{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:A7V2254YD5WHMLNE5DFW5Z5X37","short_pith_number":"pith:A7V2254Y","schema_version":"1.0","canonical_sha256":"07ebad77981f6c762da4e8cb6ee7b7dfeb5aa04bbd6b90890cba5937f281cb64","source":{"kind":"arxiv","id":"2605.16786","version":1},"attestation_state":"computed","paper":{"title":"Lever: Speculative LLM Inference on Smartphones","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Lever reduces smartphone LLM inference latency by 2.93x over flash baselines through optimized speculative decoding.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Fengzu Li, Ju Ren, Tuowei Wang, Wei Gao, Yanfan Sun","submitted_at":"2026-05-16T03:43:10Z","abstract_excerpt":"Large language models (LLMs) are increasingly needed for interactive mobile applications, but high-quality models exceed the limited DRAM available on smartphones. Flash storage can hold larger models, yet flash-backed inference is slow because autoregressive decoding repeatedly invokes the target model and incurs costly I/O. We observe that speculative decoding is a natural fit for this setting: a small draft model can remain in DRAM, while a larger flash-resident target model verifies multiple candidate tokens per invocation. However, existing methods assume server-class accelerators and fai"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.16786","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-16T03:43:10Z","cross_cats_sorted":[],"title_canon_sha256":"e169d93e2760b75e10d1dd3c072fb954d0b34eeb2ef0500afd4663bbc98af489","abstract_canon_sha256":"3f929fedd8949b62128e5a9887c5e2203e6e36e004f53e0ef44725a255407b01"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:03:21.937389Z","signature_b64":"7/iSMpI0rDKIHH/W0V7k40cA6z/lsd+QW1upW0oP4cdtbtQKq5ikP0vdzNUTbhNwke/HHPc4IT8t51LPN5ZSBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"07ebad77981f6c762da4e8cb6ee7b7dfeb5aa04bbd6b90890cba5937f281cb64","last_reissued_at":"2026-05-20T00:03:21.936503Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:03:21.936503Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Lever: Speculative LLM Inference on Smartphones","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Lever reduces smartphone LLM inference latency by 2.93x over flash baselines through optimized speculative decoding.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Fengzu Li, Ju Ren, Tuowei Wang, Wei Gao, Yanfan Sun","submitted_at":"2026-05-16T03:43:10Z","abstract_excerpt":"Large language models (LLMs) are increasingly needed for interactive mobile applications, but high-quality models exceed the limited DRAM available on smartphones. Flash storage can hold larger models, yet flash-backed inference is slow because autoregressive decoding repeatedly invokes the target model and incurs costly I/O. We observe that speculative decoding is a natural fit for this setting: a small draft model can remain in DRAM, while a larger flash-resident target model verifies multiple candidate tokens per invocation. However, existing methods assume server-class accelerators and fai"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Lever reduces inference latency by an average of 2.93x over baseline flash-offloaded inference and 1.50x over conventional speculative decoding, narrowing the latency gap between flash-backed and memory-resident LLM inference.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that jointly optimizing token-tree construction with an I/O- and compute-aware gain-cost objective, early-exit pruning, and CPU-NPU mapping will deliver the claimed speedups under real smartphone I/O latency and parallelism constraints (stated in the abstract description of the three stages).","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Lever optimizes the drafting, verification, and execution stages of speculative decoding for flash-backed LLM inference on smartphones, reporting 2.93x average latency reduction over baseline flash-offloaded inference.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Lever reduces smartphone LLM inference latency by 2.93x over flash baselines through optimized speculative decoding.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"22321fe8fb1e2e705f5dfa4b3ccc5f2c8453f4102283486d727b4110e21921db"},"source":{"id":"2605.16786","kind":"arxiv","version":1},"verdict":{"id":"aa5ea424-dbb2-41ce-b6a5-16fd26eabf79","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T21:39:51.959093Z","strongest_claim":"Lever reduces inference latency by an average of 2.93x over baseline flash-offloaded inference and 1.50x over conventional speculative decoding, narrowing the latency gap between flash-backed and memory-resident LLM inference.","one_line_summary":"Lever optimizes the drafting, verification, and execution stages of speculative decoding for flash-backed LLM inference on smartphones, reporting 2.93x average latency reduction over baseline flash-offloaded inference.