Relax Within, Balance Across: Geometry-Guided Load Balancing for Vision-Language Mixture-of-Experts

This paper argues that the standard Switch auxiliary loss, or Std-Aux, can hide large image-text routing errors because it balances only their mixed aggregate. The same router reportedly exhibits more than a fivefold change in load imbalance across image resolutions. The authors derive a fixed-profile load curve, identify the image-text load gap as the source of mix sensitivity, and propose ReBA: separate image and text balancing terms plus one equal-weight routing instance per image. Across four split backbones, ReBA lowers load on every reported benchmark input while keeping mean task accuracy comparable. The supplied abstract is truncated, so detailed benchmark values are unavailable.