Sparse routing reduces per‑token computation, yet the full expert bank must stay resident in memory. For speech‑centric omni‑modal models, conventional pooling calibration statistics mask the differing expert demands of acoustic input processing and text answer generation. We introduce PW‑Pruning, a phase‑aware extension of DO‑ACP: it builds separate importance‑weighted kernels for acoustic‑prefill and teacher‑forced answer tokens, then merges phase‑weighted log‑determinant marginal gains $$\Delta_{phase}=\log\det(K_{phase})$$ to select a single static expert bank. On the public Qwen3‑Omni backbone, retaining 32 out of 128 experts per layer improves MMAU‑mini and MMSU accuracy by 3.20 % and 4.38 % respectively after supervised fine‑tuning over three runs. Single‑phase and shuffled‑label controls indicate that semantic phase information drives these gains. Complementary experiments on a private Qwen3.5‑Omni backbone and zero‑shot English evaluation of Nemotron corroborate the findings. Across tested Qwen budgets, AU gains increase with stronger pruning, without requiring inference‑time mask switching.
Review: PW‑Pruning’s explicit separation of acoustic and textual phases achieves notable memory savings while boosting performance, offering a practical pathway for efficient deployment of speech‑text multimodal systems.