model's output genuinely deserves higher marks) and unjustified ones (where bias skews the evaluation). Using this data, they developed steering vectors through two approaches:

• Contrastive Activation Addition (CAA): A method that amplifies desired directions in neural activations while suppressing problematic ones
• Optimization-based steering: A computational approach to identify and apply corrective vectors

These lightweight interventions operate at inference time, making them practical for deployment without retraining expensive foundation models.

Results and Limitations#

The steering vectors achieved substantial reductions—up to 97%—in unjustified self-preference bias, substantially eclipsing results from simpler alternatives like prompt engineering and direct preference optimization. However, the approach revealed significant shortcomings: steering vectors proved unstable when applied to legitimate self-preference and unbiased model agreement, suggesting that self-preference operates across multiple or nonlinear neural pathways. This inconsistency points to the need for more sophisticated, robust safeguards for deploying LLMs as impartial judges.