Engineering

Quantized MoE route flips: easy to detect, hard to fix

Quantized MoE route flips silently corrupt expert selection. Detection is easy, but knowing whether to fix each flip requires causal damage analysis.

By Lars Cornelissen August 13, 2026 1 min read

𝕏 in

Donut chart showing that on fragile tokens, the standard MoE route is the best option on only 0.8 percent of tokens while a better alternative exists on 99.2 percent. Quantized MoE route flips concentrate damage on these fragile tokens. — On fragile tokens, the standard MoE router picks the best route only 0.8 percent of the time. Source: counterfactual routing analysis, arXiv 2605.07260. Data Today benchmark.

...

Bar chart comparing next-token probability gain from the best alternative route versus the standard route. Confident tokens show approximately 0.3 percentage points gain; fragile tokens show 20.4 percentage points gain. Quantized MoE route flips concentrate damage on fragile tokens. — Source: counterfactual routing analysis, arXiv 2605.07260. The fragile-token value of 20.4 pp is exact; the confident-token value of approximately 0.3 pp is approximate, reflecting the paper's finding that the standard route closely matches the best route on confident tokens. Data Today benchmark.

#moe #quantization #routing #inference #llm

Published by the Data Today newsroom · Subscribe via RSS