{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:GKZOGVERMAA35YKZACZVYHHC6P","short_pith_number":"pith:GKZOGVER","schema_version":"1.0","canonical_sha256":"32b2e354916001bee15900b35c1ce2f3fcef14f8cadcfa487ef03e4d08c0f6a6","source":{"kind":"arxiv","id":"2203.02053","version":2},"attestation_state":"computed","paper":{"title":"Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV","cs.LG","cs.MM"],"primary_cat":"cs.CL","authors_text":"James Zou, Serena Yeung, Weixin Liang, Yongchan Kwon, Yuhui Zhang","submitted_at":"2022-03-03T22:53:54Z","abstract_excerpt":"We present modality gap, an intriguing geometric phenomenon of the representation space of multi-modal models. Specifically, we show that different data modalities (e.g. images and text) are embedded at arm's length in their shared representation in multi-modal models such as CLIP. Our systematic analysis demonstrates that this gap is caused by a combination of model initialization and contrastive learning optimization. In model initialization, we show empirically and theoretically that the representation of a common deep neural network is restricted to a narrow cone. As a consequence, in a mu"},"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":"2203.02053","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2022-03-03T22:53:54Z","cross_cats_sorted":["cs.AI","cs.CV","cs.LG","cs.MM"],"title_canon_sha256":"3e9a87a6e7599f047f725fb63d8b2bae7dad6f7baf89c604bd770c2bea87ebc8","abstract_canon_sha256":"9066e7999cfddcbda6e4733c073983ccd8416122caa7a7c61d95f228219f583e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:08:26.908587Z","signature_b64":"tLm1eVKWqx+E2wL098xLTDMUkfA5rkZOMbKaxzTip34AIxOldZfjTLKNGgZEut1DeY2o4JYEWKbeRMM9mVJABg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"32b2e354916001bee15900b35c1ce2f3fcef14f8cadcfa487ef03e4d08c0f6a6","last_reissued_at":"2026-07-05T05:08:26.907963Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:08:26.907963Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV","cs.LG","cs.MM"],"primary_cat":"cs.CL","authors_text":"James Zou, Serena Yeung, Weixin Liang, Yongchan Kwon, Yuhui Zhang","submitted_at":"2022-03-03T22:53:54Z","abstract_excerpt":"We present modality gap, an intriguing geometric phenomenon of the representation space of multi-modal models. Specifically, we show that different data modalities (e.g. images and text) are embedded at arm's length in their shared representation in multi-modal models such as CLIP. Our systematic analysis demonstrates that this gap is caused by a combination of model initialization and contrastive learning optimization. In model initialization, we show empirically and theoretically that the representation of a common deep neural network is restricted to a narrow cone. As a consequence, in a mu"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2203.02053","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/2203.02053/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":"2203.02053","created_at":"2026-07-05T05:08:26.908101+00:00"},{"alias_kind":"arxiv_version","alias_value":"2203.02053v2","created_at":"2026-07-05T05:08:26.908101+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2203.02053","created_at":"2026-07-05T05:08:26.908101+00:00"},{"alias_kind":"pith_short_12","alias_value":"GKZOGVERMAA3","created_at":"2026-07-05T05:08:26.908101+00:00"},{"alias_kind":"pith_short_16","alias_value":"GKZOGVERMAA35YKZ","created_at":"2026-07-05T05:08:26.908101+00:00"},{"alias_kind":"pith_short_8","alias_value":"GKZOGVER","created_at":"2026-07-05T05:08:26.908101+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11190","citing_title":"When to Align, When to Predict: A Phase Diagram for Multimodal Learning","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08384","citing_title":"jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30625","citing_title":"Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11190","citing_title":"When to Align, When to Predict: A Phase Diagram for Multimodal Learning","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2305.14299","citing_title":"Template-assisted Contrastive Learning of Task-oriented Dialogue Sentence Embeddings","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12872","citing_title":"SMA: Submodular Modality Aligner For Data Efficient Multimodal Learning","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08384","citing_title":"jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11237","citing_title":"DeconDTN-Toolkit: A Library for Evaluation and Enhancement of Robustness to Provenance Shift","ref_index":121,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08384","citing_title":"jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06477","citing_title":"GeoStack: A Framework for Quasi-Abelian Knowledge Composition in VLMs","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14684","citing_title":"DETR-ViP: Detection Transformer with Robust Discriminative Visual Prompts","ref_index":7,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P","json":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P.json","graph_json":"https://pith.science/api/pith-number/GKZOGVERMAA35YKZACZVYHHC6P/graph.json","events_json":"https://pith.science/api/pith-number/GKZOGVERMAA35YKZACZVYHHC6P/events.json","paper":"https://pith.science/paper/GKZOGVER"},"agent_actions":{"view_html":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P","download_json":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P.json","view_paper":"https://pith.science/paper/GKZOGVER","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2203.02053&json=true","fetch_graph":"https://pith.science/api/pith-number/GKZOGVERMAA35YKZACZVYHHC6P/graph.json","fetch_events":"https://pith.science/api/pith-number/GKZOGVERMAA35YKZACZVYHHC6P/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P/action/storage_attestation","attest_author":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P/action/author_attestation","sign_citation":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P/action/citation_signature","submit_replication":"https://pith.science/pith/GKZOGVERMAA35YKZACZVYHHC6P/action/replication_record"}},"created_at":"2026-07-05T05:08:26.908101+00:00","updated_at":"2026-07-05T05:08:26.908101+00:00"}