{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:7TWKGTTAPJANMAXSMPYI3Y74TU","short_pith_number":"pith:7TWKGTTA","schema_version":"1.0","canonical_sha256":"fceca34e607a40d602f263f08de3fc9d26d96ba0f59c6a071099f50eb2fba197","source":{"kind":"arxiv","id":"2504.18246","version":2},"attestation_state":"computed","paper":{"title":"One-Pass to Reason: Token Duplication and Block-Sparse Mask for Efficient Fine-Tuning on Multi-Turn Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Prateek Jain, Ritesh Goru, Shanay Mehta","submitted_at":"2025-04-25T10:46:56Z","abstract_excerpt":"Fine-tuning Large Language Models (LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from $O\\bigl(N^{3}\\bigl)$ to $O\\bigl(N^{2}\\bigl)$ and maintaining the same memory complexity for a transformer "},"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":"2504.18246","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-04-25T10:46:56Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"5a024dd24535258a796570a7620784e82eb65e5770b1276971490c2f1adc8a38","abstract_canon_sha256":"3fe5fb83686251448b3ecbe8157d8aca9714e636f8e3022f07ede6853d9e31f6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:35:34.781365Z","signature_b64":"6ARUABU8ucmKgn3wxtrzowkIMaXrGgdVpVQ7y+zIlewcO184JfpjQSBK1AGfnHM0PwqVJsRfmjxq8qEhBII4Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fceca34e607a40d602f263f08de3fc9d26d96ba0f59c6a071099f50eb2fba197","last_reissued_at":"2026-07-05T11:35:34.780876Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:35:34.780876Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"One-Pass to Reason: Token Duplication and Block-Sparse Mask for Efficient Fine-Tuning on Multi-Turn Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Prateek Jain, Ritesh Goru, Shanay Mehta","submitted_at":"2025-04-25T10:46:56Z","abstract_excerpt":"Fine-tuning Large Language Models (LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from $O\\bigl(N^{3}\\bigl)$ to $O\\bigl(N^{2}\\bigl)$ and maintaining the same memory complexity for a transformer "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.18246","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/2504.18246/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":"2504.18246","created_at":"2026-07-05T11:35:34.780933+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.18246v2","created_at":"2026-07-05T11:35:34.780933+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.18246","created_at":"2026-07-05T11:35:34.780933+00:00"},{"alias_kind":"pith_short_12","alias_value":"7TWKGTTAPJAN","created_at":"2026-07-05T11:35:34.780933+00:00"},{"alias_kind":"pith_short_16","alias_value":"7TWKGTTAPJANMAXS","created_at":"2026-07-05T11:35:34.780933+00:00"},{"alias_kind":"pith_short_8","alias_value":"7TWKGTTA","created_at":"2026-07-05T11:35:34.780933+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2511.00413","citing_title":"Tree Training: Accelerating Agentic LLMs Training via Shared Prefix Reuse","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU","json":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU.json","graph_json":"https://pith.science/api/pith-number/7TWKGTTAPJANMAXSMPYI3Y74TU/graph.json","events_json":"https://pith.science/api/pith-number/7TWKGTTAPJANMAXSMPYI3Y74TU/events.json","paper":"https://pith.science/paper/7TWKGTTA"},"agent_actions":{"view_html":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU","download_json":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU.json","view_paper":"https://pith.science/paper/7TWKGTTA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.18246&json=true","fetch_graph":"https://pith.science/api/pith-number/7TWKGTTAPJANMAXSMPYI3Y74TU/graph.json","fetch_events":"https://pith.science/api/pith-number/7TWKGTTAPJANMAXSMPYI3Y74TU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU/action/storage_attestation","attest_author":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU/action/author_attestation","sign_citation":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU/action/citation_signature","submit_replication":"https://pith.science/pith/7TWKGTTAPJANMAXSMPYI3Y74TU/action/replication_record"}},"created_at":"2026-07-05T11:35:34.780933+00:00","updated_at":"2026-07-05T11:35:34.780933+00:00"}