Speculative decoding speeds up inference by having a small draft model propose continuations that a larger target model verifies in one pass. Tree-based variants verify multiple draft paths at once, but a new preprint argues that finite trees built from draft scores are fundamentally mismatched with the target model's preferences.
The paper, posted on arXiv as 2610.11750, frames the issue as a draft-target mismatch and asks whether better exact verification can recover the lost probability mass. The title points to an exit-guided approach, though the abstract leaves the proposed mechanism unspecified.
Because the source is a single preprint abstract, the article does not compare findings with other work. The significance is diagnostic: it identifies a structural limitation in tree-based speculative decoding and opens a question about how verification should be guided.