Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Dustin: Draft-Augmented Sparse Verification for Efficient Long-Context Generation with Speculative Decoding

About

While speculative decoding improves inference throughput for multi-batch long-context Large Language Models (LLMs), its efficiency is often limited by a verification bottleneck where Key-Value (KV) cache loading dominates latency. Existing compression methods fail in this regime: static eviction incurs accuracy loss due to saliency shift, while dynamic selection introduces prohibitive computational overhead during the verification path. We propose Dustin, a sparse verification framework designed for long-context speculative decoding. Dustin integrates lookahead signals from the draft model with historical attention from the target model to identify critical tokens with high fidelity across multi-step verification windows. To reduce recomputation latency, this approach further employs a sparse estimation scheme that restricts importance scoring to a minimal subset of attention heads. Evaluations on PG-19 and LongBench with Qwen2.5-72B demonstrate that Dustin achieves a 27.85x speedup in self-attention and a 9.17x end-to-end decoding speedup at a 32k sequence length, all with negligible accuracy degradation.

WenHung Lee, Jian-Jia Chen, Xiaolin Lin, Pei-Shuo Wang, Chi-Chih Chang, Chun-Che Yang, Ning-Chi Huang, Grace Li Zhang, Kai-Chiang Wu• 2026

Related benchmarks

TaskDatasetResultRank
Long-context language modelingLongBench--
369
Long-context UnderstandingLongBench (test)
Avg Score51.46
166
Long-context UnderstandingLongBench
Overall Average Score55.23
143
Long-Context Generation EfficiencyPG-19
Throughput408
64
Long-context generationPG-19 8K context length
Throughput553.8
32
Long-context generationPG-19 16K context length
Throughput486.5
32
Long-context generationPG-19 24K context length
Throughput393.9
32
Long-context generationPG-19 32K context length
Throughput329.6
32
Long-Context Generation EfficiencyQwen2.5-72B-Instruct Long-Context Efficiency Workload
Speedup (8K)4.76
2
Showing 9 of 9 rows

Other info

Follow for update