{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:BQEWM3B6VASAPEWZD74GYS65E3","short_pith_number":"pith:BQEWM3B6","schema_version":"1.0","canonical_sha256":"0c09666c3ea8240792d91ff86c4bdd26eed25a27ee7642bd1f386be3941da7be","source":{"kind":"arxiv","id":"2506.19830","version":1},"attestation_state":"computed","paper":{"title":"Scaling Speculative Decoding with Lookahead Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hao Zhang, Rui Ge, Yichao Fu, Zelei Shao, Zhijie Deng","submitted_at":"2025-06-24T17:48:10Z","abstract_excerpt":"Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level speculative decoding (SD) helps, but its benefit is capped, because the chance that an entire $\\gamma$-token guess is correct falls exponentially as $\\gamma$ grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with Lookahead Reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-st"},"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":"2506.19830","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-06-24T17:48:10Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"ea98f841f7808bdfe3a307ffbd008c97ba2db41011689dd92f4a966caf4ec704","abstract_canon_sha256":"7ef9650bb3ba30e78644e6cd696c3068626b43252ede7927dfb3f72a71d07c17"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:26:39.105627Z","signature_b64":"7yWnJ3HKXe3inwfBYXsXdD4zW0XXVaPVUwT6pdspbWzISsHqdd6Wf7LmTpjBjAa4KfWWkTTYaT+vr2FS6EBIAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0c09666c3ea8240792d91ff86c4bdd26eed25a27ee7642bd1f386be3941da7be","last_reissued_at":"2026-07-05T11:26:39.105121Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:26:39.105121Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Scaling Speculative Decoding with Lookahead Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hao Zhang, Rui Ge, Yichao Fu, Zelei Shao, Zhijie Deng","submitted_at":"2025-06-24T17:48:10Z","abstract_excerpt":"Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level speculative decoding (SD) helps, but its benefit is capped, because the chance that an entire $\\gamma$-token guess is correct falls exponentially as $\\gamma$ grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with Lookahead Reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-st"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.19830","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/2506.19830/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":"2506.19830","created_at":"2026-07-05T11:26:39.105180+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.19830v1","created_at":"2026-07-05T11:26:39.105180+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.19830","created_at":"2026-07-05T11:26:39.105180+00:00"},{"alias_kind":"pith_short_12","alias_value":"BQEWM3B6VASA","created_at":"2026-07-05T11:26:39.105180+00:00"},{"alias_kind":"pith_short_16","alias_value":"BQEWM3B6VASAPEWZ","created_at":"2026-07-05T11:26:39.105180+00:00"},{"alias_kind":"pith_short_8","alias_value":"BQEWM3B6","created_at":"2026-07-05T11:26:39.105180+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25354","citing_title":"Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.25354","citing_title":"Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing","ref_index":7,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3","json":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3.json","graph_json":"https://pith.science/api/pith-number/BQEWM3B6VASAPEWZD74GYS65E3/graph.json","events_json":"https://pith.science/api/pith-number/BQEWM3B6VASAPEWZD74GYS65E3/events.json","paper":"https://pith.science/paper/BQEWM3B6"},"agent_actions":{"view_html":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3","download_json":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3.json","view_paper":"https://pith.science/paper/BQEWM3B6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.19830&json=true","fetch_graph":"https://pith.science/api/pith-number/BQEWM3B6VASAPEWZD74GYS65E3/graph.json","fetch_events":"https://pith.science/api/pith-number/BQEWM3B6VASAPEWZD74GYS65E3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3/action/storage_attestation","attest_author":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3/action/author_attestation","sign_citation":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3/action/citation_signature","submit_replication":"https://pith.science/pith/BQEWM3B6VASAPEWZD74GYS65E3/action/replication_record"}},"created_at":"2026-07-05T11:26:39.105180+00:00","updated_at":"2026-07-05T11:26:39.105180+00:00"}