Reproduce “ProbMoE: Differentiable Probabilistic Routing for Mixture-of-Experts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Heng Zhao, Zilei Shao, Guy Van den Broeck, Zhe Zeng
ProbMoE introduces a differentiable probabilistic routing mechanism for Mixture-of-Experts models, formulating expert selection as an inference problem over cardinality-constrained subsets. This approach enhances routing diversity and expert utilization, achieving competitive or superior performance on tasks like mathematical reasoning while enabling adaptive compute allocation.
Key insights
- Modeling expert selection as probabilistic inference allows for a principled, differentiable learning signal to the router, overcoming the limitations of discrete top-k approximations.
- Tractable computation of conditional marginal probabilities enables effective gradient flow through the expert selection process, promoting broader expert utilization and specialization.
- Adaptive (Dynamic-k) expert allocation can reduce computational cost without significant performance degradation, demonstrating the ability to dynamically adjust compute based on token or task complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.