Why is it important to exclude the bias from the gate w...

The bias is intended as a load-control signal, not a quality signal. If the bias were included in the gate weight computation, an expert that was promoted due to low affinity would also receive a high gate weight. This would amplify the expert’s output regardless of whether its FFN has actually learned useful representations for those tokens. Keeping the gate weight equal to the affinity score ensures quality is tied to learned competence.