A new arXiv preprint investigates a practical question for compressing Mixture-of-Experts (MoE) language models: can the way experts are routed during different task phases reveal which tensors are safe to quantize at lower precision? The authors focus on per-layer, per-task-phase routing statistics, covering prompt, reasoning, and response phases, and consider using them to guide layer-wise mixed-precision quantization of routed-expert feed-forward network (FFN) tensors.
According to the paper's title, the answer is no: task-phase-conditioned expert routing is not a reliable proxy for quantization sensitivity. The abstract frames this as an investigation rather than presenting a positive method, and the title makes the negative conclusion explicit.
The source is a single arXiv abstract, so the full experimental setup and quantitative results are not available in the provided text. Still, the stated finding is a caution for approaches that assume routing behavior can stand in for direct sensitivity analysis when quantizing MoE models.