ernie-research/dcore is a research repository comparing the conditions under which a draft model's token acceptance rate in speculative decoding translates into actual throughput. Speculative decoding is an approach in which a small draft model proposes candidate next tokens first, and the main model reflects only the tokens it accepts in the actual output. The README explains that a higher acceptance rate does not necessarily translate into better throughput. The repository was created on October 20, 2026.
The focus of this repository is evaluation design. The README compares six draft methods and three families under shared training, inference, and evaluation conditions, and also presents how the reported results differed from the re-measurement and retraining results. DCore is introduced as a baseline that combines DFlash's underlying architecture with DFlare's target-layer context mixing and the D-PACE training objective. The key points are that the ranking of draft methods changes depending on the comparison conditions, and the proposal of a cost-aware draft generation method. The license is MIT and the number of stars is 19.
That said, the code and weights are scheduled to be released after the formal paper presentation, so the results cannot be directly reproduced at present, and reproducibility and implementation depth are rated low. Novelty is 8.0, while technical quality is 4.0, quality is 5.0, and momentum is 5.7. At this point, it appears to be a meaningful repository for people who want to reference evaluation design for speculative decoding.