Search papers, labs, and topics across Lattice.
This study investigates the encoding of moral content in Mixture-of-Experts (MoE) models, revealing that while they can recover moral valence with high accuracy, they are significantly more fragile than dense models. The authors demonstrate that the output dilution caused by averaging across active experts leads to a dramatic reduction in the feedforward signal, making moral representations vulnerable to activation noise. Ultimately, the findings highlight that redundant encoding in sparse architectures does not guarantee robustness, as evidenced by a 4.2-fold difference in noise tolerance compared to dense models.
MoE models may encode moral content robustly, but they are surprisingly fragile, collapsing under noise that dense models easily withstand.
Mixture-of-Experts (MoE) models appear to encode moral content as robustly as dense models, yet prove far more fragile in their encoding. In OLMoE-1B-7B, linear probes recover moral valence from nearly every expert-layer combination, with mean peak-layer accuracy above 90%. But these representations collapse under levels of activation noise that a dense model of matched size easily tolerates, with a 4.2-fold difference in robustness. We trace this to output dilution. Because the MoE block averages across active experts before contributing to the residual stream, the feedforward signal reaching downstream layers is nearly two orders of magnitude smaller than in a dense MLP. Moral information, our interest, survives aggregation intact but at a scale trivially overwhelmed by perturbation. Routing itself remains stable under noise while the vulnerability originates entirely in the diluted aggregate. Checkpoint trajectories confirm this is architectural, not learned. Experts never specialize and accuracy saturates within the first few thousand steps. In sparse architectures, redundant encoding does not imply robust encoding.
