{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2025:L6H64TZ4NHTS5VK4WN2JW2BYKY","short_pith_number":"pith:L6H64TZ4","canonical_record":{"source":{"id":"2505.24037","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-29T22:17:43Z","cross_cats_sorted":[],"title_canon_sha256":"e074ef0c95817291d61384c6b7b4afec2771abb9eacd50a73a1e73cf740346a3","abstract_canon_sha256":"f8602e5855c256345cdb1818a2157d0dd502ecfab7f30660674ec855430737b3"},"schema_version":"1.0"},"canonical_sha256":"5f8fee4f3c69e72ed55cb3749b683856397e380e90101554995c173536711ada","source":{"kind":"arxiv","id":"2505.24037","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2505.24037","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"arxiv_version","alias_value":"2505.24037v3","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.24037","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_12","alias_value":"L6H64TZ4NHTS","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_16","alias_value":"L6H64TZ4NHTS5VK4","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_8","alias_value":"L6H64TZ4","created_at":"2026-06-03T01:05:44Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2025:L6H64TZ4NHTS5VK4WN2JW2BYKY","target":"record","payload":{"canonical_record":{"source":{"id":"2505.24037","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-29T22:17:43Z","cross_cats_sorted":[],"title_canon_sha256":"e074ef0c95817291d61384c6b7b4afec2771abb9eacd50a73a1e73cf740346a3","abstract_canon_sha256":"f8602e5855c256345cdb1818a2157d0dd502ecfab7f30660674ec855430737b3"},"schema_version":"1.0"},"canonical_sha256":"5f8fee4f3c69e72ed55cb3749b683856397e380e90101554995c173536711ada","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-03T01:05:44.151733Z","signature_b64":"1o6kDCw6evYMiaRFGeHPwruEpVl11en5Ukfyr4L7va8KpIzHbB+g4klY/HgSzSeslL98Axqinyib7QoukrzqCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5f8fee4f3c69e72ed55cb3749b683856397e380e90101554995c173536711ada","last_reissued_at":"2026-06-03T01:05:44.151149Z","signature_status":"signed_v1","first_computed_at":"2026-06-03T01:05:44.151149Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2505.24037","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-03T01:05:44Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"QihetRZcOJgyxRSq1hD3V2zBmSD1I1Ft2QcJOE4MlEUVKKdWWpWDR4N36r7eqOJz7wLePbz9Jxc/bALLbXzUCA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-25T14:18:41.688040Z"},"content_sha256":"d31db61622d2158d56dad16d3025d3ef943f75c9e13351da523e7230e448a2a1","schema_version":"1.0","event_id":"sha256:d31db61622d2158d56dad16d3025d3ef943f75c9e13351da523e7230e448a2a1"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2025:L6H64TZ4NHTS5VK4WN2JW2BYKY","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Alan Ansell, Boqian Wu, Decebal Constantin Mocanu, Lu Yin, Mykola Pechenizkiy, Qiao Xiao, Shiwei Liu","submitted_at":"2025-05-29T22:17:43Z","abstract_excerpt":"Sparse large language models (LLMs) offer an attractive direction toward efficient deployment, but adapting them to downstream tasks remains challenging. The central difficulty is to enable effective task adaptation without sacrificing the efficiency advantages of sparsity. Existing fine-tuning methods are not well-suited to this setting, as they either introduce additional dense parameters or assume a fixed sparse topology, limiting their compatibility with sparse LLMs. In this paper, we propose Sparsity Evolution Fine-Tuning (SEFT), a fine-tuning framework designed specifically for sparse LL"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our experiments on various LLMs, including LLaMA families, DeepSeek, and Mistral, across a diverse set of benchmarks demonstrate that SEFT achieves stronger performance while offering superior memory and time efficiency compared to existing baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The sensitivity-driven pruning criterion can maintain the target sparsity level throughout fine-tuning while the drop-and-grow strategy successfully adapts the sparse topology to the target dataset without introducing instability or performance collapse.