Search papers, labs, and topics across Lattice.
This paper introduces Spec-AUF, a novel training approach for masked block drafters that addresses train-inference misalignment by concentrating supervision on the accepted prefix during autoregressive generation. By employing a teacher-forced learning strategy, Spec-AUF modifies the cross-entropy support to focus solely on the drafter's first predicted failure, enhancing the efficiency of the generation process without altering the inference pipeline. The results show a significant improvement in average emitted length across multiple benchmarks, indicating that this method effectively balances speed and accuracy in token generation.
Spec-AUF boosts the average emitted length of masked block drafters by 8.75%, demonstrating that targeted supervision can outperform traditional methods without complicating the inference process.
Speculative decoding accelerates autoregressive generation by drafting a block of tokens that the target model verifies left-to-right, committing only the longest accepted prefix. Block (DLM-style) drafters predict the whole block in parallel, which is fast but trained with a full-block cross-entropy that supervises every position against the gold continuation -- even though inference discards every token after the first rejection. Recent acceptance-aware objectives patch this by reweighting the full-block loss; we instead use teacher-forced learning as a motivation for how supervision should concentrate on the accepted prefix. A mask-only block drafter has no input-side channel for gold-prefix conditioning, so AUF approximates that prefix-sensitive supervision on the loss side by keeping the cross-entropy support only through the drafter's first predicted failure. AUF is a single, detached change to the CE support -- no auxiliary objective, no verifier rollouts, and no change to the inference pipeline or the exactness contract. Within fixed drafter backbones and serving settings on Qwen3-8B, AUF raises the DFlash drafter's average emitted length $\tau$, averaged over six benchmarks, from 2.40 to 2.61, with a gain on every benchmark, and transfers to Domino's two-branch head (2.56 to 2.68). Two findings sharpen the picture: the decay-only baseline reaches higher token accuracy on the shared block mask yet decodes worse, and on DFlash, once AUF truncates the support, the standard exponential position-decay weighting becomes empirically inert.
