{"paper":{"title":"Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Large reasoning models outperform standard LLMs as judges on accuracy and robustness but still carry strong evaluation biases that an explicit planning step can reduce.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hui Huang, Muyun Yang, Xuanxin Wu, Yuki Arase","submitted_at":"2026-01-07T06:19:26Z","abstract_excerpt":"This paper presents the first systematic comparison investigating whether Large Reasoning Models (LRMs) are superior judges to non-reasoning LLMs. Our empirical analysis yields four key findings: 1) LRMs outperform non-reasoning LLMs in terms of judgment accuracy, particularly on reasoning-intensive tasks; 2) LRMs demonstrate superior evaluation instruction-following capabilities; 3) LRMs exhibit enhanced robustness against adversarial attacks targeting judgment tasks; 4) However, LRMs still exhibit strong evaluation biases. To mitigate this bias vulnerability, we propose PlanJudge, a lightwei"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"LRMs outperform non-reasoning LLMs in judgment accuracy, particularly on reasoning-intensive tasks, demonstrate superior instruction-following and robustness, yet still exhibit strong evaluation biases that PlanJudge mitigates while preserving accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen tasks, adversarial attacks, and bias metrics comprehensively capture real-world judgment scenarios and that observed improvements generalize beyond the tested models and datasets.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Reasoning models judge better than non-reasoning LLMs yet retain biases; generating an evaluation plan first mitigates bias without losing accuracy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large reasoning models outperform standard LLMs as judges on accuracy and robustness but still carry strong evaluation biases that an explicit planning step can reduce.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3de6081e860dd93d181654d98c34111e51e81d105489df81c37d07e5720d9adf"},"source":{"id":"2601.03630","kind":"arxiv","version":2},"verdict":{"id":"d134c7ab-42ab-4aec-9c54-d29387a47ef1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T17:14:40.009656Z","strongest_claim":"LRMs outperform non-reasoning LLMs in judgment accuracy, particularly on reasoning-intensive tasks, demonstrate superior instruction-following and robustness, yet still exhibit strong evaluation biases that PlanJudge mitigates while preserving accuracy.","one_line_summary":"Reasoning models judge better than non-reasoning LLMs yet retain biases; generating an evaluation plan first mitigates bias without losing accuracy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen tasks, adversarial attacks, and bias metrics comprehensively capture real-world judgment scenarios and that observed improvements generalize beyond the tested models and datasets.","pith_extraction_headline":"Large reasoning models outperform standard LLMs as judges on accuracy and robustness but still carry strong evaluation biases that an explicit planning step can reduce."},"references":{"count":12,"sample":[{"doi":"","year":2025,"title":"InFindings of the Association for Computational Linguistics: ACL 2025, pages 5880–5895","work_id":"61bb32c1-ce49-4dfa-a311-d0ba4eeb1709","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev","work_id":"a48d5ada-8623-467d-baf4-93bd47703121","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Are reasoning models more prone to hallucination?","work_id":"cbb24d0e-0a95-46cb-a43e-115b3c4115d7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Planning: A detailed evaluation plan is specified based on the current evaluation task","work_id":"1630ead5-cdf4-4899-ae6d-c35802dd91ed","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"We investigate three distinct strategies for the first step of plan generation:","work_id":"903f762c-74ef-4a5e-a50c-0e26882caa22","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":12,"snapshot_sha256":"1860e2cff397bae7d9f78cbc8e77b854d9560dad92b597496023dcfbcf3fe2ac","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"}