{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:JI4JF7MPL42LLBBHIIEEFDSWV4","short_pith_number":"pith:JI4JF7MP","schema_version":"1.0","canonical_sha256":"4a3892fd8f5f34b584274208428e56af37283a1243e10100dd73ddc259f0196c","source":{"kind":"arxiv","id":"2411.10950","version":2},"attestation_state":"computed","paper":{"title":"Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Sophia Ananiadou, Zeping Yu","submitted_at":"2024-11-17T03:32:50Z","abstract_excerpt":"Understanding the mechanisms behind Large Language Models (LLMs) is crucial for designing improved models and strategies. While recent studies have yielded valuable insights into the mechanisms of textual LLMs, the mechanisms of Multi-modal Large Language Models (MLLMs) remain underexplored. In this paper, we apply mechanistic interpretability methods to analyze the visual question answering (VQA) mechanisms in the first MLLM, Llava. We compare the mechanisms between VQA and textual QA (TQA) in color answering tasks and find that: a) VQA exhibits a mechanism similar to the in-context learning "},"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":"2411.10950","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-11-17T03:32:50Z","cross_cats_sorted":[],"title_canon_sha256":"83dfb4498124283dcb14e345c948c7e146764845467160690ac33daec0178482","abstract_canon_sha256":"e2a112cff15cd5e64c04d999298e7942440a7f21c0b2cff5516f61a40890e67b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:59:55.391057Z","signature_b64":"yDwks4HSGJua4iZug5/rJYttg66qOGzrWseKAmMbRzYZ1Qs48zYFDszvbdgdT5wzqLi7GjQII1HhK0BkrHAQCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4a3892fd8f5f34b584274208428e56af37283a1243e10100dd73ddc259f0196c","last_reissued_at":"2026-07-05T09:59:55.390594Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:59:55.390594Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Sophia Ananiadou, Zeping Yu","submitted_at":"2024-11-17T03:32:50Z","abstract_excerpt":"Understanding the mechanisms behind Large Language Models (LLMs) is crucial for designing improved models and strategies. While recent studies have yielded valuable insights into the mechanisms of textual LLMs, the mechanisms of Multi-modal Large Language Models (MLLMs) remain underexplored. In this paper, we apply mechanistic interpretability methods to analyze the visual question answering (VQA) mechanisms in the first MLLM, Llava. We compare the mechanisms between VQA and textual QA (TQA) in color answering tasks and find that: a) VQA exhibits a mechanism similar to the in-context learning "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.10950","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/2411.10950/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":"2411.10950","created_at":"2026-07-05T09:59:55.390652+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.10950v2","created_at":"2026-07-05T09:59:55.390652+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.10950","created_at":"2026-07-05T09:59:55.390652+00:00"},{"alias_kind":"pith_short_12","alias_value":"JI4JF7MPL42L","created_at":"2026-07-05T09:59:55.390652+00:00"},{"alias_kind":"pith_short_16","alias_value":"JI4JF7MPL42LLBBH","created_at":"2026-07-05T09:59:55.390652+00:00"},{"alias_kind":"pith_short_8","alias_value":"JI4JF7MP","created_at":"2026-07-05T09:59:55.390652+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2603.16284","citing_title":"Locate-then-Sparsify: Attribution Guided Sparse Strategy for Visual Hallucination Mitigation","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2601.09298","citing_title":"Multi-Modal LLM based Image Captioning in ICT: Bridging the Gap Between General and Industry Domain","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05909","citing_title":"Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17941","citing_title":"From Heads to Neurons: Causal Attribution and Steering in Multi-Task Vision-Language Models","ref_index":56,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4","json":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4.json","graph_json":"https://pith.science/api/pith-number/JI4JF7MPL42LLBBHIIEEFDSWV4/graph.json","events_json":"https://pith.science/api/pith-number/JI4JF7MPL42LLBBHIIEEFDSWV4/events.json","paper":"https://pith.science/paper/JI4JF7MP"},"agent_actions":{"view_html":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4","download_json":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4.json","view_paper":"https://pith.science/paper/JI4JF7MP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.10950&json=true","fetch_graph":"https://pith.science/api/pith-number/JI4JF7MPL42LLBBHIIEEFDSWV4/graph.json","fetch_events":"https://pith.science/api/pith-number/JI4JF7MPL42LLBBHIIEEFDSWV4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4/action/storage_attestation","attest_author":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4/action/author_attestation","sign_citation":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4/action/citation_signature","submit_replication":"https://pith.science/pith/JI4JF7MPL42LLBBHIIEEFDSWV4/action/replication_record"}},"created_at":"2026-07-05T09:59:55.390652+00:00","updated_at":"2026-07-05T09:59:55.390652+00:00"}