Recycling Wasted Logits from Speculative Decoding for On-Policy Training

ycombinator · x · 2026-08-12

The post discusses how to leverage 'wasted' data in LLM serving stacks to optimize models. During speculative decoding, the target model computes logits on rejected tokens that are typically discarded. However, these represent a free, on-policy training signal readily available from actual traffic.

The author notes that because draft models are often trained on generic datasets, their distributions mismatch specific real-world use cases (like voice or coding agents). This causes the acceptance rate to drift down over time, significantly hurting throughput. By recycling these wasted logits for online training, developers can fix the distribution mismatch and sustain inference speedups.

Original post →

More from Infra

Infra channel →