{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:QJS4RCYKFKQKKNWZDJKU5TZ73A","short_pith_number":"pith:QJS4RCYK","schema_version":"1.0","canonical_sha256":"8265c88b0a2aa0a536d91a554ecf3fd80fa0d74f58fe92884ffb8cd98bfdceb4","source":{"kind":"arxiv","id":"2506.03978","version":2},"attestation_state":"computed","paper":{"title":"Structured Pruning for Diverse Best-of-N Reasoning Optimization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Bao Nguyen, Hieu Trung Nguyen, Viet Anh Nguyen","submitted_at":"2025-06-04T14:08:44Z","abstract_excerpt":"Model pruning in transformer-based language models, traditionally viewed as a means of achieving computational savings, can enhance the model's reasoning capabilities. In this work, we uncover a surprising phenomenon: the selective pruning of certain attention heads leads to improvements in reasoning performance, particularly on challenging tasks. Motivated by this observation, we propose SPRINT, a novel contrastive learning framework that dynamically selects the optimal head and layer to prune during inference. By aligning question embeddings with head embeddings, SPRINT identifies those prun"},"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":"2506.03978","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-06-04T14:08:44Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"ce262a2b4a36ed7af32e91aafbe7e6b6a86a0d1f4cc5e66a761d9e9288b0d90a","abstract_canon_sha256":"92c842b8d46aa9f315deb8fac3793793f575b8737e66a77856a8c5710c4898e0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:18:22.880380Z","signature_b64":"3UO/XIAq6y50/LfT7SCsdwuelVBHS+H1ssml464xV6jX+mv4L/Z/E7uGrTlr83+hzDoe4tWiB4PDrm2IcRhpDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8265c88b0a2aa0a536d91a554ecf3fd80fa0d74f58fe92884ffb8cd98bfdceb4","last_reissued_at":"2026-07-05T11:18:22.879889Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:18:22.879889Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Structured Pruning for Diverse Best-of-N Reasoning Optimization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Bao Nguyen, Hieu Trung Nguyen, Viet Anh Nguyen","submitted_at":"2025-06-04T14:08:44Z","abstract_excerpt":"Model pruning in transformer-based language models, traditionally viewed as a means of achieving computational savings, can enhance the model's reasoning capabilities. In this work, we uncover a surprising phenomenon: the selective pruning of certain attention heads leads to improvements in reasoning performance, particularly on challenging tasks. Motivated by this observation, we propose SPRINT, a novel contrastive learning framework that dynamically selects the optimal head and layer to prune during inference. By aligning question embeddings with head embeddings, SPRINT identifies those prun"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.03978","kind":"arxiv","version":2},"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/2506.03978/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":"2506.03978","created_at":"2026-07-05T11:18:22.879947+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.03978v2","created_at":"2026-07-05T11:18:22.879947+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.03978","created_at":"2026-07-05T11:18:22.879947+00:00"},{"alias_kind":"pith_short_12","alias_value":"QJS4RCYKFKQK","created_at":"2026-07-05T11:18:22.879947+00:00"},{"alias_kind":"pith_short_16","alias_value":"QJS4RCYKFKQKKNWZ","created_at":"2026-07-05T11:18:22.879947+00:00"},{"alias_kind":"pith_short_8","alias_value":"QJS4RCYK","created_at":"2026-07-05T11:18:22.879947+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.01194","citing_title":"VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A","json":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A.json","graph_json":"https://pith.science/api/pith-number/QJS4RCYKFKQKKNWZDJKU5TZ73A/graph.json","events_json":"https://pith.science/api/pith-number/QJS4RCYKFKQKKNWZDJKU5TZ73A/events.json","paper":"https://pith.science/paper/QJS4RCYK"},"agent_actions":{"view_html":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A","download_json":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A.json","view_paper":"https://pith.science/paper/QJS4RCYK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.03978&json=true","fetch_graph":"https://pith.science/api/pith-number/QJS4RCYKFKQKKNWZDJKU5TZ73A/graph.json","fetch_events":"https://pith.science/api/pith-number/QJS4RCYKFKQKKNWZDJKU5TZ73A/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A/action/storage_attestation","attest_author":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A/action/author_attestation","sign_citation":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A/action/citation_signature","submit_replication":"https://pith.science/pith/QJS4RCYKFKQKKNWZDJKU5TZ73A/action/replication_record"}},"created_at":"2026-07-05T11:18:22.879947+00:00","updated_at":"2026-07-05T11:18:22.879947+00:00"}