Reproduce “Exploring Weight-Space Programming: Capabilities, Generalization, and Limitations” 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
Junru Hu, Cheng Tang, Wei Li, Hongze Zhu, Chun-Mei Feng, Qi Zhang, Yirong Chen, Ming Hu
Weight-Space Programming adapts a model by directly changing selected weights, without gradients or an optimizer. An agent uses development examples to propose and test edits, then evaluates a selected checkpoint on held-out data. For Qwen3.5-0.8B, one edit changed 1,023 BF16 values and raised MATH-500 accuracy from 8.35% to 37.80%; this task-specific result does not establish general replacement of training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.