EDR: training parallel speculative drafters by directly minimizing expected decoding rounds lifts mean accepted length for DSpark and DFly across nine benchmarks
This work formulates speculative decoding with parallel and semi-autoregressive drafters as a Markov reward process and derives Expected Decoding Rounds (EDR), an objective that weights local rejection costs by state occupancies and exactly equals the expected number of decoding rounds, together with an exact temporal-difference gradient and an offline evaluator; finetuning DSpark and DFly with EDR consistently improves mean accepted length and outperforms existing objectives across nine math, code, and chat benchmarks.