mlsanityData to 14 Oct 2026

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

arXiv 2509.03647 ↗First posted 3 Sep 2025cs.CL cs.AI cs.LGcited by 3

On the map: LLM alignment / RLHF / preference optimisation → Human values and AI feedback alignment

also in 1 more
Abstract (arXiv)

Large language models (LLMs) increasingly serve as automated evaluators, yet they suffer from "self-preference bias": a tendency to favor their own outputs over those of other models. This bias undermines fairness and reliability in evaluation pipelines, particularly for tasks like preference tuning and model routing. We investigate whether lightweight steering vectors can mitigate this problem at inference time without retraining. We introduce a curated dataset that distinguishes self-preference bias into justified examples of self-preference and unjustified examples of self-preference, and we construct steering vectors using two methods: Contrastive Activation Addition (CAA) and an optimization-based approach. Our results show that steering vectors can reduce unjustified self-preference bias by up to 97%, substantially outperforming prompting and direct preference optimization baselines. Yet steering vectors are unstable on legitimate self-preference and unbiased agreement, implying self-preference spans multiple or nonlinear directions. This underscores both their promise and limits as safeguards for LLM-as-judges and motivates more robust interventions.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Human values and AI feedback alignment, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 22 of its 25 references that are in the map. First posted on arXiv 3 Sep 2025.

Builds on

22 of 25 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025Do LLM Evaluators Prefer Themselves for a Reason?4 Apr
  2. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization28 May
  3. 2024LLM Evaluators Recognize and Favor Their Own Generations15 Apr

Other references · 19

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,626
  3. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  4. 2024A Survey on LLM-as-a-Judgecited by 1,423
  5. 2018Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarizationcited by 1,317
  6. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  7. 2024Phi-4 Technical Reportcited by 711
  8. 2024Justice or Prejudice? Quantifying Biases in LLM-as-a-Judgecited by 368
  9. 2024Self-Preference Bias in LLM-as-a-Judgecited by 258
  10. 2023VisIT-Bench: A Benchmark for Vision-Language Instruction Following Inspired by Real-World Usecited by 102
  11. 2023LLMs as Narcissistic Evaluators: When Ego Inflates Evaluation Scorescited by 90
  12. 2024CodeUltraFeedback: An LLM-as-a-Judge Dataset for Aligning Large Language Models to Coding Preferencescited by 57
7 more
  1. 2025A Unified Understanding and Evaluation of Steering Methodscited by 46
  2. 2025Beyond the Surface: Measuring Self-Preference in LLM Judgmentscited by 31
  3. 2025One-shot Optimized Steering Vectors Mediate Safety-relevant Behaviors in LLMscited by 29
  4. 2025Leveraging Uncertainty Estimation for Efficient LLM Routingcited by 21
  5. 2024Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instructcited by 17
  6. 2025Rerouting LLM Routerscited by 11
  7. 2025Configurable Preference Tuning with Rubric-Guided Synthetic Datacited by 8

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Self-Preference Bias in Rubric-Based Evaluation of Large Language Modelscited by 11
  2. 2026Are LLM Evaluators Really Narcissists? Sanity Checking Self-Preference Evaluationscited by 5
  3. 2026One Axis, No Brake: Self-Knowledge Limits the Filtering of Harmful Peer Conformity in LLMscited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2026Automated Concept Discovery for LLM-as-a-Judge Preference Analysis
  2. 2026Quantifying and Mitigating Self-Preference Bias of LLM Judges
  3. 2025Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs
  4. 2024Dissecting Human and LLM Preferences
  5. 2024Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models
  6. 2024Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement
  7. 2025Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation
  8. 2025Refine-n-Judge: Curating High-Quality Preference Chains for LLM-Fine-Tuning