Speculative decoding accelerates autoregressive generation by using a small draft model to propose multiple tokens that a larger model verifies in parallel. Parallel drafters such as DFlash and DSpark reduce proposal latency, but the abstract notes that early mismatches limit how much of each block can be accepted, leaving potential speedups on the table.

TreeSpark, described in a new arXiv preprint, addresses this by learning to draft the trees you verify. Instead of using a static set of candidate continuations, the method appears to learn which tree structures are more likely to yield accepted tokens, thereby improving the efficiency of the verification step.

The abstract is truncated, so details of the learning procedure and experimental results are not available in the source. The title and opening suggest that the key contribution is a learned drafting strategy for tree-based verification, distinguishing it from earlier parallel drafters that do not adapt their proposal structure.