{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:443FAUPAS23PTLBNLV744XM3XD","short_pith_number":"pith:443FAUPA","schema_version":"1.0","canonical_sha256":"e7365051e096b6f9ac2d5d7fce5d9bb8feca0a0694133c4f7e39cd9a3417a676","source":{"kind":"arxiv","id":"2602.12506","version":3},"attestation_state":"computed","paper":{"title":"On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Anshul Shah, Arnab Mondal, Joerg Liebelt, Rosie Zhao, Xiaoyu Zhu, Xinke Deng, Yang Yang, Zhongyu Jiang","submitted_at":"2026-02-13T01:12:00Z","abstract_excerpt":"Reinforcement learning (RL) finetuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision-language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations, including misleading captions or incorrect chain-of-thought (CoT) traces, cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consist"},"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":"2602.12506","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-13T01:12:00Z","cross_cats_sorted":[],"title_canon_sha256":"355c22f110541a844b03e50e8d4b7085ec85fda70924df4db6df49b8944b13d8","abstract_canon_sha256":"9627cadb69aa69ad8e684ae540c54059152866d35ffbfe342548522d85f4aed6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:03:57.554766Z","signature_b64":"ZceJCchfLa9rnNuhZl7NT/c9ewZx9x0fpDro9NDII/W4Fl1VCu740X1s7KXvYTQP9zWeJR2F34cnATWbRYmHCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e7365051e096b6f9ac2d5d7fce5d9bb8feca0a0694133c4f7e39cd9a3417a676","last_reissued_at":"2026-05-22T01:03:57.553834Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:03:57.553834Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Anshul Shah, Arnab Mondal, Joerg Liebelt, Rosie Zhao, Xiaoyu Zhu, Xinke Deng, Yang Yang, Zhongyu Jiang","submitted_at":"2026-02-13T01:12:00Z","abstract_excerpt":"Reinforcement learning (RL) finetuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision-language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations, including misleading captions or incorrect chain-of-thought (CoT) traces, cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consist"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.12506","kind":"arxiv","version":3},"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/2602.12506/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":"2602.12506","created_at":"2026-05-22T01:03:57.553950+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.12506v3","created_at":"2026-05-22T01:03:57.553950+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.12506","created_at":"2026-05-22T01:03:57.553950+00:00"},{"alias_kind":"pith_short_12","alias_value":"443FAUPAS23P","created_at":"2026-05-22T01:03:57.553950+00:00"},{"alias_kind":"pith_short_16","alias_value":"443FAUPAS23PTLBN","created_at":"2026-05-22T01:03:57.553950+00:00"},{"alias_kind":"pith_short_8","alias_value":"443FAUPA","created_at":"2026-05-22T01:03:57.553950+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2604.06422","citing_title":"When to Call an Apple Red: Humans Follow Introspective Rules, VLMs Don't","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2604.16913","citing_title":"The Cognitive Penalty: Ablating System 1 and System 2 Reasoning in Edge-Native SLMs for Decentralized Consensus","ref_index":20,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD","json":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD.json","graph_json":"https://pith.science/api/pith-number/443FAUPAS23PTLBNLV744XM3XD/graph.json","events_json":"https://pith.science/api/pith-number/443FAUPAS23PTLBNLV744XM3XD/events.json","paper":"https://pith.science/paper/443FAUPA"},"agent_actions":{"view_html":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD","download_json":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD.json","view_paper":"https://pith.science/paper/443FAUPA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.12506&json=true","fetch_graph":"https://pith.science/api/pith-number/443FAUPAS23PTLBNLV744XM3XD/graph.json","fetch_events":"https://pith.science/api/pith-number/443FAUPAS23PTLBNLV744XM3XD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD/action/storage_attestation","attest_author":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD/action/author_attestation","sign_citation":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD/action/citation_signature","submit_replication":"https://pith.science/pith/443FAUPAS23PTLBNLV744XM3XD/action/replication_record"}},"created_at":"2026-05-22T01:03:57.553950+00:00","updated_at":"2026-05-22T01:03:57.553950+00:00"}