Verification-Aware Training for Speculative Decoding
The paper trains draft models around the point where speculative decoding actually starts throwing tokens away.
VAT simulates verification during training, then uses accept/reject patterns as supervision. It adds a lightweight verification head and changes the loss weighting so positions before the first rejection keep full weight. The authors say it can sit on top of EAGLE-3 and DFlash without changing architecture or inference. On tested Qwen3 and LLaMA-3.1 models, they report up to 11.4% longer average acceptance and up to 8.7% wall-clock speedup. HF Daily Papers' note
VAT simulates verification during training, then uses accept/reject patterns as supervision. It adds a lightweight verification head and changes the loss weighting so positions before the first rejection keep full weight. The authors say it can sit on top of EAGLE-3 and DFlash without changing architecture or inference. On tested Qwen3 and LLaMA-3.1 models, they report up to 11.4% longer average acceptance and up to 8.7% wall-clock speedup. HF Daily Papers' note
score 5