Reproduce “Probe-Space Preconditioning for Fast and Stable Zero-Order Training” 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
Francois Chaubard, Mykel J. Kochenderfer, Chris Ré
1.5-SPSA combines more probes per step with curvature-weighted updates for inference-mode zero-order fine-tuning without backpropagation’s memory costs. Directional curvature reduces sharp probes’ influence, while larger batches shift work into each step. On OPT-13B SST-2, it reaches 94.5% accuracy in 70 steps and 179,000 forward-passes, versus MeZO’s 91.4%; this does not establish faster wall-clock training or universal superiority to backpropagation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.