Reproduce “Program Bindings Inside Code Models” 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
Prateek Kumar Rajput, Yewei Song, Jacques Klein, Tegawendé F. Bissyandé
The authors test whether static-analysis roles can reconstruct code-model states while preserving behavior on a synthetic Python list-execution task. They pair token content with compiler-derived structural labels and compare the result with an equally sized positional decomposition. In StarCoder2-15B, exact-output accuracy is 79.0% versus 52.0%; structural roles do better in 10 of 14 models, with stated limits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.