Routing Sensitivity Without Controllability: A Diagnostic Study of Fairness in MoE Language Models

TL;DR

This paper introduces FARE, a diagnostic framework to evaluate the limits of routing-level fairness interventions in MoE language models, revealing structural and utility trade-offs.

Contribution

It provides a systematic analysis showing that routing sensitivity alone cannot reliably control stereotypes in MoE models, highlighting architectural constraints.

Findings

Routing-level preference shifts are often unachievable or non-robust.

Bias and knowledge are entangled within expert groups.

Routing sensitivity does not translate into improved generation fairness.

Abstract

Mixture-of-Experts (MoE) language models are universally sensitive to demographic content at the routing level, yet exploiting this sensitivity for fairness control is structurally limited. We introduce Fairness-Aware Routing Equilibrium (FARE), a diagnostic framework designed to probe the limits of routing-level stereotype intervention across diverse MoE architectures. FARE reveals that routing-level preference shifts are either unachievable (Mixtral, Qwen1.5, Qwen3), statistically non-robust (DeepSeekMoE), or accompanied by substantial utility cost (OLMoE, -4.4%p CrowS-Pairs at -6.3%p TQA). Critically, even where log-likelihood preference shifts are robust, they do not transfer to decoded generation: expanded evaluations on both non-null models yield null results across all generation metrics. Group-level expert masking reveals why: bias and core knowledge are deeply entangled within expert groups. These findings indicate that routing sensitivity is necessary but insufficient for stereotype control, and identify specific architectural conditions that can inform the design of more controllable future MoE systems.