Paper 2603.03251
Speculative Speculative Decoding
- Published
- Mar 2026
- Research lab
- Together AI
- Citations
- 11
- GitHub
- Not linked
01 In brief
Summary
The paper introduces speculative speculative decoding (SSD), a framework that parallelizes drafting and verification in speculative decoding (SD) by having the draft model predict likely verification outcomes and pre-speculate for them while verification runs.
This eliminates drafting overhead on cache hits, making SSD lossless and faster than SD.
The authors identify three challenges: predicting verification outcomes, balancing cache hit and acceptance rates, and handling cache misses.
They propose Saguaro, an optimized SSD algorithm, which uses a geometric fan-out strategy for cache construction, a novel sampling scheme to improve cache hit rates, and an optimal fallback strategy based on batch size.
Theoretical results show SSD is strictly faster than SD and provide speedup bounds.
Empirically, Saguaro achieves on average 30% faster decoding than the strongest SD baselines and up to 5x faster than autoregressive decoding on Llama-3.1-70B and Qwen-3-32B across multiple datasets, improving the throughput-latency Pareto frontier.
The work also discusses combining SSD with other SD variants like EAGLE-3 and token-tree methods, and notes limitations for throughput-bound workloads, though SSD still improves the Pareto frontier.
Future directions include scaling draft devices and cluster-level speculation sharing.
02 From the paper
Abstract
Autoregressive decoding is bottlenecked by its sequential nature. Speculative decoding has become a standard way to accelerate inference by using a fast draft model to predict upcoming tokens from a slower target model, and then verifying them in parallel with a single target model forward pass. However, speculative decoding itself relies on a sequential dependence between speculation and verification. We introduce speculative speculative decoding (SSD) to parallelize these operations. While a verification is ongoing, the draft model predicts likely verification outcomes and prepares speculations pre-emptively for them. If the actual verification outcome is then in the predicted set, a speculation can be returned immediately, eliminating drafting overhead entirely. We identify three key challenges presented by speculative speculative decoding, and suggest principled methods to solve each. The result is Saguaro, an optimized SSD algorithm. Our implementation is on average 30% faster than optimized speculative decoding baselines and up to 5x faster than autoregressive decoding with open source inference engines.