Ir al contenido principal

Detecting a Route Flip Is Easier Than Knowing Whether to Fix It: Causal Route-Mediated Damage in Quantized Mixture-of-Experts

Abstract

arXiv:2608.11212v1 Announce Type: new Abstract: Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance -- simulated 4-bit KV-cache quantization read by a protected BF16 gate -- pushes tokens across decision boundaries and flips which experts fire. This paper proposes no new mitigation; it supplies a causal apparatus, empirical findings, and a detection-limit result. A four-run apparatus prices the route-mediated fraction (RMF) of quantization dam

Transparencia: Este análisis ha sido generado con asistencia de inteligencia artificial bajo supervisión editorial de SAPIENSDATAAI.

Cookies esenciales

Necesarias para el funcionamiento del sitio. No se pueden desactivar.

Cookies analíticas

Nos permiten medir el tráfico y mejorar el sitio (Google Analytics).

Más info: Política de Cookies

Agente comercial · SAPIENSDATAAI

Cuéntanos tu proyecto sin salir de la web.