Research map / Architectures and efficiency / Attention and transformers
Dustin: Draft-Augmented Sparse Verification for Efficient Long-Context Generation with Speculative Decoding
WenHung Lee, Jian-Jia Chen, Xiaolin Lin, Pei-Shuo Wang, Chi-Chih Chang, Chun-Che Yang, Wei-Chen Wang, Hanrui Wang, Ning-Chi Huang, Grace Li Zhang, Kai-Chiang Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add draft-based lookahead attention to target-historical importance estimation (ablation) | Dustin-T · TriviaQA | accuracy (%): 77.58% -> 83.63% | Improved |
| add target-historical attention to draft-lookahead importance estimation (ablation) | Dustin-D · TriviaQA | accuracy (%): 80.51% -> 83.63% | Improved |
| add draft-based lookahead attention to target-historical importance estimation (ablation) | Dustin-T · MultiNews | score (%): 23.95% -> 25.19% | Improved |
| add target-historical attention to draft-lookahead importance estimation (ablation) | Dustin-D · MultiNews | score (%): 24.84% -> 25.19% | Improved |
| use Dustin hybrid sparse verification instead of Quest at tight 128-token KV budget | Qwen2.5-72B · LongBench avg (KV budget 128) | score: 37.01 -> 52.41 | Improved |
| use Dustin instead of SnapKV at 128-token KV budget | Llama-3.3-70B · LongBench avg (KV budget 128) | score: 50.32 -> 53.81 | Improved |
| increase KV budget from 128 to 512 (ablation) | Dustin on Qwen2.5-72B · LongBench avg | score: 52.41 -> 55.23 | Improved |
| increase context length from 8K to 32K (batch 8) | Dustin on Qwen2.5-72B · PG-19 decode speedup | speedup vs Vanilla: 3.01x -> 6.61x | Improved |
| increase context length from 8K to 32K (batch 16) | Dustin on Qwen2.5-72B · PG-19 decode speedup | speedup vs Vanilla: 4.76x -> 9.17x | Improved |
| use Dustin instead of MagicDec at 32K context, batch 8 | Llama-3.3-70B · PG-19 decode speedup | speedup vs Vanilla: 2.05x -> 6.28x | Improved |
| restrict importance estimation to selected retrieval heads (ablation) | Dustin · importance-estimation overhead | fraction of full hybrid computation: about 0.8% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match (ICLR 2026 · similarity 0.89)
- SEED: Self-Speculative Decoding via Implicit Encoder–Decoder (NeurIPS 2026 · similarity 0.89)
- MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decoding (ICLR 2025 · similarity 0.89)
- MoESD: Unveil Speculative Decoding's Potential for Accelerating Sparse MoE (NeurIPS 2025 · similarity 0.89)
- Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment (ICLR 2025 · similarity 0.89)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
