What it is
Lossless speculative decoding for diffusion language models by predicting commit order, not just tokens, then verifying drafted state trees exactly under sampling. Unlike prior DLM drafters that guess full future states or tokens, SPOT learns the missing variable: commit order. It reuses stale target token predictions, drafts state trees from output distributions only, and uses shared-noise verification to stay exact at any temperature.
Key results
- Up to 3.4x throughput vs standard decoding on SDAR-8B
- Uses 16-36% fewer target forwards than Spiffy on SDAR-8B
- SPOT-Sets adds 59% avg throughput over threshold decoding alone
For diffusion LM decoding, predicting where to commit next matters more than predicting the token itself.