{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:4VWJ56OSVU3YRW7B7G3GN6PVJH","short_pith_number":"pith:4VWJ56OS","schema_version":"1.0","canonical_sha256":"e56c9ef9d2ad3788dbe1f9b666f9f549fcb9010b386d335dee6435bed7c080e2","source":{"kind":"arxiv","id":"2505.22943","version":1},"attestation_state":"computed","paper":{"title":"Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV","cs.LG","cs.SD"],"primary_cat":"cs.CL","authors_text":"Dayoon Ko, Gunhee Kim, Heeseung Yun, Jaewoo Ahn","submitted_at":"2025-05-28T23:45:55Z","abstract_excerpt":"While pre-trained multimodal representations (e.g., CLIP) have shown impressive capabilities, they exhibit significant compositional vulnerabilities leading to counterintuitive judgments. We introduce Multimodal Adversarial Compositionality (MAC), a benchmark that leverages large language models (LLMs) to generate deceptive text samples to exploit these vulnerabilities across different modalities and evaluates them through both sample-wise attack success rate and group-wise entropy-based diversity. To improve zero-shot methods, we propose a self-training approach that leverages rejection-sampl"},"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":"2505.22943","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-28T23:45:55Z","cross_cats_sorted":["cs.AI","cs.CV","cs.LG","cs.SD"],"title_canon_sha256":"57207b9460916f862c1c89fd5323c7b47f53bfb263ab4087705be2cdd357d23a","abstract_canon_sha256":"23eb6b2ea5edc65df8661d7a5f4a2be188a7810577095522cd0c644be735c183"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:11:52.085619Z","signature_b64":"c+8+0dCu5NNoTXLlmgwGGDUukZOTxL9m+Oqo/u36GYDagYB24lGLpt0aS5/yVxpjfQkAebbcz+MNv9QSQLQRDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e56c9ef9d2ad3788dbe1f9b666f9f549fcb9010b386d335dee6435bed7c080e2","last_reissued_at":"2026-07-05T11:11:52.085083Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:11:52.085083Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV","cs.LG","cs.SD"],"primary_cat":"cs.CL","authors_text":"Dayoon Ko, Gunhee Kim, Heeseung Yun, Jaewoo Ahn","submitted_at":"2025-05-28T23:45:55Z","abstract_excerpt":"While pre-trained multimodal representations (e.g., CLIP) have shown impressive capabilities, they exhibit significant compositional vulnerabilities leading to counterintuitive judgments. We introduce Multimodal Adversarial Compositionality (MAC), a benchmark that leverages large language models (LLMs) to generate deceptive text samples to exploit these vulnerabilities across different modalities and evaluates them through both sample-wise attack success rate and group-wise entropy-based diversity. To improve zero-shot methods, we propose a self-training approach that leverages rejection-sampl"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.22943","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/2505.22943/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":"2505.22943","created_at":"2026-07-05T11:11:52.085141+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.22943v1","created_at":"2026-07-05T11:11:52.085141+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.22943","created_at":"2026-07-05T11:11:52.085141+00:00"},{"alias_kind":"pith_short_12","alias_value":"4VWJ56OSVU3Y","created_at":"2026-07-05T11:11:52.085141+00:00"},{"alias_kind":"pith_short_16","alias_value":"4VWJ56OSVU3YRW7B","created_at":"2026-07-05T11:11:52.085141+00:00"},{"alias_kind":"pith_short_8","alias_value":"4VWJ56OS","created_at":"2026-07-05T11:11:52.085141+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/4VWJ56OSVU3YRW7B7G3GN6PVJH","json":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH.json","graph_json":"https://pith.science/api/pith-number/4VWJ56OSVU3YRW7B7G3GN6PVJH/graph.json","events_json":"https://pith.science/api/pith-number/4VWJ56OSVU3YRW7B7G3GN6PVJH/events.json","paper":"https://pith.science/paper/4VWJ56OS"},"agent_actions":{"view_html":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH","download_json":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH.json","view_paper":"https://pith.science/paper/4VWJ56OS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.22943&json=true","fetch_graph":"https://pith.science/api/pith-number/4VWJ56OSVU3YRW7B7G3GN6PVJH/graph.json","fetch_events":"https://pith.science/api/pith-number/4VWJ56OSVU3YRW7B7G3GN6PVJH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH/action/storage_attestation","attest_author":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH/action/author_attestation","sign_citation":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH/action/citation_signature","submit_replication":"https://pith.science/pith/4VWJ56OSVU3YRW7B7G3GN6PVJH/action/replication_record"}},"created_at":"2026-07-05T11:11:52.085141+00:00","updated_at":"2026-07-05T11:11:52.085141+00:00"}