{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:CVEHX7XFUKF5DJRZPQYQDRXULA","short_pith_number":"pith:CVEHX7XF","schema_version":"1.0","canonical_sha256":"15487bfee5a28bd1a6397c3101c6f45813b6d8a765e6b9054f054e2bec2ead97","source":{"kind":"arxiv","id":"2410.21819","version":2},"attestation_state":"computed","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"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2410.21819","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-10-29T07:42:18Z","cross_cats_sorted":[],"title_canon_sha256":"3b78afe76893d6460aba6e3ad0163ef7cd549dbc3cff746644ae9a9fa89484a6","abstract_canon_sha256":"3939e6fde3f28e3e5448035dc81ace7f4a29cf0370051bdb1e096f49993cfa03"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:52.320468Z","signature_b64":"gFqF/U/2ogdpWmM9X2TD073OTauYLhRidoge9M5t0zo3jkR8Wgn9iTZcoP4Dip0pOEDvB7W8+dw/mfNX+Y9RDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"15487bfee5a28bd1a6397c3101c6f45813b6d8a765e6b9054f054e2bec2ead97","last_reissued_at":"2026-05-17T23:38:52.319961Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:52.319961Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2410.21819","created_at":"2026-05-17T23:38:52.320045+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.21819v2","created_at":"2026-05-17T23:38:52.320045+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.21819","created_at":"2026-05-17T23:38:52.320045+00:00"},{"alias_kind":"pith_short_12","alias_value":"CVEHX7XFUKF5","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"CVEHX7XFUKF5DJRZ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"CVEHX7XF","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":27,"sample":[{"citing_arxiv_id":"2605.21622","citing_title":"TO-Agents: A Multi-Agent AI Pipeline for Preference-Guided Topology Optimization","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2509.26464","citing_title":"Extreme Self-Preference in Language Models","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20478","citing_title":"Stage-Audit: Auditable Source-Frontier Discovery for Cross-Wiki Tables","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19529","citing_title":"Generative-Evaluative Agreement: A Necessary Validity Criterion for LLM-Enabled Adaptive Assessment","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2510.07517","citing_title":"When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2511.01490","citing_title":"Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2601.17230","citing_title":"CaseFacts: A Benchmark for Legal Fact-Checking and Precedent Retrieval","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2601.18027","citing_title":"Sentipolis: Emotion-Aware Agents for Social Simulations","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03472","citing_title":"Detecting Stealth Sycophancy in Mental-Health Dialogue with Dynamic Emotional Signature Graphs","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2604.27727","citing_title":"LLM-as-a-Judge for Human-AI Co-Creation: A Reliability-Aware Evaluation Framework for Coding","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2604.26243","citing_title":"StratMem-Bench: Evaluating Strategic Memory Use in Virtual Character Conversation Beyond Factual Recall","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10850","citing_title":"Verification Mirage: Mapping the Reliability Boundary of Self-Verification in Medical VQA","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10379","citing_title":"Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness","ref_index":46,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03858","citing_title":"MCJudgeBench: A Benchmark for Constraint-Level Judge Evaluation in Multi-Constraint Instruction Following","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24700","citing_title":"Green Shielding: A User-Centric Approach Towards Trustworthy AI","ref_index":56,"is_internal_anchor":true},{"citing_arxiv_id":"2604.22517","citing_title":"Aggregate vs. Personalized Judges in Business Idea Evaluation: Evidence from Expert Disagreement","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19598","citing_title":"Cross-Model Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Across Three Large Language Models","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12994","citing_title":"LogicEval: A Systematic Framework for Evaluating Automated Repair Techniques for Logical Vulnerabilities in Real-World Software","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2604.11287","citing_title":"Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Using a Large Language Model","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2604.07650","citing_title":"How Independent are Large Language Models? A Statistical Framework for Auditing Behavioral Entanglement and Reweighting Verifier Ensembles","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06731","citing_title":"When Routine Chats Turn Toxic: Unintended Long-Term State Poisoning in Personalized Agents","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06996","citing_title":"Self-Preference Bias in Rubric-Based Evaluation of Large Language Models","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2604.05593","citing_title":"Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge","ref_index":46,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17658","citing_title":"Towards Self-Improving Error Diagnosis in Multi-Agent Systems","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17197","citing_title":"Learning to Control Summaries with Score Ranking","ref_index":47,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA","json":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA.json","graph_json":"https://pith.science/api/pith-number/CVEHX7XFUKF5DJRZPQYQDRXULA/graph.json","events_json":"https://pith.science/api/pith-number/CVEHX7XFUKF5DJRZPQYQDRXULA/events.json","paper":"https://pith.science/paper/CVEHX7XF"},"agent_actions":{"view_html":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA","download_json":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA.json","view_paper":"https://pith.science/paper/CVEHX7XF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.21819&json=true","fetch_graph":"https://pith.science/api/pith-number/CVEHX7XFUKF5DJRZPQYQDRXULA/graph.json","fetch_events":"https://pith.science/api/pith-number/CVEHX7XFUKF5DJRZPQYQDRXULA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA/action/storage_attestation","attest_author":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA/action/author_attestation","sign_citation":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA/action/citation_signature","submit_replication":"https://pith.science/pith/CVEHX7XFUKF5DJRZPQYQDRXULA/action/replication_record"}},"created_at":"2026-05-17T23:38:52.320045+00:00","updated_at":"2026-05-17T23:38:52.320045+00:00"}