# PW-Pruning: Phase-Aware Expert Selection for Speech-Centric Omni-Modal Models

arXiv: https://arxiv.org/abs/2610.26891 · HTML: https://mlsanity.com/paper/2610.26891/

## Key facts

- No citing papers in the map yet (data through 19 Oct 2026).
- Builds on 19 of its 27 references that are in the map.
- First posted on arXiv 15 Oct 2026.

## Builds on (most cited first, 19 of 19)

- [Mixtral of Experts](https://mlsanity.com/paper/2401.04088/) (2024)
- [A Call for Clarity in Reporting BLEU Scores](https://mlsanity.com/paper/1804.08771/) (2018)
- [Common Voice: A Massively-Multilingual Speech Corpus](https://mlsanity.com/paper/1912.06670/) (2019)
- [Qwen3-Omni Technical Report](https://mlsanity.com/paper/2509.17765/) (2025)
- [FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech](https://mlsanity.com/paper/2205.12446/) (2022)
- [MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark](https://mlsanity.com/paper/2410.19168/) (2024)
- [Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models](https://mlsanity.com/paper/2402.14800/) (2024)
- [MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix](https://mlsanity.com/paper/2505.13032/) (2025)
- [MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark](https://mlsanity.com/paper/2506.04779/) (2025)
- [Task-Specific Expert Pruning for Sparse Mixture-of-Experts](https://mlsanity.com/paper/2206.00277/) (2022)
- [Opencpop: A High-Quality Open Source Chinese Popular Song Corpus for Singing Voice Synthesis](https://mlsanity.com/paper/2201.07429/) (2022)
- [REAP the Experts: Why Pruning Prevails for One-Shot MoE compression](https://mlsanity.com/paper/2510.13999/) (2025)
- [DisPose: Disentangling Pose Guidance for Controllable Human Image Animation](https://mlsanity.com/paper/2412.09349/) (2024)
- [HeartMuLa: A Family of Open Sourced Music Foundation Models](https://mlsanity.com/paper/2601.10547/) (2026)
- [Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations](https://mlsanity.com/paper/2507.12197/) (2025)
- [VideoGuard: Protecting Video Content from Unauthorized Editing](https://mlsanity.com/paper/2508.03480/) (2025)
- [UVCG: Leveraging Temporal Consistency for Universal Video Protection](https://mlsanity.com/paper/2411.17746/) (2024)
- [Pruning and Distilling Mixture-of-Experts into Dense Language Models](https://mlsanity.com/paper/2605.28207/) (2026)
- [Comprehend and Talk: Text to Speech Synthesis via Dual Language Modeling](https://mlsanity.com/paper/2509.22062/) (2025)

## Abstract (arXiv)

Sparse routing reduces per-token computation but leaves the full expert bank resident in memory. We study expert pruning for speech-centric omni-modal models, where pooling calibration statistics can obscure the different expert demands of acoustic-input processing and text-answer generation. We propose PW-Pruning, a phase-aware extension of DO-ACP that constructs separate importance-weighted kernels for acoustic-prefill and teacher-forced answer tokens, then combines phase-weighted log-determinant marginal gains to select one static expert bank. On Qwen3-Omni, retaining 32 of 128 experts per layer improves MMAU-mini and MMSU accuracy over matched pooled selection by 3.20 and 4.38 percentage points, respectively, after supervised fine-tuning, averaged over three runs. Single-phase and shuffled-label controls on this public backbone suggest that semantic phase information contributes to these gains. Experiments on a non-public Qwen3.5-Omni backbone and English-only zero-shot evaluation of Nemotron provide complementary evidence. Across the tested Qwen budgets, AU gains are larger under stronger pruning, without inference-time mask switching.