{"paper":{"title":"PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Decomposing confidence rewards into perception and reasoning clusters improves vision-language model training.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chang D. Yoo, Chong Luo, Eunseop Yoon, Gwanhyeong Koo, Hee Suk Yoon, Ji Woo Hong, Mark Hasegawa-Johnson, Qi Dai, SooHwan Eom","submitted_at":"2026-05-13T12:55:18Z","abstract_excerpt":"Reinforcement Learning with Verifiable Rewards (RLVR) traditionally relies on a sparse, outcome-based signal. Recent work shows that providing a fine-grained, model-intrinsic signal (rewarding the confidence growth in the ground-truth answer) effectively improves language reasoning training by providing step-level guidance without costly external models. While effective for unimodal text, we find that naively applying this global reward to vision-language (V-L) reasoning is a suboptimal strategy, as the task is a heterogeneous mix of sparse visual perception and dense textual reasoning. This g"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate that PDCR outperforms the naive, global-reward formulation and sparse-reward baselines on key V-L reasoning benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The unsupervised clustering driven by the model-internal Visual Dependence Score accurately partitions steps into perception and reasoning without any labeled supervision or external verification.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PDCR improves vision-language reasoning by computing separate normalized confidence advantages for perception steps and reasoning steps after unsupervised decomposition.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Decomposing confidence rewards into perception and reasoning clusters improves vision-language model training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"714cd5196fa51bd7896d97721fa492f8bbceff8d1586e8100f653db7bb716ee2"},"source":{"id":"2605.13467","kind":"arxiv","version":1},"verdict":{"id":"d9383de8-b038-4d3c-bb25-311e4729b4d7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:16:10.963422Z","strongest_claim":"We demonstrate that PDCR outperforms the naive, global-reward formulation and sparse-reward baselines on key V-L reasoning benchmarks.","one_line_summary":"PDCR improves vision-language reasoning by computing separate normalized confidence advantages for perception steps and reasoning steps after unsupervised decomposition.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The unsupervised clustering driven by the model-internal Visual Dependence Score accurately partitions steps into perception and reasoning without any labeled supervision or external verification.","pith_extraction_headline":"Decomposing confidence rewards into perception and reasoning clusters improves vision-language model training."},"references":{"count":79,"sample":[{"doi":"","year":2023,"title":"Gpt-4v(ision) system card. 2023. 1","work_id":"a549d3c1-1572-4949-8f9e-de9af53432a9","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Qwen2.5-VL Technical Report","work_id":"69dffacb-bfe8-442d-be86-48624c60426f","ref_index":2,"cited_arxiv_id":"2502.13923","is_internal_anchor":true},{"doi":"","year":2025,"title":"Vrprm: Process reward modeling via visual reasoning.arXiv preprint arXiv:2508.03556, 2025","work_id":"673ea4d5-7e09-4f7f-acf3-106a8e69f259","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"arXiv preprint arXiv:2505.23558 , year=","work_id":"cc9e5e98-c226-4e12-bf74-ff438225c4c0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Ultrafeedback: Boosting language models with scaled ai feedback","work_id":"bc448bac-e570-4911-b8b6-fc2bf33a532a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":79,"snapshot_sha256":"eaa94caf858d0e6c7ba30d5f326c505e9c792d75883c64c48a8460a16dd29639","internal_anchors":12},"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"}