ReAnchor: Reusing Rejected Verifier States for Diffusion Drafting

Key points

• Meanwhile, target verification computes hidden states even at positions beyond the first rejection, but these states are typically discarded rather than reused.1
• We propose ReAnchor, a training-free method that reanchors a pretrained drafter at shifted positions using discarded verifier states to generate parallel draft blocks.1
• These blocks improve predictions within the original block and extend coverage beyond its endpoint.1
• At sampling temperature of T=0, ReAnchor achieves its largest average gains on Gemma 4-12B, improving acceptance length by 19.4% and decoding speed by 17.9% across eight benchmarks.1

Sources (1)

• [1]ReAnchor: Reusing Rejected Verifier States for Diffusion Drafting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 02:57 PM

  Block diffusion drafters accelerate speculative decoding by generating multiple draft tokens in a single forward pass.

  Meanwhile, target verification computes hidden states even at positions beyond the first rejection, but these states are typically discarded rather than reused.

Extractive summary: sentences quoted from the sources.