An arXiv preprint, "From Spectral to Frobenius: When Muon Loses Its Spectral Implicit Bias," examines how the Muon optimizer's implicit bias changes with the problem setting. Implicit bias refers to the direction that an optimization algorithm tends to select among many possible solutions, even without explicit regularization.
Prior work established that gradient descent on linearly separable problems converges to the Frobenius max-margin direction. More recently, Fan et al. (2025) showed that spectral-norm steepest descent exhibits spectral max-margin behavior, meaning the geometry of the norm used by the optimizer matters. The new preprint connects these two lines of work by asking when Muon behaves like one or the other.
According to the abstract, Muon does not always retain the spectral implicit bias. The paper's contribution is to characterize the transition from spectral to Frobenius behavior. This matters for practitioners using Muon, since the implicit bias can affect generalization and solution structure, though the full details of the analysis are not available in the abstract.