{"paper":{"title":"Permutation-Consensus Listwise Judging for Robust Factuality Evaluation","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Rerunning listwise factuality prompts over multiple candidate orderings and aggregating the results produces up to 7-point gains over direct judging.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Elsa Fan, Justin Tang, Nathan Huang, Tianyi Huang, Wenqian Chen","submitted_at":"2026-03-20T23:35:14Z","abstract_excerpt":"Large language models (LLMs) are now widely used as judges, yet their decisions can change under presentation choices that should be irrelevant. We study one such source of instability: candidate-order sensitivity in listwise factuality evaluation, where several answers can look similarly polished while differing substantially in hallucination risk. We introduce PCFJudge, an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus deci"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On RewardBench 2 Factuality, PCFJudge improves over direct judging by up to 7 absolute points.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That order sensitivity is a dominant, fixable source of error in listwise factuality judging and that simple aggregation over permutations does not introduce compensating biases or reduce discriminative power.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PCFJudge improves direct LLM factuality judging by up to 7 points on RewardBench 2 Factuality by aggregating results over multiple candidate permutations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Rerunning listwise factuality prompts over multiple candidate orderings and aggregating the results produces up to 7-point gains over direct judging.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c68f66b662db5835cf046036d05e1ddccf3636dfabb6850a40a5f541e1652e11"},"source":{"id":"2603.20562","kind":"arxiv","version":3},"verdict":{"id":"798d4e1b-8339-4951-a6f9-5627e1118dd8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T07:40:00.308621Z","strongest_claim":"On RewardBench 2 Factuality, PCFJudge improves over direct judging by up to 7 absolute points.","one_line_summary":"PCFJudge improves direct LLM factuality judging by up to 7 points on RewardBench 2 Factuality by aggregating results over multiple candidate permutations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That order sensitivity is a dominant, fixable source of error in listwise factuality judging and that simple aggregation over permutations does not introduce compensating biases or reduce discriminative power.","pith_extraction_headline":"Rerunning listwise factuality prompts over multiple candidate orderings and aggregating the results produces up to 7-point gains over direct judging."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.20562/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"18923d2f25a092660d10d8c7010dd1c664130752d6fe7974b1e19df3d06303ab"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}