{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:MFNARZR7EENZ4FTMQI3CDO2GMK","short_pith_number":"pith:MFNARZR7","schema_version":"1.0","canonical_sha256":"615a08e63f211b9e166c823621bb46629402e274fb151b00a8f1aaa75ef5ba12","source":{"kind":"arxiv","id":"2305.14897","version":2},"attestation_state":"computed","paper":{"title":"Text encoders bottleneck compositionality in contrastive vision-language models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CV","cs.LG"],"primary_cat":"cs.CL","authors_text":"Amita Kamath, Jack Hessel, Kai-Wei Chang","submitted_at":"2023-05-24T08:48:44Z","abstract_excerpt":"Performant vision-language (VL) models like CLIP represent captions using a single vector. How much information about language is lost in this bottleneck? We first curate CompPrompts, a set of increasingly compositional image captions that VL models should be able to capture (e.g., single object, to object+property, to multiple interacting objects). Then, we train text-only recovery probes that aim to reconstruct captions from single-vector text representations produced by several VL models. This approach does not require images, allowing us to test on a broader range of scenes compared to pri"},"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":"2305.14897","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-05-24T08:48:44Z","cross_cats_sorted":["cs.CV","cs.LG"],"title_canon_sha256":"6d75d7dcebf24379fcc041c94405f3df9286d972c759ea149907997987530a86","abstract_canon_sha256":"a045296359831860c0851d98305b15a6d92cab413c56a788db64d7eda87bd9c3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:06:27.639495Z","signature_b64":"R061rESuSVw7vPHAw44IAyuS05+YWftQNOFjNiSW9slb41UnUWn8Ga49Z74fuT4MpNsUtTdCsHzxBYA8NEo4AQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"615a08e63f211b9e166c823621bb46629402e274fb151b00a8f1aaa75ef5ba12","last_reissued_at":"2026-07-05T07:06:27.639008Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:06:27.639008Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Text encoders bottleneck compositionality in contrastive vision-language models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CV","cs.LG"],"primary_cat":"cs.CL","authors_text":"Amita Kamath, Jack Hessel, Kai-Wei Chang","submitted_at":"2023-05-24T08:48:44Z","abstract_excerpt":"Performant vision-language (VL) models like CLIP represent captions using a single vector. How much information about language is lost in this bottleneck? We first curate CompPrompts, a set of increasingly compositional image captions that VL models should be able to capture (e.g., single object, to object+property, to multiple interacting objects). Then, we train text-only recovery probes that aim to reconstruct captions from single-vector text representations produced by several VL models. This approach does not require images, allowing us to test on a broader range of scenes compared to pri"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2305.14897","kind":"arxiv","version":2},"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/2305.14897/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":"2305.14897","created_at":"2026-07-05T07:06:27.639068+00:00"},{"alias_kind":"arxiv_version","alias_value":"2305.14897v2","created_at":"2026-07-05T07:06:27.639068+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2305.14897","created_at":"2026-07-05T07:06:27.639068+00:00"},{"alias_kind":"pith_short_12","alias_value":"MFNARZR7EENZ","created_at":"2026-07-05T07:06:27.639068+00:00"},{"alias_kind":"pith_short_16","alias_value":"MFNARZR7EENZ4FTM","created_at":"2026-07-05T07:06:27.639068+00:00"},{"alias_kind":"pith_short_8","alias_value":"MFNARZR7","created_at":"2026-07-05T07:06:27.639068+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.13288","citing_title":"Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2505.19261","citing_title":"Enhancing Text-to-Image Diffusion Transformer via Split-Text Conditioning","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2512.13511","citing_title":"Adapting MLLMs for Nuanced Video Retrieval","ref_index":40,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11301","citing_title":"LatentRouter: Can We Choose the Right Multimodal Model Before Seeing Its Answer?","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01402","citing_title":"Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression","ref_index":138,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01402","citing_title":"Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression","ref_index":138,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK","json":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK.json","graph_json":"https://pith.science/api/pith-number/MFNARZR7EENZ4FTMQI3CDO2GMK/graph.json","events_json":"https://pith.science/api/pith-number/MFNARZR7EENZ4FTMQI3CDO2GMK/events.json","paper":"https://pith.science/paper/MFNARZR7"},"agent_actions":{"view_html":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK","download_json":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK.json","view_paper":"https://pith.science/paper/MFNARZR7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2305.14897&json=true","fetch_graph":"https://pith.science/api/pith-number/MFNARZR7EENZ4FTMQI3CDO2GMK/graph.json","fetch_events":"https://pith.science/api/pith-number/MFNARZR7EENZ4FTMQI3CDO2GMK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK/action/storage_attestation","attest_author":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK/action/author_attestation","sign_citation":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK/action/citation_signature","submit_replication":"https://pith.science/pith/MFNARZR7EENZ4FTMQI3CDO2GMK/action/replication_record"}},"created_at":"2026-07-05T07:06:27.639068+00:00","updated_at":"2026-07-05T07:06:27.639068+00:00"}