Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLM Evaluators

Name: Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLM Evaluators
Uploaded: 2026-10-08T03:01:28.766315+00:00
Duration: 337 s
Description: Researchers developed lightweight steering vectors to reduce self-preference bias in LLM evaluators by up to 97% at inference time. By analyzing activation differences between biased and unbiased outputs, they found these vectors effectively correct illegitimate self-preference, though they struggled with stability across other preference types.

Researchers developed lightweight steering vectors to reduce self-preference bias in LLM evaluators by up to 97% at inference time. By analyzing activation differences between biased and unbiased outputs, they found these vectors effectively correct illegitimate self-preference, though they struggled with stability across other preference types.

Open the full episode