{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VJVXNYRZUBOQIHTIZZTYYEWERQ","short_pith_number":"pith:VJVXNYRZ","schema_version":"1.0","canonical_sha256":"aa6b76e239a05d041e68ce678c12c48c1e5b965e45654d4b68aeb4f68fb0bb1e","source":{"kind":"arxiv","id":"2502.04959","version":3},"attestation_state":"computed","paper":{"title":"No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Andrew D. Bagdanov, Bart{\\l}omiej Twardowski, Daniel Marczak, Joost Van De Weijer, Sebastian Cygert, Simone Magistri","submitted_at":"2025-02-07T14:22:56Z","abstract_excerpt":"Model merging integrates the weights of multiple task-specific models into a single multi-task model. Despite recent interest in the problem, a significant performance gap between the combined and single-task models remains. In this paper, we investigate the key characteristics of task matrices -- weight update matrices applied to a pre-trained model -- that enable effective merging. We show that alignment between singular components of task-specific and merged matrices strongly correlates with performance improvement over the pre-trained model. Based on this, we propose an isotropic merging f"},"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":"2502.04959","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-07T14:22:56Z","cross_cats_sorted":[],"title_canon_sha256":"56b357dd68ae610a2b00cc25b51ca5a7cf4b70daf922f8f2ad40b2a9bf74526c","abstract_canon_sha256":"f7f6c79a0915eb379cd2bf35daee8f189263056a64bda04a457d78ad58148bbe"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:19:51.173970Z","signature_b64":"IMtLJQkwuCk1e81YJoJuCk5IppYhE7mYpWRt4RwzMnTohJSvsoCBGJx7YRM6Nnj0KRWCm6n22SKp/M9ZERmJBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"aa6b76e239a05d041e68ce678c12c48c1e5b965e45654d4b68aeb4f68fb0bb1e","last_reissued_at":"2026-07-05T11:19:51.173472Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:19:51.173472Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Andrew D. Bagdanov, Bart{\\l}omiej Twardowski, Daniel Marczak, Joost Van De Weijer, Sebastian Cygert, Simone Magistri","submitted_at":"2025-02-07T14:22:56Z","abstract_excerpt":"Model merging integrates the weights of multiple task-specific models into a single multi-task model. Despite recent interest in the problem, a significant performance gap between the combined and single-task models remains. In this paper, we investigate the key characteristics of task matrices -- weight update matrices applied to a pre-trained model -- that enable effective merging. We show that alignment between singular components of task-specific and merged matrices strongly correlates with performance improvement over the pre-trained model. Based on this, we propose an isotropic merging f"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.04959","kind":"arxiv","version":3},"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/2502.04959/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":"2502.04959","created_at":"2026-07-05T11:19:51.173547+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.04959v3","created_at":"2026-07-05T11:19:51.173547+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.04959","created_at":"2026-07-05T11:19:51.173547+00:00"},{"alias_kind":"pith_short_12","alias_value":"VJVXNYRZUBOQ","created_at":"2026-07-05T11:19:51.173547+00:00"},{"alias_kind":"pith_short_16","alias_value":"VJVXNYRZUBOQIHTI","created_at":"2026-07-05T11:19:51.173547+00:00"},{"alias_kind":"pith_short_8","alias_value":"VJVXNYRZ","created_at":"2026-07-05T11:19:51.173547+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18627","citing_title":"PACT: Preserving Anchored Cores in Task-vectors for Model Merging","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02252","citing_title":"ResMerge: Residual-based Spectral Merging of Large Language Models","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15622","citing_title":"Position: Zeroth-Order Optimization in Deep Learning Is Underexplored, Not Underpowered","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22823","citing_title":"PivotMerge: Bridging Heterogeneous Multimodal Pre-training via Post-Alignment Model Merging","ref_index":38,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ","json":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ.json","graph_json":"https://pith.science/api/pith-number/VJVXNYRZUBOQIHTIZZTYYEWERQ/graph.json","events_json":"https://pith.science/api/pith-number/VJVXNYRZUBOQIHTIZZTYYEWERQ/events.json","paper":"https://pith.science/paper/VJVXNYRZ"},"agent_actions":{"view_html":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ","download_json":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ.json","view_paper":"https://pith.science/paper/VJVXNYRZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.04959&json=true","fetch_graph":"https://pith.science/api/pith-number/VJVXNYRZUBOQIHTIZZTYYEWERQ/graph.json","fetch_events":"https://pith.science/api/pith-number/VJVXNYRZUBOQIHTIZZTYYEWERQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ/action/storage_attestation","attest_author":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ/action/author_attestation","sign_citation":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ/action/citation_signature","submit_replication":"https://pith.science/pith/VJVXNYRZUBOQIHTIZZTYYEWERQ/action/replication_record"}},"created_at":"2026-07-05T11:19:51.173547+00:00","updated_at":"2026-07-05T11:19:51.173547+00:00"}