[R] KALAVAI: Predicting When Independent Specialist Fusion Works (gain = 0.82 × divergence − 2.72, R² = 0.856, tested 410M–6.9B)

Hey all, I've been working on this for a few months and just put the paper on arXiv: https://arxiv.org/abs/2603.22755 Project page: https://murailabs.com/kalavai/ Code + scripts: https://github.com/mechramc/Kalavai The basic idea: take a base checkpoint, give copies to a bunch of people, each person