Reproduce “On numerical issues of divergences between nearly equal distributions” 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
Brandon Amos
The paper presents eight numerical reformulations for categorical KL divergence, generalized Jensen–Shannon divergence, Monte Carlo KL estimates, and entropy. They preserve small log-ratio differences and rewrite cancellation-prone sums using nonnegative remainders. In an FP32 forward-KL example, the reformulation retains the value and gradient as distributions approach equality in the tested sweep, but cannot recover differences already lost in input rounding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.