Reproduce “DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shreyas Kulkarni, Uttam Gurram, Madhav Kanda, Krishna Teja Chitty Venkata, Shichu Chen, Kartik Balasubramaniam, Supriyo Chakraborty, Sambit Sahu
DStill trains a one-pass speculative decoding drafter from teacher refinement trajectories to predict its endpoint from the root. Its predictions remain factorized. On GPT-OSS-120B across six benchmarks, it raises mean acceptance length from 4.45 to 4.79 and mean speedup from 3.45× to 3.68× versus DFlash, at concurrency 1, temperature 0, reasoning on/medium.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.