Block diffusion drafters accelerate speculative decoding by producing several draft tokens in a single forward pass. However, each pass is constrained to a fixed-length block derived from a single target-generated anchor, which limits how far the drafting process can proceed.
ReAnchor, as described in a new arXiv announcement, proposes to reuse rejected verifier states to overcome this limitation. By re-anchoring with states that were previously rejected, the method may allow drafting to continue further than a single fixed block permits.
The abstract is brief and does not detail the exact mechanism, but the core idea is clear: instead of discarding rejected verifier states, ReAnchor repurposes them to extend the drafting horizon. This positions the method as a potential efficiency gain for speculative decoding systems.