Generalize CSNB evaluation across MoE router gates and layers

Evaluate whether the limitation of Cauchy–Schwarz Norm Bounds for Silent Data Corruption detection extends to other Mixture-of-Experts router gates with different numbers of experts or to other linear layers within Mixture-of-Experts models.

Background

The paper finds that Cauchy–Schwarz Norm Bounds (CSNB) provide fault-detection capability comparable to V-ABFT for the evaluated dense linear layers but substantially lower capability for the Mixtral-8x7B router gate. The authors attribute this difference partly to the router’s small number of output columns and to the global CSNB bound’s lack of direct tightening as the number of output columns decreases.

Because the evaluation covers only one Mixture-of-Experts model and its router gate, it remains unresolved whether the observed limitation is specific to that configuration or generalizes to router gates with different numbers of experts and to other linear layers in Mixture-of-Experts architectures.

References

Whether this limitation extends to other MoE router gates with different numbers of experts or to other linear layers within MoE models has not been evaluated in this work.

— Revisiting Cauchy-Schwarz Norm Bounds for SDC Detection in LLM Inference  (2610.18557 - Altenbernd et al., 13 Oct 2026) in Section Discussion, paragraph “Verification depends on the protected computation”