{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:NPHMDM3JNPAVX4NMEFIBLAYYJW","short_pith_number":"pith:NPHMDM3J","schema_version":"1.0","canonical_sha256":"6bcec1b3696bc15bf1ac21501583184dad00ec776a08b30d629b90c4ef5e4eb5","source":{"kind":"arxiv","id":"2504.08729","version":1},"attestation_state":"computed","paper":{"title":"Steering CLIP's vision transformer with sparse autoencoders","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Blake Aaron Richards, Danilo Bzdok, Ethan Goldfarb, Lorenz Hufe, Praneet Suresh, Robert Graham, Sonia Joseph, Wojciech Samek, Yossi Gandelsman","submitted_at":"2025-04-11T17:56:09Z","abstract_excerpt":"While vision models are highly capable, their internal mechanisms remain poorly understood -- a challenge which sparse autoencoders (SAEs) have helped address in language, but which remains underexplored in vision. We address this gap by training SAEs on CLIP's vision transformer and uncover key differences between vision and language processing, including distinct sparsity patterns for SAEs trained across layers and token types. We then provide the first systematic analysis on the steerability of CLIP's vision transformer by introducing metrics to quantify how precisely SAE features can be st"},"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":"2504.08729","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-04-11T17:56:09Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"3540603e5628ac1e1d135780fa5b709a2eb135263cebca0c8164569036c9b11f","abstract_canon_sha256":"2f6d0fd2b8c7c53cd242f45c2f51b6e537bc0aa1b0c0e143411093fdf5f9c369"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:47:53.284726Z","signature_b64":"K8NdY4ANTbjogbLTioVTSI3U1iIHwMbRtfivk/CRhhGOm0/xG53Y4kkwIz4iDKfZFEoWkOc6VMc1fYKix/pJCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6bcec1b3696bc15bf1ac21501583184dad00ec776a08b30d629b90c4ef5e4eb5","last_reissued_at":"2026-07-05T10:47:53.284242Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:47:53.284242Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Steering CLIP's vision transformer with sparse autoencoders","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Blake Aaron Richards, Danilo Bzdok, Ethan Goldfarb, Lorenz Hufe, Praneet Suresh, Robert Graham, Sonia Joseph, Wojciech Samek, Yossi Gandelsman","submitted_at":"2025-04-11T17:56:09Z","abstract_excerpt":"While vision models are highly capable, their internal mechanisms remain poorly understood -- a challenge which sparse autoencoders (SAEs) have helped address in language, but which remains underexplored in vision. We address this gap by training SAEs on CLIP's vision transformer and uncover key differences between vision and language processing, including distinct sparsity patterns for SAEs trained across layers and token types. We then provide the first systematic analysis on the steerability of CLIP's vision transformer by introducing metrics to quantify how precisely SAE features can be st"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.08729","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/2504.08729/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":"2504.08729","created_at":"2026-07-05T10:47:53.284303+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.08729v1","created_at":"2026-07-05T10:47:53.284303+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.08729","created_at":"2026-07-05T10:47:53.284303+00:00"},{"alias_kind":"pith_short_12","alias_value":"NPHMDM3JNPAV","created_at":"2026-07-05T10:47:53.284303+00:00"},{"alias_kind":"pith_short_16","alias_value":"NPHMDM3JNPAVX4NM","created_at":"2026-07-05T10:47:53.284303+00:00"},{"alias_kind":"pith_short_8","alias_value":"NPHMDM3J","created_at":"2026-07-05T10:47:53.284303+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18667","citing_title":"Can neurons speak? Semantic narration of vision at single-cell resolution","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08031","citing_title":"Vision-Language Asymmetry in Bistable Image Captioning","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07451","citing_title":"TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignment","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26396","citing_title":"At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15961","citing_title":"Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01247","citing_title":"Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00899","citing_title":"LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11467","citing_title":"From Attribution to Action: A Human-Centered Application of Activation Steering","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08846","citing_title":"Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs","ref_index":32,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW","json":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW.json","graph_json":"https://pith.science/api/pith-number/NPHMDM3JNPAVX4NMEFIBLAYYJW/graph.json","events_json":"https://pith.science/api/pith-number/NPHMDM3JNPAVX4NMEFIBLAYYJW/events.json","paper":"https://pith.science/paper/NPHMDM3J"},"agent_actions":{"view_html":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW","download_json":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW.json","view_paper":"https://pith.science/paper/NPHMDM3J","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.08729&json=true","fetch_graph":"https://pith.science/api/pith-number/NPHMDM3JNPAVX4NMEFIBLAYYJW/graph.json","fetch_events":"https://pith.science/api/pith-number/NPHMDM3JNPAVX4NMEFIBLAYYJW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW/action/storage_attestation","attest_author":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW/action/author_attestation","sign_citation":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW/action/citation_signature","submit_replication":"https://pith.science/pith/NPHMDM3JNPAVX4NMEFIBLAYYJW/action/replication_record"}},"created_at":"2026-07-05T10:47:53.284303+00:00","updated_at":"2026-07-05T10:47:53.284303+00:00"}