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SEFT dynamically adjusts sparse connections in pruned LLMs via weight drop-and-grow and sensitivity-driven pruning to adapt to tasks while preserving sparsity level.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2399b9aaa30a0f13b5d34192abe6ea3f8d7df4340c7601e11cd3d65a4660468c"},"source":{"id":"2505.24037","kind":"arxiv","version":3},"verdict":{"id":"bf57f0ad-4003-4dbd-b414-1418ed4ee413","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T12:08:13.333131Z","strongest_claim":"Our experiments on various LLMs, including LLaMA families, DeepSeek, and Mistral, across a diverse set of benchmarks demonstrate that SEFT achieves stronger performance while offering superior memory and time efficiency compared to existing baselines.","one_line_summary":"SEFT dynamically adjusts sparse connections in pruned LLMs via weight drop-and-grow and sensitivity-driven pruning to adapt to tasks while preserving sparsity level.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The sensitivity-driven pruning criterion can maintain the target sparsity level throughout fine-tuning while the drop-and-grow strategy successfully adapts the sparse topology to the target dataset without introducing instability or performance collapse.","pith_extraction_headline":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2505.24037/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"},"verdict_id":"bf57f0ad-4003-4dbd-b414-1418ed4ee413"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-03T01:05:44Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"JUoEKHZpsZBsHVHlzI1QJK5BJEEXPZM4AI1xdob7NoNo2FZVYz4CUj+J4IXfMNxnfKcf+qkwgHwaSkYXAAVnBA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-25T14:18:41.688511Z"},"content_sha256":"deedea3cf688bd672d4c2ac23c72083f9ab814bbaafe4c7e1817de8660760bf8","schema_version":"1.0","event_id":"sha256:deedea3cf688bd672d4c2ac23c72083f9ab814bbaafe4c7e1817de8660760bf8"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/bundle.json","state_url":"https://pith.science/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-25T14:18:41Z","links":{"resolver":"https://pith.science/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY","bundle":"https://pith.science/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/bundle.json","state":"https://pith.science/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/state.json","well_known_bundle":"https://pith.science/.well-known/pith/L6H64TZ4NHTS5VK4WN2JW2BYKY/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:L6H64TZ4NHTS5VK4WN2JW2BYKY","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"f8602e5855c256345cdb1818a2157d0dd502ecfab7f30660674ec855430737b3","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-29T22:17:43Z","title_canon_sha256":"e074ef0c95817291d61384c6b7b4afec2771abb9eacd50a73a1e73cf740346a3"},"schema_version":"1.0","source":{"id":"2505.24037","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2505.24037","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"arxiv_version","alias_value":"2505.24037v3","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.24037","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_12","alias_value":"L6H64TZ4NHTS","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_16","alias_value":"L6H64TZ4NHTS5VK4","created_at":"2026-06-03T01:05:44Z"},{"alias_kind":"pith_short_8","alias_value":"L6H64TZ4","created_at":"2026-06-03T01:05:44Z"}],"graph_snapshots":[{"event_id":"sha256:deedea3cf688bd672d4c2ac23c72083f9ab814bbaafe4c7e1817de8660760bf8","target":"graph","created_at":"2026-06-03T01:05:44Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Our experiments on various LLMs, including LLaMA families, DeepSeek, and Mistral, across a diverse set of benchmarks demonstrate that SEFT achieves stronger performance while offering superior memory and time efficiency compared to existing baselines."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The sensitivity-driven pruning criterion can maintain the target sparsity level throughout fine-tuning while the drop-and-grow strategy successfully adapts the sparse topology to the target dataset without introducing instability or performance collapse."