vLLM extends adaptive verification to draft models without confidence heads

On October 14, vLLM reported an online-fitted acceptance estimator that adds confidence scoring to every draft-model-based speculator type, so draft models without a confidence head, such as EAGLE and MTP, can also use Adaptive Verification. The estimator derives features from the draft model's logits, predicts the target model's acceptance probability through online fitting, and feeds that prediction directly into adaptive verification as the confidence score.

Written by AI from the coverage · updated 1 hour ago

Not enough continuous observations to draw a trend yet.