{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:ZW67IC5TVTEVM75APPZUYYMUZL","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"e44c8d1bf337f28a9eef6d07f8920fbb29ab4a8cc9ed4ba42ac5d50a4bbe688a","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2025-05-28T17:39:15Z","title_canon_sha256":"1d06872d1f805a26fa9fccdb0e1669fc666f60fd1dd3785d349b212af66fdf32"},"schema_version":"1.0","source":{"id":"2505.22618","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2505.22618","created_at":"2026-05-17T23:38:49Z"},{"alias_kind":"arxiv_version","alias_value":"2505.22618v3","created_at":"2026-05-17T23:38:49Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.22618","created_at":"2026-05-17T23:38:49Z"},{"alias_kind":"pith_short_12","alias_value":"ZW67IC5TVTEV","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"ZW67IC5TVTEVM75A","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"ZW67IC5T","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:2210cb9b14156e39d3c440e569e921154ccb64622f06af5f332b0300933f2090","target":"graph","created_at":"2026-05-17T23:38:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to 27.6× throughput improvement with minimal accuracy loss, closing the performance gap with autoregressive models."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the block-wise approximate KV cache introduces only negligible performance drop and that a single confidence threshold can be chosen to preserve generation quality across benchmarks without post-hoc per-task retuning."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Fast-dLLM adds reusable KV cache blocks and selective parallel decoding to diffusion LLMs, closing most of the speed gap with autoregressive models without retraining."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Diffusion LLMs can reach up to 27 times higher throughput by adding a reusable block-wise KV cache and decoding only high-confidence tokens in parallel."}],"snapshot_sha256":"f24406c2f002b1a64e35976b50236fbcc4633fac5ca074731963f35337f28e30"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d77a8a9c1f6ad21895b1a43885db0c963b200bcfefb761bd05082ee3cdeef337"},"paper":{"abstract_excerpt":"Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation with parallel decoding capabilities. However, the practical inference speed of open-sourced Diffusion LLMs often lags behind autoregressive models due to the lack of Key-Value (KV) Cache and quality degradation when decoding multiple tokens simultaneously. To bridge this gap, we introduce a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop. Additionally, we identify the root cause of gen","authors_text":"Chengyue Wu, Enze Xie, Hao Zhang, Ligeng Zhu, Ping Luo, Shizhe Diao, Shuchen Xue, Song Han, Zhijian Liu","cross_cats":[],"headline":"Diffusion LLMs can reach up to 27 times higher throughput by adding a reusable block-wise KV cache and decoding only high-confidence tokens in parallel.","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2025-05-28T17:39:15Z","title":"Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding"},"references":{"count":44,"internal_anchors":3,"resolved_work":44,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Chiu, Zhihan Yang, Zhixuan Qi, Jiaqi Han, Subham Sekhar Sahoo, and V olodymyr Kuleshov","work_id":"a7d122b4-332a-4b4e-b20f-261ba8497fa0","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Structured denoising diffusion models in discrete state-spaces","work_id":"73976884-b26f-4171-a159-b0f0a8b9ce9e","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"A continuous time framework for discrete denoising models","work_id":"f67f008a-e2de-44f8-a5a0-6dd58dc2befd","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Fast sampling via de-randomization for discrete diffusion models","work_id":"19c1e430-1409-412d-b13e-f87d1e5714a7","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Discrete flow matching","work_id":"d3810acb-2472-492b-b3b2-2cefdf738e9c","year":2024}],"snapshot_sha256":"01b65f53997ee7bc50ebae18ef6441d398d39a42222c7c90bd024da7edd75c84"},"source":{"id":"2505.22618","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-16T04:22:39.565036Z","id":"e45f5b97-466a-4ff5-bba7-06dba8cc791c","model_set":{"reader":"grok-4.3"},"one_line_summary":"Fast-dLLM adds reusable KV cache blocks and selective parallel decoding to diffusion LLMs, closing most of the speed gap with autoregressive models without retraining.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Diffusion LLMs can reach up to 27 times higher throughput by adding a reusable block-wise KV cache and decoding only high-confidence tokens in parallel.","strongest_claim":"Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to 27.6× throughput improvement with minimal accuracy loss, closing the performance gap with autoregressive models.","weakest_assumption":"That the block-wise approximate KV cache introduces only negligible performance drop and that a single confidence threshold can be chosen to preserve generation quality across benchmarks without post-hoc per-task retuning."}},"verdict_id":"e45f5b97-466a-4ff5-bba7-06dba8cc791c"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:34310bb02736a204f39da2e2ee1774f4774609342cabc6038b280f888241aefb","target":"record","created_at":"2026-05-17T23:38:49Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"e44c8d1bf337f28a9eef6d07f8920fbb29ab4a8cc9ed4ba42ac5d50a4bbe688a","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2025-05-28T17:39:15Z","title_canon_sha256":"1d06872d1f805a26fa9fccdb0e1669fc666f60fd1dd3785d349b212af66fdf32"},"schema_version":"1.0","source":{"id":"2505.22618","kind":"arxiv","version":3}},"canonical_sha256":"cdbdf40bb3acc9567fa07bf34c6194cac65e7a7edf1940d6d865ac759c9a4607","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"cdbdf40bb3acc9567fa07bf34c6194cac65e7a7edf1940d6d865ac759c9a4607","first_computed_at":"2026-05-17T23:38:49.113794Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:49.113794Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"HGJkjoPa92YnH+CMVS+NVDvdiOQRbsWOMyMujYN8D2pWzh5fQZ+tTRm/X7wc5y0JSB6cGvp5fphvBHQZ1tKRDw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:49.114586Z","signed_message":"canonical_sha256_bytes"},"source_id":"2505.22618","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:34310bb02736a204f39da2e2ee1774f4774609342cabc6038b280f888241aefb","sha256:2210cb9b14156e39d3c440e569e921154ccb64622f06af5f332b0300933f2090"],"state_sha256":"4570a6a60f3381333167dd2fd780f042989e2741e71330f721a973789089913e"}