EDR Training Objective Improves Speculative Decoding Draft Models
Training parallel drafters has been stuck with block-local surrogates that ignore cross-round coupling; EDR closes that gap with an objective that provably matches global decoding cost. Teams running LLM inference can apply it to existing drafters like DSpark and DFly to cut round counts without adding hyperparameters. The offline evaluator also lets researchers compare drafters on shared rollouts without running full speculative decoding.
