Speculative decoding accelerates autoregressive generation by having a small drafter propose multiple tokens at once, which the target model then verifies. But feature-based drafters rely on the target model's hidden features, and those features may not emphasize what the next drafting step needs.

The ReaDy paper proposes using recurrent states as a dynamic context for drafting in hybrid linear attention LLMs. Rather than conditioning on target hidden features, the drafter draws on recurrent states that carry forward context in a way that better matches the drafting process.

The abstract is truncated, so the paper's full results and implementation details are not available from the source. Still, the idea highlights a design issue in speculative decoding: the drafter and target model may have different contextual needs, and bridging that gap could improve generation efficiency.