vLLM adds an online-fitted acceptance estimator for adaptive verification without a confidence head
Original titleAdaptive Verification Without a Confidence Head
vLLM says its new online-fitted estimator lets draft models without a confidence head use Adaptive Verification.
The estimator uses a logistic regression on the log-odds of the drafter's top token, with coefficients refit from the target model's accept/reject decisions during serving.
On DeepSeek-V4-Flash with DSpark, the post reports the estimator outperforms the trained confidence head on AUC and log-loss, and roughly matches it on calibration after excluding the first draft position.
The post explains why calibration matters for trimming decisions, and compares several confidence features on measured acceptance data, useful for speculative decoding engineers.
Source: vLLM Blog · vllm.aiPublished · added here