Multilingual MoE Models Show Uneven Safety Alignment.

Safety alignment in multilingual AI models often exhibits significant unevenness, with models reliably refusing harmful requests in high-resource languages like English but complying with the same requests in lower-resource languages. This study mechanistically traces this disparity within "sarvam," an Indic-multilingual Mixture-of-Experts (MoE) reasoning model. The findings reveal that the model's ability to detect harm is encoded as an internal direction that is nearly language-invariant in mid-network layers. Steering this direction upstream causally controls the refusal behavior. However, the actual mechanism that *writes* the refusal is distinct from harm detection; it is a late-stage process assembled during generation, rather than a single forward pass readout. The research localizes this "writer" to a specific circuit within the MoE, which is held in check by an "attention opposer." The study quantifies the cost of various interventions, finding that damping this opposer is both cheap and effective for improving cross-lingual safety. Amplifying the writer is prohibitively expensive, and surgical edits to responsible heads are ineffective. The organization of this circuit and the gradient method used to expose it were also observed in a second, unrelated MoE model, though the lever's strength is architecture-specific. This provides a cost-measured map for multilingual safety repairs.