{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:QIJS7POPXR6RYWPL2KUE7Y3VYC","short_pith_number":"pith:QIJS7POP","schema_version":"1.0","canonical_sha256":"82132fbdcfbc7d1c59ebd2a84fe375c09ce915c88ee618fca9b944316fd2c959","source":{"kind":"arxiv","id":"2605.13277","version":1},"attestation_state":"computed","paper":{"title":"Utility-Oriented Visual Evidence Selection for Multimodal Retrieval-Augmented Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Ranking visual evidence by information gain on a latent helpfulness variable matches its answer-space utility in multimodal RAG","cross_cats":["cs.AI","cs.CV","cs.IR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Haofeng Zhang, Weiqing Luo, Xiao Wang, Zhiyuan Yu, Ziyi Huang, Zongye Hu","submitted_at":"2026-05-13T09:54:31Z","abstract_excerpt":"Visual evidence selection is a critical component of multimodal retrieval-augmented generation (RAG), yet existing methods typically rely on semantic relevance or surface-level similarity, which are often misaligned with the actual utility of visual evidence for downstream reasoning. We reformulate multimodal evidence selection from an information-theoretic perspective by defining evidence utility as the information gain induced on a model's output distribution. To overcome the intractability of answer-space optimization, we introduce a latent notion of evidence helpfulness and theoretically s"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13277","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T09:54:31Z","cross_cats_sorted":["cs.AI","cs.CV","cs.IR","cs.LG"],"title_canon_sha256":"9274f354ba30343f0b53f3f8d3261e368967506e66d00baef791b534caabfefa","abstract_canon_sha256":"a0eb75b8a0c660ed2a4f81960274c6969fbfa096c82856ae7542a61892f5e7d0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:44:49.229283Z","signature_b64":"9ivqBOo/Hszp01f7pKiKn3UcoEo1bGmRz1ia9RDJ48C/OK/LFLAM7BgHYTr80eqoC6Y7/BraZHjnNO7er9+gCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"82132fbdcfbc7d1c59ebd2a84fe375c09ce915c88ee618fca9b944316fd2c959","last_reissued_at":"2026-05-18T02:44:49.228866Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:44:49.228866Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Utility-Oriented Visual Evidence Selection for Multimodal Retrieval-Augmented Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Ranking visual evidence by information gain on a latent helpfulness variable matches its answer-space utility in multimodal RAG","cross_cats":["cs.AI","cs.CV","cs.IR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Haofeng Zhang, Weiqing Luo, Xiao Wang, Zhiyuan Yu, Ziyi Huang, Zongye Hu","submitted_at":"2026-05-13T09:54:31Z","abstract_excerpt":"Visual evidence selection is a critical component of multimodal retrieval-augmented generation (RAG), yet existing methods typically rely on semantic relevance or surface-level similarity, which are often misaligned with the actual utility of visual evidence for downstream reasoning. We reformulate multimodal evidence selection from an information-theoretic perspective by defining evidence utility as the information gain induced on a model's output distribution. To overcome the intractability of answer-space optimization, we introduce a latent notion of evidence helpfulness and theoretically s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ranking evidence by information gain on this latent variable is equivalent to answer-space utility","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"under mild assumptions","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Evidence utility is defined as information gain on the model's output distribution, with ranking by gain on a latent helpfulness variable shown equivalent to answer-space utility under mild assumptions, enabling a training-free surrogate framework that outperforms baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Ranking visual evidence by information gain on a latent helpfulness variable matches its answer-space utility in multimodal RAG","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"43252e453d2c92aee16fccff049ee089523766acbfc4bfe88e0dd91c033d5d93"},"source":{"id":"2605.13277","kind":"arxiv","version":1},"verdict":{"id":"79da5243-cbc0-4d46-955a-e9a4fcc5c9ea","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:06:00.400435Z","strongest_claim":"ranking evidence by information gain on this latent variable is equivalent to answer-space utility","one_line_summary":"Evidence utility is defined as information gain on the model's output distribution, with ranking by gain on a latent helpfulness variable shown equivalent to answer-space utility under mild assumptions, enabling a training-free surrogate framework that outperforms baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"under mild assumptions","pith_extraction_headline":"Ranking visual evidence by information gain on a latent helpfulness variable matches its answer-space utility in multimodal RAG"},"references":{"count":43,"sample":[{"doi":"","year":null,"title":"VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks","work_id":"ec376227-cc62-4775-9e03-bed3862151d0","ref_index":1,"cited_arxiv_id":"2410.05160","is_internal_anchor":true},{"doi":"","year":null,"title":"Vlm2vec-v2: Advancing multimodal em- bedding for videos, images, and visual documents","work_id":"645b5239-8642-4bd5-a661-7ff075ceb306","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=","work_id":"f31acd76-d74b-4edb-aed0-20b9256ab082","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"E5-V: Universal Embeddings with Multimodal Large Language Models","work_id":"df2e178e-bd96-48f4-8431-51c61fbc63fd","ref_index":4,"cited_arxiv_id":"2407.12580","is_internal_anchor":true},{"doi":"","year":null,"title":"GME: Improving Universal Multimodal Retrieval by Multimodal","work_id":"a99a6453-de04-4fcb-8376-b10bf7dd68c7","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"69ca431a7ca9f161e76ff13a3e48b127b21d43f61226233ae9f239837e85eda3","internal_anchors":11},"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":"2605.13277","created_at":"2026-05-18T02:44:49.228932+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13277v1","created_at":"2026-05-18T02:44:49.228932+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13277","created_at":"2026-05-18T02:44:49.228932+00:00"},{"alias_kind":"pith_short_12","alias_value":"QIJS7POPXR6R","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"QIJS7POPXR6RYWPL","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"QIJS7POP","created_at":"2026-05-18T12:33:37.589309+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/QIJS7POPXR6RYWPL2KUE7Y3VYC","json":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC.json","graph_json":"https://pith.science/api/pith-number/QIJS7POPXR6RYWPL2KUE7Y3VYC/graph.json","events_json":"https://pith.science/api/pith-number/QIJS7POPXR6RYWPL2KUE7Y3VYC/events.json","paper":"https://pith.science/paper/QIJS7POP"},"agent_actions":{"view_html":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC","download_json":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC.json","view_paper":"https://pith.science/paper/QIJS7POP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13277&json=true","fetch_graph":"https://pith.science/api/pith-number/QIJS7POPXR6RYWPL2KUE7Y3VYC/graph.json","fetch_events":"https://pith.science/api/pith-number/QIJS7POPXR6RYWPL2KUE7Y3VYC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC/action/storage_attestation","attest_author":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC/action/author_attestation","sign_citation":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC/action/citation_signature","submit_replication":"https://pith.science/pith/QIJS7POPXR6RYWPL2KUE7Y3VYC/action/replication_record"}},"created_at":"2026-05-18T02:44:49.228932+00:00","updated_at":"2026-05-18T02:44:49.228932+00:00"}