{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:LLYV2JHHEXPLM52DAFNQWQTAQY","short_pith_number":"pith:LLYV2JHH","schema_version":"1.0","canonical_sha256":"5af15d24e725deb67743015b0b4260863e7596e9ba07145098d3afb11b6499aa","source":{"kind":"arxiv","id":"2605.29343","version":1},"attestation_state":"computed","paper":{"title":"Draft-OPD: On-Policy Distillation for Speculative Draft Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bowen Zhou, Ganqu Cui, Haodi Lei, Haoran Zhang, Ning Ding, Qianjia Cheng, Shunkai Zhang, Xiaoye Qu, Yafy Li, Yu Cheng, Yun Luo","submitted_at":"2026-05-28T04:30:22Z","abstract_excerpt":"Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under "},"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":"2605.29343","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-28T04:30:22Z","cross_cats_sorted":[],"title_canon_sha256":"1778396558b535df7faaf202f0cc56dbf60917963d27f0701e0439d8fd4722dc","abstract_canon_sha256":"37f0c856f8985d47134305222ba33b497c008ba5829bd0b562d64beee28e1cc6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T01:05:33.919210Z","signature_b64":"23V5WVYcC2g778P0y+MlrQYlvFpmD3D/HxeLa/3So5Arq4QlzYnJaQLSwhhW1srrEsm/44OSp10zmVAJFiMsAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5af15d24e725deb67743015b0b4260863e7596e9ba07145098d3afb11b6499aa","last_reissued_at":"2026-05-29T01:05:33.918476Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T01:05:33.918476Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Draft-OPD: On-Policy Distillation for Speculative Draft Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bowen Zhou, Ganqu Cui, Haodi Lei, Haoran Zhang, Ning Ding, Qianjia Cheng, Shunkai Zhang, Xiaoye Qu, Yafy Li, Yu Cheng, Yun Luo","submitted_at":"2026-05-28T04:30:22Z","abstract_excerpt":"Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.29343","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/2605.29343/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":"2605.29343","created_at":"2026-05-29T01:05:33.918597+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.29343v1","created_at":"2026-05-29T01:05:33.918597+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.29343","created_at":"2026-05-29T01:05:33.918597+00:00"},{"alias_kind":"pith_short_12","alias_value":"LLYV2JHHEXPL","created_at":"2026-05-29T01:05:33.918597+00:00"},{"alias_kind":"pith_short_16","alias_value":"LLYV2JHHEXPLM52D","created_at":"2026-05-29T01:05:33.918597+00:00"},{"alias_kind":"pith_short_8","alias_value":"LLYV2JHH","created_at":"2026-05-29T01:05:33.918597+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/LLYV2JHHEXPLM52DAFNQWQTAQY","json":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY.json","graph_json":"https://pith.science/api/pith-number/LLYV2JHHEXPLM52DAFNQWQTAQY/graph.json","events_json":"https://pith.science/api/pith-number/LLYV2JHHEXPLM52DAFNQWQTAQY/events.json","paper":"https://pith.science/paper/LLYV2JHH"},"agent_actions":{"view_html":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY","download_json":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY.json","view_paper":"https://pith.science/paper/LLYV2JHH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.29343&json=true","fetch_graph":"https://pith.science/api/pith-number/LLYV2JHHEXPLM52DAFNQWQTAQY/graph.json","fetch_events":"https://pith.science/api/pith-number/LLYV2JHHEXPLM52DAFNQWQTAQY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY/action/storage_attestation","attest_author":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY/action/author_attestation","sign_citation":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY/action/citation_signature","submit_replication":"https://pith.science/pith/LLYV2JHHEXPLM52DAFNQWQTAQY/action/replication_record"}},"created_at":"2026-05-29T01:05:33.918597+00:00","updated_at":"2026-05-29T01:05:33.918597+00:00"}