{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:KK4FGJ3IIGDYUTD47K3AGRDZXR","short_pith_number":"pith:KK4FGJ3I","schema_version":"1.0","canonical_sha256":"52b853276841878a4c7cfab6034479bc6dccfe844f53838ea7fae36bf554917e","source":{"kind":"arxiv","id":"2508.20279","version":1},"attestation_state":"computed","paper":{"title":"How Multimodal LLMs Solve Image Tasks: A Lens on Visual Grounding, Task Reasoning, and Answer Decoding","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.CV","authors_text":"Yong Jae Lee, Zhuoran Yu","submitted_at":"2025-08-27T21:22:01Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) have demonstrated strong performance across a wide range of vision-language tasks, yet their internal processing dynamics remain underexplored. In this work, we introduce a probing framework to systematically analyze how MLLMs process visual and textual inputs across layers. We train linear classifiers to predict fine-grained visual categories (e.g., dog breeds) from token embeddings extracted at each layer, using a standardized anchor question. To uncover the functional roles of different layers, we evaluate these probes under three types of controlled"},"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":"2508.20279","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-08-27T21:22:01Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"002023e19f810158d6944db945b8f6ac8d82143619fef49e164eff2b3941a138","abstract_canon_sha256":"8204e0a2935430c7e2f3b6030a9833c7108180a6f5bdd23b7d20692aea59f265"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:00:33.471806Z","signature_b64":"h7eISUHcWl8e5b9YZkWdRsfpvLBkdBkxq16CcZDy7HW9lUOK5dmsToAb7W8dII/IYdW7m+I/j8daHSoV6DkODA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"52b853276841878a4c7cfab6034479bc6dccfe844f53838ea7fae36bf554917e","last_reissued_at":"2026-07-05T12:00:33.471300Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:00:33.471300Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How Multimodal LLMs Solve Image Tasks: A Lens on Visual Grounding, Task Reasoning, and Answer Decoding","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.CV","authors_text":"Yong Jae Lee, Zhuoran Yu","submitted_at":"2025-08-27T21:22:01Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) have demonstrated strong performance across a wide range of vision-language tasks, yet their internal processing dynamics remain underexplored. In this work, we introduce a probing framework to systematically analyze how MLLMs process visual and textual inputs across layers. We train linear classifiers to predict fine-grained visual categories (e.g., dog breeds) from token embeddings extracted at each layer, using a standardized anchor question. To uncover the functional roles of different layers, we evaluate these probes under three types of controlled"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.20279","kind":"arxiv","version":1},"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/2508.20279/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":"2508.20279","created_at":"2026-07-05T12:00:33.471366+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.20279v1","created_at":"2026-07-05T12:00:33.471366+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.20279","created_at":"2026-07-05T12:00:33.471366+00:00"},{"alias_kind":"pith_short_12","alias_value":"KK4FGJ3IIGDY","created_at":"2026-07-05T12:00:33.471366+00:00"},{"alias_kind":"pith_short_16","alias_value":"KK4FGJ3IIGDYUTD4","created_at":"2026-07-05T12:00:33.471366+00:00"},{"alias_kind":"pith_short_8","alias_value":"KK4FGJ3I","created_at":"2026-07-05T12:00:33.471366+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08009","citing_title":"From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs","ref_index":198,"is_internal_anchor":true},{"citing_arxiv_id":"2606.20077","citing_title":"The Hidden Evolution of Disguised Visual Context inside the VLM","ref_index":75,"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":110,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR","json":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR.json","graph_json":"https://pith.science/api/pith-number/KK4FGJ3IIGDYUTD47K3AGRDZXR/graph.json","events_json":"https://pith.science/api/pith-number/KK4FGJ3IIGDYUTD47K3AGRDZXR/events.json","paper":"https://pith.science/paper/KK4FGJ3I"},"agent_actions":{"view_html":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR","download_json":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR.json","view_paper":"https://pith.science/paper/KK4FGJ3I","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.20279&json=true","fetch_graph":"https://pith.science/api/pith-number/KK4FGJ3IIGDYUTD47K3AGRDZXR/graph.json","fetch_events":"https://pith.science/api/pith-number/KK4FGJ3IIGDYUTD47K3AGRDZXR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR/action/storage_attestation","attest_author":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR/action/author_attestation","sign_citation":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR/action/citation_signature","submit_replication":"https://pith.science/pith/KK4FGJ3IIGDYUTD47K3AGRDZXR/action/replication_record"}},"created_at":"2026-07-05T12:00:33.471366+00:00","updated_at":"2026-07-05T12:00:33.471366+00:00"}