Reproduce “Alignment via Generalization” 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
Shai Shalev-Shwartz, Amnon Shashua
Aligned-PAC treats alignment as a generalization problem, representing requirements as proxies with separate losses, example distributions, and thresholds. For a fixed hypothesis, its multilevel estimator measures the gap between training-proxy and population-proxy violations using O(ε⁻² log(1/δ)) fresh examples and proxy draws, with matching worst-case lower bounds. Guarantees apply only to the selected proxy distribution, not human intent or unseen-proxy generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.