Reproduce “Adversarial Speculative Reasoning Improves Long Horizon Agents” 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
Hongyin Luo, Wei Fang, Lijie Yang, Owen Stepan, Daniel Suit, Jeff Krapf, Dana Wensberg, Jack O'Brien
The authors use speculative decoding to steer selected decisions into less noisy context positions. A trained draft is more likely to be rejected at those points, so the target resamples at the start of a fresh verification block while preserving its token distribution. On DeepSWE, GLM-5.3-NVFP4 pass rate rises from 60.6% to 67.6%; not all benchmark gains are statistically established.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.