A new study finds that Mixture-of-Experts AI models give more reliable self-explanations when their internal components are actively pushed apart in representation space.
Mixture-of-Experts architectures split processing across multiple specialized sub-networks called experts. Recent work tried to make these models more interpretable by assigning each expert a semantic role — synergy, redundancy, uniqueness, and so on. The problem, this paper argues, is that those role labels are mostly decorative: if two experts end up learning similar internal representations anyway, their explanations bleed together and lose meaning. The researchers tested a fix called representation-level decorrelation regularization, which adds a training objective that penalizes experts for overlapping in latent space, forcing genuinely distinct specialization. Across multiple multimodal benchmarks, the technique improved explanation faithfulness — measured by comprehensiveness, sufficiency, and a curve-area metric called AOPC — without hurting task performance.
The finding matters because "explainability" has become a compliance word as much as a technical one. If the explanations a model produces don't actually reflect how it reached a decision, they're worse than useless — they're cover. This work suggests that slapping role names onto experts is not enough and that the fix has to happen at the level of learned representations, not architecture labels. Notably, the gains showed up in standard sparse MoE models too, not just the role-based variant the researchers started with.
The code is public on GitHub, which is a reasonable sign of confidence. What remains untested is whether the same regularization holds up at the scale of the largest production MoE models, where the number of experts — and the pressure to let them drift together — is orders of magnitude higher.