AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
subtract an outcome-matched proxy baseline from self-vote probabilityself-preference measurement · MATH500, Qwen2.5-72BILSP (%): 54.1 -> 20.0Improved
subtract outcome-matched proxy baselineself-preference measurement · MATH500, Llama-3.2-3BILSP (%): 30.6 -> -5.1Improved
subtract outcome-matched proxy baselineself-preference measurement · MATH500relative self-preference: -98.76%Improved
subtract outcome-matched proxy baselineself-preference measurement · objective tasks (math, code)relative self-preference: -89.26%Improved
subtract outcome-matched proxy baselineself-preference measurement · subjective tasks (translation, AlpacaEval, truthfulness)relative self-preference: -82.57%, -79.19%, -67.57%Improved
subtract outcome-matched proxy baselineself-preference measurement · MMLUILSP (%): 62.0 -> 19.3Mixed
subtract outcome-matched proxy baselineself-preference measurement · reproduced prior experimentssignificant examples: 51% retain significanceMixed
add chain-of-thought reasoning before the judge verdictLLM judge · MATH500, Qwen2.5-7BILSP (%): 18.9 vs 20.0 baseline-adjustedHurt
add chain-of-thought reasoning before the judge verdictLLM judge · MMLUILSP (%): 10.3Hurt
add chain-of-thought reasoning before the judge verdictLLM judge · MATH500, MBPP-Plus, MMLUILSP (%)Mixed
restrict proxies to out-of-family models instead of all proxies (ablation)Evaluator Quality Baseline · sensitivity analysisILSP mean diff: 0.0197 -> 0.0227No clear effect
increase number of matched proxies per example (ablation)Evaluator Quality Baseline · all datasetsmean test statisticNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy