{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:72ACJEDMQRHOO2JVCUDG2M2WZK","short_pith_number":"pith:72ACJEDM","schema_version":"1.0","canonical_sha256":"fe8024906c844ee7693515066d3356caa677d1f5fae77070f9e460f3c1f4a436","source":{"kind":"arxiv","id":"2605.15961","version":1},"attestation_state":"computed","paper":{"title":"Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Ankit Sonthalia, Arnas Uselis, Fabian Morelli, Seong Joon Oh","submitted_at":"2026-05-15T13:54:21Z","abstract_excerpt":"Large-scale pre-trained vision-language models like CLIP demonstrate remarkable zero-shot performance across diverse tasks. However, fine-tuning these models to improve downstream performance often degrades robustness against distribution shifts. Recent approaches have attempted to mitigate this trade-off, but often rely on computationally expensive text-guidance. We propose a novel method for robust fine-tuning, SAE-FT, which operates only on the model's visual representations. SAE-FT regularizes changes to these representations by penalizing the addition and removal of semantically meaningfu"},"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":"2605.15961","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-05-15T13:54:21Z","cross_cats_sorted":[],"title_canon_sha256":"a035f534f25ed218da3145e5da2d20bf8eb021b44527f8ff1104ce259388294e","abstract_canon_sha256":"57b65d836c5cc4f2deebd6b7aa3bcfc98410f4b540c53c611fbc7751b16eff1a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:01:46.725127Z","signature_b64":"yqjeC2IcJQd/hDmAYSp7tcBZREWklb1BtvCwChKd5Tm9tn2EXCvpoRGPljnrJdq+CTv96Qet90MEgMfC8mKAAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fe8024906c844ee7693515066d3356caa677d1f5fae77070f9e460f3c1f4a436","last_reissued_at":"2026-05-20T00:01:46.724636Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:01:46.724636Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Ankit Sonthalia, Arnas Uselis, Fabian Morelli, Seong Joon Oh","submitted_at":"2026-05-15T13:54:21Z","abstract_excerpt":"Large-scale pre-trained vision-language models like CLIP demonstrate remarkable zero-shot performance across diverse tasks. However, fine-tuning these models to improve downstream performance often degrades robustness against distribution shifts. Recent approaches have attempted to mitigate this trade-off, but often rely on computationally expensive text-guidance. We propose a novel method for robust fine-tuning, SAE-FT, which operates only on the model's visual representations. SAE-FT regularizes changes to these representations by penalizing the addition and removal of semantically meaningfu"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.15961","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/2605.15961/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T17:33:44.875206Z","status":"skipped","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T17:01:55.701666Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"cf4636f586bd86dd0a8edabadd0dfcf46b24b86818a29e8d0d64817ecd2dbefe"},"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":"2605.15961","created_at":"2026-05-20T00:01:46.724715+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.15961v1","created_at":"2026-05-20T00:01:46.724715+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.15961","created_at":"2026-05-20T00:01:46.724715+00:00"},{"alias_kind":"pith_short_12","alias_value":"72ACJEDMQRHO","created_at":"2026-05-20T00:01:46.724715+00:00"},{"alias_kind":"pith_short_16","alias_value":"72ACJEDMQRHOO2JV","created_at":"2026-05-20T00:01:46.724715+00:00"},{"alias_kind":"pith_short_8","alias_value":"72ACJEDM","created_at":"2026-05-20T00:01:46.724715+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/72ACJEDMQRHOO2JVCUDG2M2WZK","json":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK.json","graph_json":"https://pith.science/api/pith-number/72ACJEDMQRHOO2JVCUDG2M2WZK/graph.json","events_json":"https://pith.science/api/pith-number/72ACJEDMQRHOO2JVCUDG2M2WZK/events.json","paper":"https://pith.science/paper/72ACJEDM"},"agent_actions":{"view_html":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK","download_json":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK.json","view_paper":"https://pith.science/paper/72ACJEDM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.15961&json=true","fetch_graph":"https://pith.science/api/pith-number/72ACJEDMQRHOO2JVCUDG2M2WZK/graph.json","fetch_events":"https://pith.science/api/pith-number/72ACJEDMQRHOO2JVCUDG2M2WZK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK/action/storage_attestation","attest_author":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK/action/author_attestation","sign_citation":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK/action/citation_signature","submit_replication":"https://pith.science/pith/72ACJEDMQRHOO2JVCUDG2M2WZK/action/replication_record"}},"created_at":"2026-05-20T00:01:46.724715+00:00","updated_at":"2026-05-20T00:01:46.724715+00:00"}