{"paper":{"title":"Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Style bias dominates LLM judges at 0.76-0.92 strength across models, dwarfing position bias, while combined debiasing improves agreement for some judges by over 11 points.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Sadman Kabir Soumik","submitted_at":"2026-04-25T07:18:30Z","abstract_excerpt":"LLM-as-a-Judge has become the dominant paradigm for evaluating language model outputs, yet LLM judges exhibit systematic biases that compromise evaluation reliability. We present a comprehensive empirical study comparing nine debiasing strategies across five judge models from four provider families (Google, Anthropic, OpenAI, Meta), three benchmarks (MT-Bench n=400, LLMBar n=200, custom n=375), and four bias types. Our headline practical finding is that a mid-tier model with the right debiasing can outperform frontier judges at a fraction of the cost: Gemini 2.5 Flash with the Combined Budget "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Style bias is the dominant bias (0.76-0.92 across all models), far exceeding position bias (<= 0.04), yet has received minimal research attention.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The chosen benchmarks (MT-Bench, LLMBar, and the custom set) and bias measurement protocols accurately isolate the targeted biases without introducing confounding artifacts from the test data itself or from how responses were generated.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Style bias dominates LLM-as-a-Judge systems far more than position bias, with debiasing strategies providing model-dependent gains and public tools released for replication.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Style bias dominates LLM judges at 0.76-0.92 strength across models, dwarfing position bias, while combined debiasing improves agreement for some judges by over 11 points.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2bd7de3189d8e0eec5c84f313280246433f95c57d0da21e811a399e27bbd3a16"},"source":{"id":"2604.23178","kind":"arxiv","version":2},"verdict":{"id":"6e8af9b5-9826-4494-a886-382390b93bb8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T08:11:19.583183Z","strongest_claim":"Style bias is the dominant bias (0.76-0.92 across all models), far exceeding position bias (<= 0.04), yet has received minimal research attention.","one_line_summary":"Style bias dominates LLM-as-a-Judge systems far more than position bias, with debiasing strategies providing model-dependent gains and public tools released for replication.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The chosen benchmarks (MT-Bench, LLMBar, and the custom set) and bias measurement protocols accurately isolate the targeted biases without introducing confounding artifacts from the test data itself or from how responses were generated.","pith_extraction_headline":"Style bias dominates LLM judges at 0.76-0.92 strength across models, dwarfing position bias, while combined debiasing improves agreement for some judges by over 11 points."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.23178/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-21T09:37:57.140419Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T23:24:23.663274Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"15c156ba6924d6fd6919298bff656bcddecd689c43cba5e7ba01d474861a1b0f"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}