MoE Models Show Fragile Representations Despite Redundancy.

A new study reveals a critical vulnerability in Mixture-of-Experts (MoE) models: their internal representations, while seemingly robust due to redundancy across multiple experts, are surprisingly fragile to noise. Researchers found that even though linear probes could recover information, such as moral valence, from nearly every expert-layer combination with high accuracy, these representations collapsed under noise levels that dense models of comparable size easily tolerated. This fragility is attributed to "output dilution." The mechanism by which MoE blocks average outputs from active experts before feeding them into the residual stream significantly reduces the signal strength. This diluted signal, nearly two orders of magnitude smaller than in a dense MLP, makes the encoded information highly susceptible to even minor perturbations. The research confirms that this issue is architectural, not a learned behavior, as it appears early in training and routing stability is not the problem. The implication is that redundant encoding in sparse architectures does not automatically translate to robust encoding, posing a challenge for the reliability of MoE models in noisy or adversarial environments.