Dense-model quality versus pruning robustness at larger scales

Characterize the tradeoff between dense-model quality and structured-pruning robustness for BMuon at larger model scales.

Background

The language-model experiments show that BMuon can improve robustness to post-training structured MLP-channel pruning while sacrificing unpruned performance relative to Muon. This demonstrates a practical tradeoff between dense accuracy and compressibility.

The paper does not determine how this tradeoff behaves as model size, training scale, or data scale increases. Characterizing it at larger scales would clarify whether the observed loss in dense-model quality is intrinsic, how it depends on the budget and geometry parameters, and whether the robustness gains persist.