Researchers introduce Verification-Aware Training (VAT), a framework that improves speculative decoding for large language models. VAT adds a lightweight verification head that trains the draft model to predict which token positions will survive sequential verification, paired with adaptive weighting that maintains full emphasis until the first rejection point. Evaluated with EAGLE-3 and DFlash across multiple model sizes, VAT achieved acceptance-length improvements of up to 11.4 percent and wall-clock speedups of up to 8.7 percent.
