Does Role Specialization Matter for Explanation Faithfulness in Mixture-of-Experts?

Technical Summary: Does Role Specialization Matter for Explanation Faithfulness in Mixture-of-Experts?

Problem Statement
Mixture-of-Experts (MoE) architectures have recently been extended with role-based mechanisms to enhance interpretability in multimodal settings. These approaches assign semantic roles (e.g., uniqueness, redundancy, synergy) to individual expert components to decompose cross-modal interactions. However, the authors observe a critical gap: while structural role decomposition is explicitly defined, it remains unclear whether this structure preserves explanation faithfulness. In practice, experts assigned different semantic roles often learn highly overlapping representations due to shared encoders and identical task objectives. This inter-expert representation overlap weakens effective role separation, potentially rendering role assignments nominal and causing feature attributions to fail in reflecting the actual evidence used for predictions. The paper investigates whether reducing this representation overlap improves the faithfulness of explanations without compromising task performance.

Methodology
The authors propose a controlled study using a feature-level extension of the Interpretable Multimodal Interaction-aware MoE (I2MoE), termed FL-I2MoE, as their backbone. Instead of redesigning the architecture, they treat inter-expert representation overlap as a controllable factor.

1. Framework: The model processes multimodal inputs through shared encoders, feeding feature-level representations into $E$ experts ($M$ uniqueness, 1 synergy, 1 redundancy). Each expert produces class logits and a latent representation ($z_e$).
2. Regularization Strategy: To reduce overlap without altering the routing mechanism or attribution procedure, the authors introduce representation-level decorrelation regularization. They define an anchor latent representation for each expert (the output on the original input) and apply three distinct objectives to minimize similarity between expert pairs:
   • Rep-Cos: Minimizes squared cosine similarity between $L_2$-normalized latent vectors (instance-level directional alignment).
   • Rep-CKA: Minimizes Centered Kernel Alignment (CKA) between centered representation matrices (population-level structural similarity).
   • Rep-Barlow: Minimizes off-diagonal entries of the cross-correlation matrix between standardized latent dimensions (cross-feature redundancy).
3. Training Objective: The total loss combines the task loss ($L_{task}$), role-specific interaction losses derived from Partial Information Decomposition ($L_{interaction}$), and the chosen decorrelation loss ($L_{decor}$):
   $$L = L_{task} + \alpha L_{interaction} + \lambda L_{decor}$$
   Hyperparameter $\lambda$ is swept to identify regimes that reduce overlap while preserving predictive performance (within a 2% margin).

Key Contributions

• Identification of a Gap: The paper identifies that structural role decomposition alone is insufficient to guarantee faithful explanations if inter-expert representations remain overlapping.
• Decorrelation Mechanism: It introduces representation-level decorrelation as a mechanism to enforce clearer specialization among experts while retaining the original architecture and routing logic.
• Generalizability: The study demonstrates that the benefits of representation separation extend beyond explicitly role-based architectures (like I2MoE) to standard sparse MoE baselines (e.g., GShardGate), suggesting a general mechanism for enhancing explanation faithfulness.

Experimental Results
The authors evaluated their approach on three multimodal benchmarks: ENRICO (UI classification), MIMIC-IV (mortality prediction), and MMIMDb (movie genre classification).

• Task Performance: Across all datasets and decorrelation variants (Rep-Cos, Rep-CKA, Rep-Barlow), the models maintained predictive performance within 2% of the baseline, confirming that separation does not inherently degrade utility.
• Explanation Faithfulness: Using perturbation-based metrics (Comprehensiveness and Sufficiency) and their Area Over the Perturbation Curve (AOPC) summaries, the decorrelated models consistently outperformed the baseline. Specifically, the gap between "identified masking" (removing features highlighted by the explanation) and "random masking" increased, indicating that the explanations better captured decision-relevant evidence.
• Representation Geometry: Analysis showed that regularization reduced pairwise CKA and correlation between experts. In several cases, silhouette scores improved, indicating clearer cluster-level separation.
• Role Alignment: On a synthetic benchmark with known interaction structures, reducing overlap via Rep-CKA improved the alignment between routing weights and the ground-truth semantic roles (e.g., the synergy expert receiving higher weights for synergistic inputs).
• Ablation: An ablation study revealed that while role-based interaction learning alone did not improve faithfulness over a standard MoE, and decorrelation alone offered partial benefits, the combination of both yielded the highest faithfulness gains.

Significance and Claims
The paper concludes that structural role decomposition is not a sufficient condition for explanation faithfulness in MoE systems. The primary finding is that inter-expert representation overlap acts as a bottleneck that obscures effective specialization. By explicitly reducing this overlap through decorrelation regularization, the authors demonstrate that explanation faithfulness can be significantly improved while preserving task performance.

The authors modestly claim that their findings suggest representation-level separation provides a more general mechanism for enhancing faithfulness, applicable even to standard sparse MoE architectures lacking explicit semantic roles. They emphasize that the effectiveness of specific decorrelation objectives (Cosine vs. CKA vs. Barlow) varies by dataset and metric, implying that the relationship between separation and faithfulness is influenced by modality structure and interaction complexity. The work supports the hypothesis that for role-based MoE to be truly interpretable, the latent representations of the roles must be geometrically distinct.