{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:UZ2J6ZWGSBSBCDLR4TBICFOFJE","short_pith_number":"pith:UZ2J6ZWG","schema_version":"1.0","canonical_sha256":"a6749f66c69064110d71e4c28115c54909306232c7fe5018912257da3fd9d5ce","source":{"kind":"arxiv","id":"2608.08802","version":1},"attestation_state":"computed","paper":{"title":"Improving Generalization Robustness of Multimodal RLVR","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Bin Xu, Bohan Zhuang, Chenrui Zhou, Fan Wang, Jiajun Song, Jiasheng Tang, Lama Moukheiber, Pengfei Zhou, Wangbo Zhao, Xiaopeng Peng, Yang You, Yixing Ma, Zhenglin Wan, Zhiwei Tang","submitted_at":"2026-08-09T16:36:42Z","abstract_excerpt":"Reinforcement Learning with Verifiable Rewards (RLVR) makes Multimodal Large Language Models more accurate, but the gains are brittle: simply paraphrasing a question or changing the prompt template can degrade them, which challenges reliable deployment in high-stakes scenarios like medical VQA. We trace this to two issues of the standard RL objective. First, the binary verifier conflates format with content, so the reward signal cannot tell a wrong answer apart from a misformatted one. Second, the training distribution covers only a thin slice of the real-world prompts that the model might mee"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2608.08802","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-08-09T16:36:42Z","cross_cats_sorted":[],"title_canon_sha256":"a6a66f22b8f1bb55cca390fdd5c521c927f99b36d72459cc841404c413339ceb","abstract_canon_sha256":"852e706141423ff992834211430e23b34545bec454d115374ede8e1774a21f76"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-11T01:25:23.099120Z","signature_b64":"ru10o48sisVxCuqAi0s+W7vU7k3zANyFlkc4SnLWZuTCSlcOWumBhdg0SgtEesFgToib1lIgvBpIGm7JcfHpBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a6749f66c69064110d71e4c28115c54909306232c7fe5018912257da3fd9d5ce","last_reissued_at":"2026-08-11T01:25:23.096423Z","signature_status":"signed_v1","first_computed_at":"2026-08-11T01:25:23.096423Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Improving Generalization Robustness of Multimodal RLVR","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Bin Xu, Bohan Zhuang, Chenrui Zhou, Fan Wang, Jiajun Song, Jiasheng Tang, Lama Moukheiber, Pengfei Zhou, Wangbo Zhao, Xiaopeng Peng, Yang You, Yixing Ma, Zhenglin Wan, Zhiwei Tang","submitted_at":"2026-08-09T16:36:42Z","abstract_excerpt":"Reinforcement Learning with Verifiable Rewards (RLVR) makes Multimodal Large Language Models more accurate, but the gains are brittle: simply paraphrasing a question or changing the prompt template can degrade them, which challenges reliable deployment in high-stakes scenarios like medical VQA. We trace this to two issues of the standard RL objective. First, the binary verifier conflates format with content, so the reward signal cannot tell a wrong answer apart from a misformatted one. Second, the training distribution covers only a thin slice of the real-world prompts that the model might mee"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.08802","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2608.08802/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2608.08802","created_at":"2026-08-11T01:25:23.097433+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.08802v1","created_at":"2026-08-11T01:25:23.097433+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.08802","created_at":"2026-08-11T01:25:23.097433+00:00"},{"alias_kind":"pith_short_12","alias_value":"UZ2J6ZWGSBSB","created_at":"2026-08-11T01:25:23.097433+00:00"},{"alias_kind":"pith_short_16","alias_value":"UZ2J6ZWGSBSBCDLR","created_at":"2026-08-11T01:25:23.097433+00:00"},{"alias_kind":"pith_short_8","alias_value":"UZ2J6ZWG","created_at":"2026-08-11T01:25:23.097433+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE","json":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE.json","graph_json":"https://pith.science/api/pith-number/UZ2J6ZWGSBSBCDLR4TBICFOFJE/graph.json","events_json":"https://pith.science/api/pith-number/UZ2J6ZWGSBSBCDLR4TBICFOFJE/events.json","paper":"https://pith.science/paper/UZ2J6ZWG"},"agent_actions":{"view_html":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE","download_json":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE.json","view_paper":"https://pith.science/paper/UZ2J6ZWG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.08802&json=true","fetch_graph":"https://pith.science/api/pith-number/UZ2J6ZWGSBSBCDLR4TBICFOFJE/graph.json","fetch_events":"https://pith.science/api/pith-number/UZ2J6ZWGSBSBCDLR4TBICFOFJE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE/action/storage_attestation","attest_author":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE/action/author_attestation","sign_citation":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE/action/citation_signature","submit_replication":"https://pith.science/pith/UZ2J6ZWGSBSBCDLR4TBICFOFJE/action/replication_record"}},"created_at":"2026-08-11T01:25:23.097433+00:00","updated_at":"2026-08-11T01:25:23.097433+00:00"}