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that jointly optimizing token-tree construction with an I/O- and compute-aware gain-cost objective, early-exit pruning, and CPU-NPU mapping will deliver the claimed speedups under real smartphone I/O latency and parallelism constraints (stated in the abstract description of the three stages).","pith_extraction_headline":"Lever reduces smartphone LLM inference latency by 2.93x over flash baselines through optimized speculative decoding."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16786/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T22:01:19.651802Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:50:55.144814Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T19:01:56.298520Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.434219Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"033653024c54694c2ab0b8aaa7f00a7e56e2289d5ac785aec6ad7a1a8e2a07fb"},"references":{"count":40,"sample":[{"doi":"","year":2024,"title":"Llm in a flash: Efficient large language model inference with limited memory","work_id":"a9c527dd-dab9-4723-894f-c136b3e1ccaf","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Hydra: Sequentially-dependent draft heads for medusa decoding","work_id":"c49cff9c-dcd0-4434-ad6b-51534aca2b2f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Program Synthesis with Large Language Models","work_id":"fd241a05-03b9-4de2-9588-9d77ce176125","ref_index":3,"cited_arxiv_id":"2108.07732","is_internal_anchor":true},{"doi":"","year":2024,"title":"Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads","work_id":"f6202cd1-1a78-4c19-8242-688acf4952b6","ref_index":4,"cited_arxiv_id":"2401.10774","is_internal_anchor":true},{"doi":"","year":2023,"title":"Accelerating Large Language Model Decoding with Speculative Sampling","work_id":"b53b11e5-8613-439d-9d9d-0e2a9090d79f","ref_index":5,"cited_arxiv_id":"2302.01318","is_internal_anchor":true}],"resolved_work":40,"snapshot_sha256":"cf7636b1bc1a7d1a0adbce931c825d59311bc69e86f4a77b84ddec39c92c4cee","internal_anchors":8},"formal_canon":{"evidence_count":2,"snapshot_sha256":"fd5d216180b46a71e3f4c9ed90582bafdf96906dc54352ca62d6495bfa29330d"},"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":"2605.16786","created_at":"2026-05-20T00:03:21.936661+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.16786v1","created_at":"2026-05-20T00:03:21.936661+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.16786","created_at":"2026-05-20T00:03:21.936661+00:00"},{"alias_kind":"pith_short_12","alias_value":"A7V2254YD5WH","created_at":"2026-05-20T00:03:21.936661+00:00"},{"alias_kind":"pith_short_16","alias_value":"A7V2254YD5WHMLNE","created_at":"2026-05-20T00:03:21.936661+00:00"},{"alias_kind":"pith_short_8","alias_value":"A7V2254Y","created_at":"2026-05-20T00:03:21.936661+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37","json":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37.json","graph_json":"https://pith.science/api/pith-number/A7V2254YD5WHMLNE5DFW5Z5X37/graph.json","events_json":"https://pith.science/api/pith-number/A7V2254YD5WHMLNE5DFW5Z5X37/events.json","paper":"https://pith.science/paper/A7V2254Y"},"agent_actions":{"view_html":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37","download_json":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37.json","view_paper":"https://pith.science/paper/A7V2254Y","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.16786&json=true","fetch_graph":"https://pith.science/api/pith-number/A7V2254YD5WHMLNE5DFW5Z5X37/graph.json","fetch_events":"https://pith.science/api/pith-number/A7V2254YD5WHMLNE5DFW5Z5X37/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37/action/timestamp_anchor","attest_storage":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37/action/storage_attestation","attest_author":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37/action/author_attestation","sign_citation":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37/action/citation_signature","submit_replication":"https://pith.science/pith/A7V2254YD5WHMLNE5DFW5Z5X37/action/replication_record"}},"created_at":"2026-05-20T00:03:21.936661+00:00","updated_at":"2026-05-20T00:03:21.936661+00:00"}