{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:6XKPDF4KK2G6KRNHVO332L6I2T","short_pith_number":"pith:6XKPDF4K","schema_version":"1.0","canonical_sha256":"f5d4f1978a568de545a7abb7bd2fc8d4c7db6a8f94e6137f4b24625049a5959b","source":{"kind":"arxiv","id":"2605.14005","version":1},"attestation_state":"computed","paper":{"title":"Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Mistletoe attacks collapse speculative decoding speedup by slashing average accepted draft length while leaving output quality unchanged.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Bin Chen, Chang Da, Fan Mo, Hao Fang, Kuofeng Gao, Shuoyang Sun, Shu-Tao Xia, Xinhao Zhong, Yi Sun","submitted_at":"2026-05-13T18:11:42Z","abstract_excerpt":"Speculative decoding has become a widely adopted technique for accelerating large language model (LLM) inference by drafting multiple candidate tokens and verifying them with a target model in parallel. Its efficiency, however, critically depends on the average accepted length $\\tau$, i.e., how many draft tokens survive each verification step. In this work, we identify a new mechanism-level vulnerability in model-based speculative decoding: the drafter is trained to approximate the target model distribution, but this approximation is inevitably imperfect. Such a drafter-target mismatch creates"},"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":false},"canonical_record":{"source":{"id":"2605.14005","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T18:11:42Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"62e3f32e581681bde61c43d619770be3d8632e779f42938b1439fcf33bd969bd","abstract_canon_sha256":"e3cc60dc8e75596b52e76d36c94fea39cac98fd01f9c4769032acbd6ce9efc56"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:13.119827Z","signature_b64":"q4bZ3IzIY8zrFukkWtJ3LK39Rp8Ju2TaZoJiM0GhDzEBD5/ts2uNX16/IXGL9ojB0lsadrQjFArR8P34UJyvBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f5d4f1978a568de545a7abb7bd2fc8d4c7db6a8f94e6137f4b24625049a5959b","last_reissued_at":"2026-05-17T23:39:13.119332Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:13.119332Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Mistletoe attacks collapse speculative decoding speedup by slashing average accepted draft length while leaving output quality unchanged.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Bin Chen, Chang Da, Fan Mo, Hao Fang, Kuofeng Gao, Shuoyang Sun, Shu-Tao Xia, Xinhao Zhong, Yi Sun","submitted_at":"2026-05-13T18:11:42Z","abstract_excerpt":"Speculative decoding has become a widely adopted technique for accelerating large language model (LLM) inference by drafting multiple candidate tokens and verifying them with a target model in parallel. Its efficiency, however, critically depends on the average accepted length $\\tau$, i.e., how many draft tokens survive each verification step. In this work, we identify a new mechanism-level vulnerability in model-based speculative decoding: the drafter is trained to approximate the target model distribution, but this approximation is inevitably imperfect. Such a drafter-target mismatch creates"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Mistletoe substantially reduces average accepted length τ, collapses speedup, and lowers averaged token throughput, while preserving output quality and perplexity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that small perturbations can preserve the target model's visible behavior while substantially reducing draft-token acceptability via null-space projection of degradation gradients.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Mistletoe is a stealthy attack that collapses the speedup of speculative decoding by reducing average accepted length τ without changing output semantics or perplexity.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Mistletoe attacks collapse speculative decoding speedup by slashing average accepted draft length while leaving output quality unchanged.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2f1cbfda2e142373fc47385d6b8996bd062310aaf74c9ccf979a927cec33e339"},"source":{"id":"2605.14005","kind":"arxiv","version":1},"verdict":{"id":"d61135cf-1ae7-49a8-92d2-51daa0fbcdf0","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:36:12.382649Z","strongest_claim":"Mistletoe substantially reduces average accepted length τ, collapses speedup, and lowers averaged token throughput, while preserving output quality and perplexity.","one_line_summary":"Mistletoe is a stealthy attack that collapses the speedup of speculative decoding by reducing average accepted length τ without changing output semantics or perplexity.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that small perturbations can preserve the target model's visible behavior while substantially reducing draft-token acceptability via null-space projection of degradation gradients.","pith_extraction_headline":"Mistletoe attacks collapse speculative decoding speedup by slashing average accepted draft length while leaving output quality unchanged."},"references":{"count":15,"sample":[{"doi":"","year":null,"title":"Hydra: Sequentially-dependent draft heads for medusa decoding","work_id":"e51f556b-4111-49b3-96b0-6191c097ea33","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads","work_id":"f6202cd1-1a78-4c19-8242-688acf4952b6","ref_index":2,"cited_arxiv_id":"2401.10774","is_internal_anchor":true},{"doi":"","year":null,"title":"Accelerating Large Language Model Decoding with Speculative Sampling","work_id":"b53b11e5-8613-439d-9d9d-0e2a9090d79f","ref_index":3,"cited_arxiv_id":"2302.01318","is_internal_anchor":true},{"doi":"","year":null,"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","ref_index":4,"cited_arxiv_id":"2107.03374","is_internal_anchor":true},{"doi":"","year":2023,"title":"Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.See https://vicuna","work_id":"9ae0c422-f9e3-4f20-bcc9-3ec30d9df7dd","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":15,"snapshot_sha256":"09a2ee1f7aa0e9a362be8c39addfcbea2a1046604e15f93a46853eaf012e9ccc","internal_anchors":8},"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":"2605.14005","created_at":"2026-05-17T23:39:13.119423+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14005v1","created_at":"2026-05-17T23:39:13.119423+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14005","created_at":"2026-05-17T23:39:13.119423+00:00"},{"alias_kind":"pith_short_12","alias_value":"6XKPDF4KK2G6","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"6XKPDF4KK2G6KRNH","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"6XKPDF4K","created_at":"2026-05-18T12:33:37.589309+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/6XKPDF4KK2G6KRNHVO332L6I2T","json":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T.json","graph_json":"https://pith.science/api/pith-number/6XKPDF4KK2G6KRNHVO332L6I2T/graph.json","events_json":"https://pith.science/api/pith-number/6XKPDF4KK2G6KRNHVO332L6I2T/events.json","paper":"https://pith.science/paper/6XKPDF4K"},"agent_actions":{"view_html":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T","download_json":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T.json","view_paper":"https://pith.science/paper/6XKPDF4K","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14005&json=true","fetch_graph":"https://pith.science/api/pith-number/6XKPDF4KK2G6KRNHVO332L6I2T/graph.json","fetch_events":"https://pith.science/api/pith-number/6XKPDF4KK2G6KRNHVO332L6I2T/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T/action/storage_attestation","attest_author":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T/action/author_attestation","sign_citation":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T/action/citation_signature","submit_replication":"https://pith.science/pith/6XKPDF4KK2G6KRNHVO332L6I2T/action/replication_record"}},"created_at":"2026-05-17T23:39:13.119423+00:00","updated_at":"2026-05-17T23:39:13.119423+00:00"}