χisao — GPU optimizer finds all modes in multimodal black-box functions
New arXiv paper introduces χisao, a GPU-native population optimizer designed to find all modes of a multimodal black-box function in parallel, rather than sequentially like basin-hopping or CMA-ES.
The key insight: oscillate between convergence and anticonvergence cycles to escape local traps while freezing confirmed peaks, so samples don't waste compute revisiting the same modes.
Relevant for Bayesian inference, scientific computing, and any optimization task where finding the full landscape matters more than a single global optimum.