{"paper":{"title":"Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning with format and accuracy rewards enables explicit reasoning chains to guide image segmentation.","cross_cats":["cs.MM"],"primary_cat":"cs.CV","authors_text":"Bei Yu, Bohao Peng, Fanbin Lu, Jiaya Jia, Yuqi Liu, Zhisheng Zhong, Zihao Yue","submitted_at":"2025-03-09T08:48:51Z","abstract_excerpt":"Traditional methods for reasoning segmentation rely on supervised fine-tuning with categorical labels and simple descriptions, limiting its out-of-domain generalization and lacking explicit reasoning processes. To address these limitations, we propose Seg-Zero, a novel framework that demonstrates remarkable generalizability and derives explicit chain-of-thought reasoning through cognitive reinforcement. Seg-Zero introduces a decoupled architecture consisting of a reasoning model and a segmentation model. The reasoning model interprets user intentions, generates explicit reasoning chains, and p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Seg-Zero-7B achieves a zero-shot performance of 57.5 on the ReasonSeg benchmark, surpassing the prior LISA-7B by 18%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the format-plus-accuracy reward mechanism, applied only through reinforcement learning without any explicit reasoning supervision, reliably produces useful and generalizable chain-of-thought reasoning rather than superficial patterns that happen to score well on the training distribution.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Seg-Zero uses cognitive reinforcement learning on a decoupled reasoning-plus-segmentation architecture to produce explicit reasoning chains and reach 57.5 zero-shot accuracy on ReasonSeg, beating prior supervised LISA-7B by 18%.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning with format and accuracy rewards enables explicit reasoning chains to guide image segmentation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d4d9daea391d23569cc2e7031b0e2fd29746279130e22f10d41b8cc64f67839d"},"source":{"id":"2503.06520","kind":"arxiv","version":2},"verdict":{"id":"3f881baf-43c2-4cd6-b18e-dfa457c6dc42","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T12:27:04.945088Z","strongest_claim":"Seg-Zero-7B achieves a zero-shot performance of 57.5 on the ReasonSeg benchmark, surpassing the prior LISA-7B by 18%.","one_line_summary":"Seg-Zero uses cognitive reinforcement learning on a decoupled reasoning-plus-segmentation architecture to produce explicit reasoning chains and reach 57.5 zero-shot accuracy on ReasonSeg, beating prior supervised LISA-7B by 18%.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the format-plus-accuracy reward mechanism, applied only through reinforcement learning without any explicit reasoning supervision, reliably produces useful and generalizable chain-of-thought reasoning rather than superficial patterns that happen to score well on the training distribution.","pith_extraction_headline":"Reinforcement learning with format and accuracy rewards enables explicit reasoning chains to guide image segmentation."},"references":{"count":45,"sample":[{"doi":"","year":2017,"title":"Segnet: A deep convolutional encoder-decoder architecture for image segmentation","work_id":"c5380531-81cc-4c05-ab24-d89795bc0a27","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":"One token to seg them all: Language instructed reasoning seg- mentation in videos","work_id":"ab23c2cb-5301-4a98-8234-5ffcf0ecf252","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs","work_id":"57309a7b-831c-447b-92aa-115c5bb5aefb","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Rethinking Atrous Convolution for Semantic Image Segmentation","work_id":"6f5d4c68-8df6-4794-b125-a10bfe8d5876","ref_index":5,"cited_arxiv_id":"1706.05587","is_internal_anchor":true}],"resolved_work":45,"snapshot_sha256":"63d13e76de97e481fcfc316ef5dca57081624cce43973d00957b82d449c616ae","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"}