Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts
The paper explores the specialization of experts in Vision Mixture-of-Experts (MoE) models. It highlights the importance of analyzing expert tuning and representation beyond just routing categories. The findings suggest that expert specialization is more complex and involves broader tuning to various visual and semantic dimensions.
- ▪Mixture-of-Experts models are analyzed to understand expert specialization beyond routing categories.
- ▪The study employs a contrastive objective on natural images to characterize expert tuning using visual neuroscience tools.
- ▪Results indicate that an animate-inanimate distinction is a dominant factor in expert partitioning, stable across independently trained models.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20610 (cs) [Submitted on 20 May 2026] Title:Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts Authors:Gene Tangtartharakul, Katherine R. Storrs View a PDF of the paper titled Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts, by Gene Tangtartharakul and Katherine R. Storrs View PDF HTML (experimental) Abstract:Mixture-of-Experts (MoE) models are often interpreted by analysing which categories are routed to which experts. However, routing alone does not reveal what each expert actually encodes.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
