{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:U2PFYIWF3ES64TFBGJEEMU5TWW","short_pith_number":"pith:U2PFYIWF","schema_version":"1.0","canonical_sha256":"a69e5c22c5d925ee4ca132484653b3b5b225a6def992e0d7ea835dfd87962eaf","source":{"kind":"arxiv","id":"2506.05344","version":2},"attestation_state":"computed","paper":{"title":"SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jiahui Wang, Jiwen Lu, Yongming Rao, Zuyan Liu","submitted_at":"2025-06-05T17:59:55Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this "},"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":"2506.05344","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-06-05T17:59:55Z","cross_cats_sorted":[],"title_canon_sha256":"448e33a4b27fb42c5d4614c8111cc64152dd56190528dcb7aae394b5ebfd7293","abstract_canon_sha256":"43f21a6d6f9965979b0b2b7e4023c7d6a9ad262075d079a29587c68bf4e05c07"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:32:30.220655Z","signature_b64":"hP6sktJHoIFnpf9O6SxeiKcn88kZkfegiOa/lsCEqFJd45B4KqLFvht5BSuJA2NuPLUEprdYkHUamKaT9UKuDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a69e5c22c5d925ee4ca132484653b3b5b225a6def992e0d7ea835dfd87962eaf","last_reissued_at":"2026-07-05T11:32:30.220130Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:32:30.220130Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jiahui Wang, Jiwen Lu, Yongming Rao, Zuyan Liu","submitted_at":"2025-06-05T17:59:55Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.05344","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/2506.05344/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":"2506.05344","created_at":"2026-07-05T11:32:30.220196+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.05344v2","created_at":"2026-07-05T11:32:30.220196+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.05344","created_at":"2026-07-05T11:32:30.220196+00:00"},{"alias_kind":"pith_short_12","alias_value":"U2PFYIWF3ES6","created_at":"2026-07-05T11:32:30.220196+00:00"},{"alias_kind":"pith_short_16","alias_value":"U2PFYIWF3ES64TFB","created_at":"2026-07-05T11:32:30.220196+00:00"},{"alias_kind":"pith_short_8","alias_value":"U2PFYIWF","created_at":"2026-07-05T11:32:30.220196+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2511.22663","citing_title":"AIA: Rethinking Architecture Decoupling Strategy In Unified Multimodal Model","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14710","citing_title":"Vision-Core Guided Contrastive Learning for Balanced Multi-modal Prognosis Prediction of Stroke","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05887","citing_title":"HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW","json":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW.json","graph_json":"https://pith.science/api/pith-number/U2PFYIWF3ES64TFBGJEEMU5TWW/graph.json","events_json":"https://pith.science/api/pith-number/U2PFYIWF3ES64TFBGJEEMU5TWW/events.json","paper":"https://pith.science/paper/U2PFYIWF"},"agent_actions":{"view_html":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW","download_json":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW.json","view_paper":"https://pith.science/paper/U2PFYIWF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.05344&json=true","fetch_graph":"https://pith.science/api/pith-number/U2PFYIWF3ES64TFBGJEEMU5TWW/graph.json","fetch_events":"https://pith.science/api/pith-number/U2PFYIWF3ES64TFBGJEEMU5TWW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW/action/storage_attestation","attest_author":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW/action/author_attestation","sign_citation":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW/action/citation_signature","submit_replication":"https://pith.science/pith/U2PFYIWF3ES64TFBGJEEMU5TWW/action/replication_record"}},"created_at":"2026-07-05T11:32:30.220196+00:00","updated_at":"2026-07-05T11:32:30.220196+00:00"}