{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:P25W5XUJLCEYCUNCAY7MEGIFHG","short_pith_number":"pith:P25W5XUJ","schema_version":"1.0","canonical_sha256":"7ebb6ede8958898151a2063ec2190539ab959a8524351569bb2f05d5f38b9d9b","source":{"kind":"arxiv","id":"2608.09928","version":1},"attestation_state":"computed","paper":{"title":"Multimodal Model Diffing for Feature Discovery and Control","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Ashkan Khakzar, Christian Schroeder de Witt, Constantin Venhoff, Hunar Batra, Lachin Naghashyar, Philip Torr, Ronald Clark","submitted_at":"2026-08-10T17:59:30Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering a"},"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":"2608.09928","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-08-10T17:59:30Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"3040544af5d8537e0c50e6fab7a1d7ba037202478117997e904a3b558ef03aa5","abstract_canon_sha256":"71fe550a65d9b16c4c24f33b1c525606318ae45d38ef24dc1d13e841933b8630"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-11T02:25:28.862626Z","signature_b64":"8YhSuUXYpgUHtx25n7tENVksxQf5DFqEX4XVCgu6NU3Qh7YPWcFG9z50i3c5xvQ/Q82GLB2oRSHK3yJXLGAXDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7ebb6ede8958898151a2063ec2190539ab959a8524351569bb2f05d5f38b9d9b","last_reissued_at":"2026-08-11T02:25:28.861018Z","signature_status":"signed_v1","first_computed_at":"2026-08-11T02:25:28.861018Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Multimodal Model Diffing for Feature Discovery and Control","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Ashkan Khakzar, Christian Schroeder de Witt, Constantin Venhoff, Hunar Batra, Lachin Naghashyar, Philip Torr, Ronald Clark","submitted_at":"2026-08-10T17:59:30Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.09928","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/2608.09928/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":"2608.09928","created_at":"2026-08-11T02:25:28.861609+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.09928v1","created_at":"2026-08-11T02:25:28.861609+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.09928","created_at":"2026-08-11T02:25:28.861609+00:00"},{"alias_kind":"pith_short_12","alias_value":"P25W5XUJLCEY","created_at":"2026-08-11T02:25:28.861609+00:00"},{"alias_kind":"pith_short_16","alias_value":"P25W5XUJLCEYCUNC","created_at":"2026-08-11T02:25:28.861609+00:00"},{"alias_kind":"pith_short_8","alias_value":"P25W5XUJ","created_at":"2026-08-11T02:25:28.861609+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/P25W5XUJLCEYCUNCAY7MEGIFHG","json":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG.json","graph_json":"https://pith.science/api/pith-number/P25W5XUJLCEYCUNCAY7MEGIFHG/graph.json","events_json":"https://pith.science/api/pith-number/P25W5XUJLCEYCUNCAY7MEGIFHG/events.json","paper":"https://pith.science/paper/P25W5XUJ"},"agent_actions":{"view_html":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG","download_json":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG.json","view_paper":"https://pith.science/paper/P25W5XUJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.09928&json=true","fetch_graph":"https://pith.science/api/pith-number/P25W5XUJLCEYCUNCAY7MEGIFHG/graph.json","fetch_events":"https://pith.science/api/pith-number/P25W5XUJLCEYCUNCAY7MEGIFHG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG/action/storage_attestation","attest_author":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG/action/author_attestation","sign_citation":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG/action/citation_signature","submit_replication":"https://pith.science/pith/P25W5XUJLCEYCUNCAY7MEGIFHG/action/replication_record"}},"created_at":"2026-08-11T02:25:28.861609+00:00","updated_at":"2026-08-11T02:25:28.861609+00:00"}