Frontier MoE Models Lack Robust Expert Modularity.

Mixture-of-Experts (MoE) models, which route tokens to a subset of specialized experts, are often hypothesized to possess functional modularity, where individual experts are responsible for specific capabilities or languages. This research rigorously tested this hypothesis on Command A+, a frontier open-weights MoE model. The study employed a pre-registered causal test, building a routing-mass atlas and ablating expert families at inference time. They measured whether ablating a family selectively impaired its hypothesized function, comparing it against a size-matched random-expert null. Crucially, the same families were tested across four metrics and an independent corpus with bootstrap confidence intervals. The findings are cautionary: robust functional modularity proved rare and highly sensitive to the measurement approach. Out of six pre-registered families, only one (Arabic language) demonstrated clean, selective modularity that held up across independent corpora and conservative statistical criteria. Other families showed real causal effects but failed selectivity, with their apparent modularity shifting based on the corpus, metric, and statistical bar used. A positive control on Qwen3-30B-A3B successfully recovered its known disjoint structure, validating the methodology. The results were consistent even with the un-quantized BF16 model, ruling out quantization artifacts.