{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:L6UVHNQFDCEPLHKKCXGBWBPWRE","short_pith_number":"pith:L6UVHNQF","schema_version":"1.0","canonical_sha256":"5fa953b6051888f59d4a15cc1b05f6891b79fdacc898c61c5550c4cd9b67ab40","source":{"kind":"arxiv","id":"2604.23321","version":2},"attestation_state":"computed","paper":{"title":"MMEB-V3: Measuring the Performance Gaps of Omni-Modality Embedding Models","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Multimodal embedding models cannot reliably enforce instructions to retrieve a target modality.","cross_cats":[],"primary_cat":"cs.IR","authors_text":"Haohang Huang, Kai Zou, Mingyi Su, Ping Nie, Rui Meng, Tomas Pfister, Wei Zhang, Wenhu Chen, Xiaoyu Shen, Xuan Lu, Xuan Zhang, Ziyan Jiang","submitted_at":"2026-04-25T14:15:05Z","abstract_excerpt":"Multimodal embedding models aim to map heterogeneous inputs, such as text, images, videos, and audio, into a shared semantic space. However, existing methods and benchmarks remain largely limited to partial modality coverage, making it difficult to systematically evaluate full-modality representation learning. In this work, we take a step toward the full-modality setting. We introduce MMEB-V3, a comprehensive benchmark that evaluates embeddings across text, image, video, audio, as well as agent-centric scenarios. To enable more fine-grained diagnosis, we further construct OmniSET (Omni-modalit"},"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":"2604.23321","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.IR","submitted_at":"2026-04-25T14:15:05Z","cross_cats_sorted":[],"title_canon_sha256":"678ff0458d9da90fb2f680c5e2661b6135e70cdd9e3e7fbe0f4a65a61a516e43","abstract_canon_sha256":"4f743aebd65f4c363e803bfa00b911bb7422b8406f87f1c69069e16f555368ce"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5fa953b6051888f59d4a15cc1b05f6891b79fdacc898c61c5550c4cd9b67ab40","last_reissued_at":"2026-07-30T01:19:51.917729Z","signature_status":"unsigned_v0","first_computed_at":"2026-07-30T01:19:51.917729Z"},"graph_snapshot":{"paper":{"title":"MMEB-V3: Measuring the Performance Gaps of Omni-Modality Embedding Models","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Multimodal embedding models cannot reliably enforce instructions to retrieve a target modality.","cross_cats":[],"primary_cat":"cs.IR","authors_text":"Haohang Huang, Kai Zou, Mingyi Su, Ping Nie, Rui Meng, Tomas Pfister, Wei Zhang, Wenhu Chen, Xiaoyu Shen, Xuan Lu, Xuan Zhang, Ziyan Jiang","submitted_at":"2026-04-25T14:15:05Z","abstract_excerpt":"Multimodal embedding models aim to map heterogeneous inputs, such as text, images, videos, and audio, into a shared semantic space. However, existing methods and benchmarks remain largely limited to partial modality coverage, making it difficult to systematically evaluate full-modality representation learning. In this work, we take a step toward the full-modality setting. We introduce MMEB-V3, a comprehensive benchmark that evaluates embeddings across text, image, video, audio, as well as agent-centric scenarios. To enable more fine-grained diagnosis, we further construct OmniSET (Omni-modalit"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Current multimodal embeddings are not yet capable of reliably enforcing modality constraints specified by instructions, and consequently fail to exhibit consistent modality-aware retrieval behavior.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the OmniSET tuples truly represent semantic equivalence across modalities without residual modality-specific cues that could explain the observed retrieval failures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MMEB-V3 benchmark shows omni-modality embedding models fail to enforce instruction-specified modality constraints and exhibit asymmetric, query-biased retrieval.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Multimodal embedding models cannot reliably enforce instructions to retrieve a target modality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"75d66dd98295e3823bbf0d421697ee30efe51f1e1dd626bdc70e4fb4bf1f3d1d"},"source":{"id":"2604.23321","kind":"arxiv","version":2},"verdict":{"id":"060feda4-d27e-49d7-9138-f5fd1fd11c5d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T07:24:36.472390Z","strongest_claim":"Current multimodal embeddings are not yet capable of reliably enforcing modality constraints specified by instructions, and consequently fail to exhibit consistent modality-aware retrieval behavior.","one_line_summary":"MMEB-V3 benchmark shows omni-modality embedding models fail to enforce instruction-specified modality constraints and exhibit asymmetric, query-biased retrieval.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the OmniSET tuples truly represent semantic equivalence across modalities without residual modality-specific cues that could explain the observed retrieval failures.","pith_extraction_headline":"Multimodal embedding models cannot reliably enforce instructions to retrieve a target modality."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.23321/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-21T09:35:15.146851Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T23:16:00.296303Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"a8fc80721ae5b80c1df92d15597d70ef8c90364e9d2000a28de29fc25830d74e"},"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":"2604.23321","created_at":"2026-07-30T01:19:51.922798+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.23321v2","created_at":"2026-07-30T01:19:51.922798+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.23321","created_at":"2026-07-30T01:19:51.922798+00:00"},{"alias_kind":"pith_short_12","alias_value":"L6UVHNQFDCEP","created_at":"2026-07-30T01:19:51.922798+00:00"},{"alias_kind":"pith_short_16","alias_value":"L6UVHNQFDCEPLHKK","created_at":"2026-07-30T01:19:51.922798+00:00"},{"alias_kind":"pith_short_8","alias_value":"L6UVHNQF","created_at":"2026-07-30T01:19:51.922798+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2606.11700","citing_title":"CompRank: Efficient LLM Reranking via Token-Level Compression and Decoding-Free Scoring","ref_index":10,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE","json":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE.json","graph_json":"https://pith.science/api/pith-number/L6UVHNQFDCEPLHKKCXGBWBPWRE/graph.json","events_json":"https://pith.science/api/pith-number/L6UVHNQFDCEPLHKKCXGBWBPWRE/events.json","paper":"https://pith.science/paper/L6UVHNQF"},"agent_actions":{"view_html":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE","download_json":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE.json","view_paper":"https://pith.science/paper/L6UVHNQF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.23321&json=true","fetch_graph":"https://pith.science/api/pith-number/L6UVHNQFDCEPLHKKCXGBWBPWRE/graph.json","fetch_events":"https://pith.science/api/pith-number/L6UVHNQFDCEPLHKKCXGBWBPWRE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE/action/storage_attestation","attest_author":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE/action/author_attestation","sign_citation":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE/action/citation_signature","submit_replication":"https://pith.science/pith/L6UVHNQFDCEPLHKKCXGBWBPWRE/action/replication_record"}},"created_at":"2026-07-30T01:19:51.922798+00:00","updated_at":"2026-07-30T01:19:51.922798+00:00"}