Infrastructure & deployment
Trajectory-Level Speculative Decoding for Diffusion Language Models
The authors propose speculating over denoising trajectories—not just token sequences—to address low-throughput decoding in diffusion language models.
Why it matters: Trajectory-level speculation targets the throughput lost when low-confidence dLLM decoding becomes single-token-at-a-time. The reported gains come with trajectory drift as the fundamental cost of increased parallelism, so efficiency and accuracy must be considered together.