Reproduce “Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting” 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
Xinwei Qiang, Xiang Fang, Chang Chen, Zaifeng Pan, Yue Guan, Yufei Ding
The paper measures rejection forced by limited context in block speculative decoding, separating this information floor from a drafter’s model gap. Revealing one predecessor removes 86–100% of the order-0 floor across measured positions, but substantial model gaps remain. In Qwen3-4B SGLang tests, the complete prefix-attention system raises accepted length from 4.75 to 4.94 and speedup from 3.34× to 3.52×.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.