Ziang Wu, Peng Jin, Qishen Yin, Munan Ning, Hao Li, Peizhen Zhang, Li Yuan
4 min
Abstract
Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux balances only the mixed load, so large image and text load errors can cancel at one mix. On our main model, the same trained router shows more than a fivefold change in load imbalance across image resolutions. We hold the image and text load profiles fixed and derive the exact load curve as the token mix varies. The image-text load gap controls sensitivity to the token mix. Physical preprocessing can also change the conditional profiles. The fixed-profile law excludes such changes. To design a remedy, we examine the router input structure. Image and text occupy distinct regions, while visual tokens group strongly by source image. The modality boundary motivates separate image and text terms. The image boundary motivates one equal-weight routing instance per image. ReBA, or Relax Within, Balance Across, implements both choices. Across four split backbones, ReBA lowers load on every reported benchmark input while keeping mean task accuracy comparable to Std-Aux. ReBA also lowers average load over the tested range and worst physical load under resolution and tiling shifts. Code is available at https://github.com/ZiangWu-77/ReBA.
Sam: Exactly. That's the core of why the model stays stable when resolution changes. By counting votes rather than raw tokens, the routing decisions don't get skewed just because one image happens to be more detailed than another.
Alex: It sounds like a meaningful improvement—not about more processing power, but better organization of the power you already have.
Sam: That's a fair way to put it. By aligning what the model is trained to optimize with the actual structure of the data it handles, they've created a more stable and predictable system. The experiments showed this lowers what the paper calls the "load gap"—the difference between how busy the busiest specialist is versus the least busy—consistently across different model sizes.
Alex: And does this better organization come at a cost to actual performance?
Sam: The paper reports that task accuracy remains comparable to the standard approach. The primary gain is in stability and efficiency—making sure resources are used as intended. It doesn't make the model more capable in the traditional sense, but it does make it more reliable under the kind of varied, real-world conditions where these systems are actually deployed.
Alex: So it's less about building a smarter system and more about making sure the system you've built behaves consistently—even when the inputs change in ways it didn't see during training.
Sam: That's precisely it. And for anyone building or deploying these systems at scale, that kind of predictability matters quite a lot. A model that performs well on average but collapses under certain input conditions is a much harder problem to manage than one that's simply somewhat less capable overall.
Alex: Thanks for walking us through that, Sam. And thank you for listening to ResearchPod.