Reproduce “Draft-Model Speculative Decoding Can Emit Zero-Probability Tokens” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanyu Yang
The authors audit whether speculative decoding preserves a target model’s output distribution in inference engines. They compare speculative paths with each engine’s non-speculative sampler on small fixtures. In TensorRT-LLM 1.2.1’s tested draft-target path, top-k 3 produced TV distance 0.574, and 45.0% of sequences had zero reference probability; these results do not establish failure across models or paths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.