{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:TXE6TRFJRNG2KJ2UKNLADZV7T5","short_pith_number":"pith:TXE6TRFJ","schema_version":"1.0","canonical_sha256":"9dc9e9c4a98b4da52754535601e6bf9f5075e9754232800fe3384b53861b1835","source":{"kind":"arxiv","id":"2506.08227","version":1},"attestation_state":"computed","paper":{"title":"A Good CREPE needs more than just Sugar: Investigating Biases in Compositional Vision-Language Benchmarks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jenia Jitsev, Matthias Bethge, Mehdi Cherti, Samuel Albanie, Shyamgopal Karthik, Vishaal Udandarao","submitted_at":"2025-06-09T20:53:43Z","abstract_excerpt":"We investigate 17 benchmarks (e.g. SugarCREPE, VALSE) commonly used for measuring compositional understanding capabilities of vision-language models (VLMs). We scrutinize design choices in their construction, including data source (e.g. MS-COCO) and curation procedures (e.g. constructing negative images/captions), uncovering several inherent biases across most benchmarks. We find that blind heuristics (e.g. token-length, log-likelihood under a language model) perform on par with CLIP models, indicating that these benchmarks do not effectively measure compositional understanding. We demonstrate"},"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":"2506.08227","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-06-09T20:53:43Z","cross_cats_sorted":[],"title_canon_sha256":"2fc4fc97b8373a5b3ee53e318d097f01e348dbef4cf1b5294f298a4292f94cfe","abstract_canon_sha256":"762c98df81dc101dc55fb51622b7c1d37538b09a93fbf4290cd92ec3c9fcf6dc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:18:46.507991Z","signature_b64":"Qbfh5Q1KDAJ3rl+IgJajKwuipgXhtPoncxImSW69EsmcWNAgDamJc+nfWxVo/pvd/VnPx7M9A+2oVZjiEpOlAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9dc9e9c4a98b4da52754535601e6bf9f5075e9754232800fe3384b53861b1835","last_reissued_at":"2026-07-05T11:18:46.507534Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:18:46.507534Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Good CREPE needs more than just Sugar: Investigating Biases in Compositional Vision-Language Benchmarks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jenia Jitsev, Matthias Bethge, Mehdi Cherti, Samuel Albanie, Shyamgopal Karthik, Vishaal Udandarao","submitted_at":"2025-06-09T20:53:43Z","abstract_excerpt":"We investigate 17 benchmarks (e.g. SugarCREPE, VALSE) commonly used for measuring compositional understanding capabilities of vision-language models (VLMs). We scrutinize design choices in their construction, including data source (e.g. MS-COCO) and curation procedures (e.g. constructing negative images/captions), uncovering several inherent biases across most benchmarks. We find that blind heuristics (e.g. token-length, log-likelihood under a language model) perform on par with CLIP models, indicating that these benchmarks do not effectively measure compositional understanding. We demonstrate"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.08227","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/2506.08227/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":"2506.08227","created_at":"2026-07-05T11:18:46.507586+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.08227v1","created_at":"2026-07-05T11:18:46.507586+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.08227","created_at":"2026-07-05T11:18:46.507586+00:00"},{"alias_kind":"pith_short_12","alias_value":"TXE6TRFJRNG2","created_at":"2026-07-05T11:18:46.507586+00:00"},{"alias_kind":"pith_short_16","alias_value":"TXE6TRFJRNG2KJ2U","created_at":"2026-07-05T11:18:46.507586+00:00"},{"alias_kind":"pith_short_8","alias_value":"TXE6TRFJ","created_at":"2026-07-05T11:18:46.507586+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25432","citing_title":"Brevity is the Soul of Inference Efficiency: Inducing Concision in VLMs via Data Curation","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11496","citing_title":"Revisiting Compositionality in Dual-Encoder Vision-Language Models: The Role of Inference","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5","json":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5.json","graph_json":"https://pith.science/api/pith-number/TXE6TRFJRNG2KJ2UKNLADZV7T5/graph.json","events_json":"https://pith.science/api/pith-number/TXE6TRFJRNG2KJ2UKNLADZV7T5/events.json","paper":"https://pith.science/paper/TXE6TRFJ"},"agent_actions":{"view_html":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5","download_json":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5.json","view_paper":"https://pith.science/paper/TXE6TRFJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.08227&json=true","fetch_graph":"https://pith.science/api/pith-number/TXE6TRFJRNG2KJ2UKNLADZV7T5/graph.json","fetch_events":"https://pith.science/api/pith-number/TXE6TRFJRNG2KJ2UKNLADZV7T5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5/action/storage_attestation","attest_author":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5/action/author_attestation","sign_citation":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5/action/citation_signature","submit_replication":"https://pith.science/pith/TXE6TRFJRNG2KJ2UKNLADZV7T5/action/replication_record"}},"created_at":"2026-07-05T11:18:46.507586+00:00","updated_at":"2026-07-05T11:18:46.507586+00:00"}