{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:GBPHB5LSPBS36LHQLJGTFEMHKP","short_pith_number":"pith:GBPHB5LS","schema_version":"1.0","canonical_sha256":"305e70f5727865bf2cf05a4d32918753c0aaa87eb4305fc2fada7c66d2013ce7","source":{"kind":"arxiv","id":"2405.04304","version":5},"attestation_state":"computed","paper":{"title":"Dynamic Speculation Lookahead Accelerates Speculative Decoding of Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Daniel Korat, Jonathan Mamou, Moshe Berchansky, Moshe Wasserblat, Nadav Timor, Oren Pereg, Roy Schwartz","submitted_at":"2024-05-07T13:27:52Z","abstract_excerpt":"Speculative decoding is commonly used for reducing the inference latency of large language models. Its effectiveness depends highly on the speculation lookahead (SL)-the number of tokens generated by the draft model at each iteration. In this work we show that the common practice of using the same SL for all iterations (static SL) is suboptimal. We introduce DISCO (DynamIc SpeCulation lookahead Optimization), a novel method for dynamically selecting the SL. Our experiments with four datasets show that DISCO reaches an average speedup of 10% compared to the best static SL baseline, while genera"},"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":"2405.04304","kind":"arxiv","version":5},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-05-07T13:27:52Z","cross_cats_sorted":[],"title_canon_sha256":"497ec7f4fcad1f5ec81a3b5a96f3169870fceb23eaa7df23c02374bab08a40d4","abstract_canon_sha256":"63c6edc3a1367bd7b60d04955138cf28aa93a51e761d71eb5f30a8b78199e1bb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:32:07.660795Z","signature_b64":"DL93ANYYxlAiKAc5qimAZMjgLHAyo7SWwt3vHdw6BSQbqRuOiAN+usq09iPK2KZomLrYXXO8FtW70cJacMspAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"305e70f5727865bf2cf05a4d32918753c0aaa87eb4305fc2fada7c66d2013ce7","last_reissued_at":"2026-07-05T09:32:07.660205Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:32:07.660205Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dynamic Speculation Lookahead Accelerates Speculative Decoding of Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Daniel Korat, Jonathan Mamou, Moshe Berchansky, Moshe Wasserblat, Nadav Timor, Oren Pereg, Roy Schwartz","submitted_at":"2024-05-07T13:27:52Z","abstract_excerpt":"Speculative decoding is commonly used for reducing the inference latency of large language models. Its effectiveness depends highly on the speculation lookahead (SL)-the number of tokens generated by the draft model at each iteration. In this work we show that the common practice of using the same SL for all iterations (static SL) is suboptimal. We introduce DISCO (DynamIc SpeCulation lookahead Optimization), a novel method for dynamically selecting the SL. Our experiments with four datasets show that DISCO reaches an average speedup of 10% compared to the best static SL baseline, while genera"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.04304","kind":"arxiv","version":5},"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/2405.04304/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":"2405.04304","created_at":"2026-07-05T09:32:07.660270+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.04304v5","created_at":"2026-07-05T09:32:07.660270+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.04304","created_at":"2026-07-05T09:32:07.660270+00:00"},{"alias_kind":"pith_short_12","alias_value":"GBPHB5LSPBS3","created_at":"2026-07-05T09:32:07.660270+00:00"},{"alias_kind":"pith_short_16","alias_value":"GBPHB5LSPBS36LHQ","created_at":"2026-07-05T09:32:07.660270+00:00"},{"alias_kind":"pith_short_8","alias_value":"GBPHB5LS","created_at":"2026-07-05T09:32:07.660270+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.15051","citing_title":"An Interpretable Latency Model for Speculative Decoding in LLM Serving","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00487","citing_title":"TAPS: Target-Aware Prefix Tree Selection for Diffusion-Drafted Speculative Decoding","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20104","citing_title":"Draft Less, Retrieve More: Hybrid Tree Construction for Speculative Decoding","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14978","citing_title":"Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09731","citing_title":"SMART: When is it Actually Worth Expanding a Speculative Tree?","ref_index":25,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP","json":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP.json","graph_json":"https://pith.science/api/pith-number/GBPHB5LSPBS36LHQLJGTFEMHKP/graph.json","events_json":"https://pith.science/api/pith-number/GBPHB5LSPBS36LHQLJGTFEMHKP/events.json","paper":"https://pith.science/paper/GBPHB5LS"},"agent_actions":{"view_html":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP","download_json":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP.json","view_paper":"https://pith.science/paper/GBPHB5LS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.04304&json=true","fetch_graph":"https://pith.science/api/pith-number/GBPHB5LSPBS36LHQLJGTFEMHKP/graph.json","fetch_events":"https://pith.science/api/pith-number/GBPHB5LSPBS36LHQLJGTFEMHKP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP/action/storage_attestation","attest_author":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP/action/author_attestation","sign_citation":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP/action/citation_signature","submit_replication":"https://pith.science/pith/GBPHB5LSPBS36LHQLJGTFEMHKP/action/replication_record"}},"created_at":"2026-07-05T09:32:07.660270+00:00","updated_at":"2026-07-05T09:32:07.660270+00:00"}