Reproduce “Reciprocal Guidance: Orchestrating Draft and Verify Budgets for Advancing the Diffusion-AR Self-Speculation Frontier” 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
Linye Wei, Shutian Zheng, Haoyu Zeng, Meng Li
RecGuide improves serving throughput in self-speculative decoding by preparing likely correction and continuation branches during verification and adapting request-specific draft block sizes to load. On NLD-8B with AIME25 at concurrency 1 and block size 16, it reaches 4.38× speedup over autoregressive decoding, versus 2.42× for LinearSS. Tests cover two model sizes and eight benchmarks on one B200.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.