{"paper":{"title":"Revealing Interpretable Failure Modes of VLMs","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"REVELIO uncovers interpretable concept compositions that cause consistent failures in vision-language models.","cross_cats":["cs.LG","cs.RO"],"primary_cat":"cs.AI","authors_text":"Gagandeep Singh, Isha Chaudhary, Kavya Sachdeva, Sayan Ranu, Vedaant V Jain","submitted_at":"2026-05-12T19:25:17Z","abstract_excerpt":"Vision-Language Models (VLMs) are increasingly used in safety-critical applications because of their broad reasoning capabilities and ability to generalize with minimal task-specific engineering. Despite these advantages, they can exhibit catastrophic failures in specific real-world situations, constituting failure modes.\n  We introduce REVELIO, a framework for systematically uncovering interpretable failure modes in VLMs. We define a failure mode as a composition of interpretable, domain-relevant concepts-such as pedestrian proximity or adverse weather conditions-under which a target VLM cons"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We introduce REVELIO, a framework for systematically uncovering interpretable failure modes in VLMs... uncovering previously unreported vulnerabilities in state-of-the-art VLMs.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the searched concept compositions correspond to genuine, consistent real-world failure modes rather than artifacts of the simulation or search heuristics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"REVELIO uncovers interpretable failure modes in VLMs by searching combinatorial concept spaces with diversity-aware beam search and Gaussian-process Thompson sampling, revealing vulnerabilities in autonomous driving and indoor robotics.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"REVELIO uncovers interpretable concept compositions that cause consistent failures in vision-language models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5ee097f560c6dabd233b8134cd5222b6336b1b5393f3603f4a39ca5d910995b8"},"source":{"id":"2605.12674","kind":"arxiv","version":1},"verdict":{"id":"963a4c7c-278c-48e5-8987-37b50bf4cf64","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:24:46.550617Z","strongest_claim":"We introduce REVELIO, a framework for systematically uncovering interpretable failure modes in VLMs... uncovering previously unreported vulnerabilities in state-of-the-art VLMs.","one_line_summary":"REVELIO uncovers interpretable failure modes in VLMs by searching combinatorial concept spaces with diversity-aware beam search and Gaussian-process Thompson sampling, revealing vulnerabilities in autonomous driving and indoor robotics.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the searched concept compositions correspond to genuine, consistent real-world failure modes rather than artifacts of the simulation or search heuristics.","pith_extraction_headline":"REVELIO uncovers interpretable concept compositions that cause consistent failures in vision-language models."},"references":{"count":222,"sample":[{"doi":"","year":2025,"title":"Claude Sonnet","work_id":"f55b445a-ab6e-4e4d-b496-c3a74994d3ed","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Large language model-assisted autonomous vehicle recovery from immobilization, 2025","work_id":"0ffe6562-93d4-4fef-8d21-e5d1e61a4cba","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1998,"title":"The use of mmr, diversity-based reranking for reordering documents and producing summaries","work_id":"98357628-52da-45c9-8f07-1e7de5d019ca","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Gonzalez, and Ion Stoica","work_id":"7b049cb6-6a65-4610-bb33-d5863d8ae9eb","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Kevin Kai-Chun Chang, Ekin Beyazit, Alberto Sangiovanni-Vincentelli, Tichakorn Wongpirom- sarn, and Sanjit A. Seshia. Scenicrules: An autonomous driving benchmark with multi-objective specifications a","work_id":"34756ee5-9047-48e6-94a2-0bdd3a67767f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":222,"snapshot_sha256":"7612c06159f671a97d092a461973de24bd8ddc6736d7d64bca477774a45c3ecc","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"72f7a87453b88f76e969da74fd45d0ed54029c0c5b87df5101475dcdfa1bd8c9"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}