This is the official repository for the paper: EDSD: Entropy-Driven Design for Faster Speculative Decoding (ACL 2026). EDSD is an entropy-driven framework for compute- and data-efficient drafter training and design in speculative decoding, reducing training cost while improving acceptance and robustness.
Our implementation is now publicly available at:
https://github.com/KerwinKai/SpecForge/tree/add_edsd
The current implementation is developed on top of SpecForge. We are in the process of preparing our contribution to the SpecForge project, with the goal of making EDSD more accessible to the broader community.
