On Monday, 27 July 2026, at 11:30 AM, the Center for Computational & Data Sciences (CCDS), Independent University, Bangladesh (IUB), hosted a weekly paper presentation on speculative decoding and accelerating LLM inference with draft models.
Presenter: Fahim Ahmed, Research Assistant, CCDS, IUB
Papers Presented:
The session covered speculative decoding, a family of methods that speed up inference in a large target model by using a smaller draft model to propose upcoming tokens. The first paper introduced the core formulation: the draft model generates several tokens ahead, and the target model verifies them in a single parallel pass, accepting or rejecting each one under a sampling scheme that leaves the output distribution unchanged. This yields lower latency without any loss in generation quality and without retraining or modifying the target model. The second paper turned to a structural limitation of that design, namely the sequential dependence between drafting and verification, where the draft model must wait for verification before continuing. By relaxing this dependency, the work aims for further speedups beyond what the original method achieves.
We thank all participants for their valuable contributions and insightful discussions.
Tags: CCDS Weekly Paper Presentation, Artificial Intelligence, Machine Learning, LLM, Speculative Decoding, Inference Efficiency, Research, CCDS, IUB