Research map / Large language models / Language, safety and interpretability
Are LLM Evaluators Really Narcissists? Sanity Checking Self-Preference Evaluations
Dani Roytburg, Matthew Bozoukov, Matthew Nguyen, Jou Barzdukas, Mackenzie Puig-Hall, Narmeen Fatimah Oozeer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| subtract an outcome-matched proxy baseline from self-vote probability | self-preference measurement · MATH500, Qwen2.5-72B | ILSP (%): 54.1 -> 20.0 | Improved |
| subtract outcome-matched proxy baseline | self-preference measurement · MATH500, Llama-3.2-3B | ILSP (%): 30.6 -> -5.1 | Improved |
| subtract outcome-matched proxy baseline | self-preference measurement · MATH500 | relative self-preference: -98.76% | Improved |
| subtract outcome-matched proxy baseline | self-preference measurement · objective tasks (math, code) | relative self-preference: -89.26% | Improved |
| subtract outcome-matched proxy baseline | self-preference measurement · subjective tasks (translation, AlpacaEval, truthfulness) | relative self-preference: -82.57%, -79.19%, -67.57% | Improved |
| subtract outcome-matched proxy baseline | self-preference measurement · MMLU | ILSP (%): 62.0 -> 19.3 | Mixed |
| subtract outcome-matched proxy baseline | self-preference measurement · reproduced prior experiments | significant examples: 51% retain significance | Mixed |
| add chain-of-thought reasoning before the judge verdict | LLM judge · MATH500, Qwen2.5-7B | ILSP (%): 18.9 vs 20.0 baseline-adjusted | Hurt |
| add chain-of-thought reasoning before the judge verdict | LLM judge · MMLU | ILSP (%): 10.3 | Hurt |
| add chain-of-thought reasoning before the judge verdict | LLM judge · MATH500, MBPP-Plus, MMLU | ILSP (%) | Mixed |
| restrict proxies to out-of-family models instead of all proxies (ablation) | Evaluator Quality Baseline · sensitivity analysis | ILSP mean diff: 0.0197 -> 0.0227 | No clear effect |
| increase number of matched proxies per example (ablation) | Evaluator Quality Baseline · all datasets | mean test statistic | No clear effect |
Closest papers by meaning
- LLM Evaluators Recognize and Favor Their Own Generations (NeurIPS 2024 · similarity 0.90)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.88)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.88)
- Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.88)
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.87)
- SELF-[IN]CORRECT: LLMs Struggle with Discriminating Self-Generated Responses (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
