Theoretical Foundation of Hard Top-k Primal Routing
Year: 2026
Type: report
Abstract: A pervasive misconception in deep learning router design is that discrete Top-k selection operates as an ad-hoc, non-differentiable heuristic necessitating continuous relaxations such as softmax smoothing. Here, we formalize the token-to-expert dispatch mechanism through convex optimization and network flow theory, demonstrating that the hard Top-k operator over duality-adjusted affinities is the exact, closed-form solution to a linear program, incurring a zero integrality gap.
Author: Rehn, Carl Johan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234443023