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"SEFT dynamically adjusts sparse connections in pruned LLMs via weight drop-and-grow and sensitivity-driven pruning to adapt to tasks while preserving sparsity level."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient."}],"snapshot_sha256":"2399b9aaa30a0f13b5d34192abe6ea3f8d7df4340c7601e11cd3d65a4660468c"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2505.24037/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Sparse large language models (LLMs) offer an attractive direction toward efficient deployment, but adapting them to downstream tasks remains challenging. The central difficulty is to enable effective task adaptation without sacrificing the efficiency advantages of sparsity. Existing fine-tuning methods are not well-suited to this setting, as they either introduce additional dense parameters or assume a fixed sparse topology, limiting their compatibility with sparse LLMs. In this paper, we propose Sparsity Evolution Fine-Tuning (SEFT), a fine-tuning framework designed specifically for sparse LL","authors_text":"Alan Ansell, Boqian Wu, Decebal Constantin Mocanu, Lu Yin, Mykola Pechenizkiy, Qiao Xiao, Shiwei Liu","cross_cats":[],"headline":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-29T22:17:43Z","title":"Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.24037","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-19T12:08:13.333131Z","id":"bf57f0ad-4003-4dbd-b414-1418ed4ee413","model_set":{"reader":"grok-4.3"},"one_line_summary":"SEFT dynamically adjusts sparse connections in pruned LLMs via weight drop-and-grow and sensitivity-driven pruning to adapt to tasks while preserving sparsity level.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Dynamically evolving the sparse connections of pruned LLMs during fine-tuning recovers performance lost to pruning while keeping models efficient.","strongest_claim":"Our experiments on various LLMs, including LLaMA families, DeepSeek, and Mistral, across a diverse set of benchmarks demonstrate that SEFT achieves stronger performance while offering superior memory and time efficiency compared to existing baselines.","weakest_assumption":"The sensitivity-driven pruning criterion can maintain the target sparsity level throughout fine-tuning while the drop-and-grow strategy successfully adapts the sparse topology to the target dataset without introducing instability or performance collapse."}},"verdict_id":"bf57f0ad-4003-4dbd-b414-1418ed4ee413"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:d31db61622d2158d56dad16d3025d3ef943f75c9e13351da523e7230e448a2a1","target":"record","created_at":"2026-06-03T01:05:44Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"f8602e5855c256345cdb1818a2157d0dd502ecfab7f30660674ec855430737b3","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-29T22:17:43Z","title_canon_sha256":"e074ef0c95817291d61384c6b7b4afec2771abb9eacd50a73a1e73cf740346a3"},"schema_version":"1.0","source":{"id":"2505.24037","kind":"arxiv","version":3}},"canonical_sha256":"5f8fee4f3c69e72ed55cb3749b683856397e380e90101554995c173536711ada","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"5f8fee4f3c69e72ed55cb3749b683856397e380e90101554995c173536711ada","first_computed_at":"2026-06-03T01:05:44.151149Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-03T01:05:44.151149Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"1o6kDCw6evYMiaRFGeHPwruEpVl11en5Ukfyr4L7va8KpIzHbB+g4klY/HgSzSeslL98Axqinyib7QoukrzqCg==","signature_status":"signed_v1","signed_at":"2026-06-03T01:05:44.151733Z","signed_message":"canonical_sha256_bytes"},"source_id":"2505.24037","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:d31db61622d2158d56dad16d3025d3ef943f75c9e13351da523e7230e448a2a1","sha256:deedea3cf688bd672d4c2ac23c72083f9ab814bbaafe4c7e1817de8660760bf8"],"state_sha256":"079285b385056ff3e6c97c5a27498eb24a410aafd447e98a7be1029a5c3b041c"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"1zMmpg0uldjyWC/+rXjCvPtdqpdwh2OC/nCqinlBmcqyBCUHaKH6gsoqgx5bTS8uTO9N+hQJ11V354JNjPaJBw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-25T14:18:41.690914Z","bundle_sha256":"5818e689efe4a533dadd6777b7ed0c360cbf4f9740cd9d7cda6d18ece92890ae"}}