{"paper":{"title":"Self-Preference Bias in LLM-as-a-Judge","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LLMs as judges give higher scores to low-perplexity outputs than humans, even for non-self-generated text.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Koki Wataoka, Ryokan Ri, Tsubasa Takahashi","submitted_at":"2024-10-29T07:42:18Z","abstract_excerpt":"Automated evaluation leveraging large language models (LLMs), commonly referred to as LLM evaluators or LLM-as-a-judge, has been widely used in measuring the performance of dialogue systems. However, the self-preference bias in LLMs has posed significant risks, including promoting specific styles or policies intrinsic to the LLMs. Despite the importance of this issue, there is a lack of established methods to measure the self-preference bias quantitatively, and its underlying causes are poorly understood. In this paper, we introduce a novel quantitative metric to measure the self-preference bi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"LLMs assign significantly higher evaluations to outputs with lower perplexity than human evaluators, regardless of whether the outputs were self-generated. This suggests that the essence of the bias lies in perplexity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the introduced quantitative metric isolates self-preference bias from other confounding factors in LLM judgments and that the observed correlation with perplexity is causal rather than correlational.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLMs judge their own outputs higher because they assign better scores to lower-perplexity text, even when the text is not self-generated.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLMs as judges give higher scores to low-perplexity outputs than humans, even for non-self-generated text.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b88afe9a505f5bed9eb2fa59f15e8d2e711563d960609587b928592d4625f10f"},"source":{"id":"2410.21819","kind":"arxiv","version":2},"verdict":{"id":"0c1f9f11-198c-40b0-b960-8a279a2b9b36","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T14:41:39.613451Z","strongest_claim":"LLMs assign significantly higher evaluations to outputs with lower perplexity than human evaluators, regardless of whether the outputs were self-generated. This suggests that the essence of the bias lies in perplexity.","one_line_summary":"LLMs judge their own outputs higher because they assign better scores to lower-perplexity text, even when the text is not self-generated.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the introduced quantitative metric isolates self-preference bias from other confounding factors in LLM judgments and that the observed correlation with perplexity is causal rather than correlational.","pith_extraction_headline":"LLMs as judges give higher scores to low-perplexity outputs than humans, even for non-self-generated text."},"references":{"count":19,"sample":[{"doi":"","year":2022,"title":"Daniel Deutsch, Rotem Dror, and Dan Roth","work_id":"9c5e3e50-9dfd-46b0-8bcd-68a014329574","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2022.emnlp-main.753","year":2022,"title":"On the limitations of reference-free evaluations of generated text","work_id":"78772036-6fc7-459c-9c27-78fc06e14ba3","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Bowman, and Shi Feng","work_id":"ec9ad2bb-47a2-46c7-b8e9-05f69142decd","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Chenghao Yang, Sida Li, and Ari Holtzman","work_id":"ae09097d-58a5-40bc-8f92-14d333fc4a2d","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Large language models are inconsistent and biased evaluators","work_id":"36f79a6c-dedd-407f-a53d-20491052e910","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":19,"snapshot_sha256":"bb4713df88645d3ad40abc47c71b26be069a96e88e0a57f199ede6bfcafe082a","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"cf90e20267b64a74744ee5afa5fe7365ee7ac929677e5ffcf204c56b1e214bf6"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}