ARTFEED — Contemporary Art Intelligence

Quantized MoE Routing Flips: Causal Analysis of Route-Mediated Damage

ai-technology · 2026-08-13

A new arXiv paper (2608.11212) investigates the impact of numerical disturbances on Top-k Mixture-of-Experts (MoE) routing, specifically focusing on 4-bit KV-cache quantization. The study reveals that such disturbances can cause tokens to cross decision boundaries, flipping which experts are activated. The paper introduces a causal apparatus to quantify the route-mediated fraction (RMF) of quantization damage, decomposes it by mechanism via token-level attribution, and validates findings across three architectures using pre-registered probes. On OLMoE-1B-7B, the pilot results show that approximately a third of the damage is routing-mediated (RMF ~ 0.31). The deployable router margin can detect that a flip occurred (AUC 0.772) but cannot distinguish harmful from helpful flips. The paper does not propose a mitigation but provides empirical findings and a detection-limit result.

Key facts

  • arXiv:2608.11212v1
  • Top-k Mixture-of-Experts (MoE) routing is discontinuous
  • Simulated 4-bit KV-cache quantization read by a protected BF16 gate
  • Four-run apparatus prices the route-mediated fraction (RMF)
  • Token-level attribution decomposes damage by mechanism
  • Pre-registered probes across three architectures
  • OLMoE-1B-7B at 4-bit KV: RMF ~ 0.31
  • Router margin detects flip (AUC 0.772) but cannot tell harmful from helpful

Entities

Institutions

  • arXiv

Sources