Distilled Large Language Model-Driven Dynamic Sparse Expert Activation Mechanism

Large-scale vision-language mixture-of-experts (VL-MoE) models provide strong multimodal capability, but efficient deployment on memory-constrained platforms remains difficult. Existing MoE offloading systems are largely designed for text-centric workloads and become much less effective for visual-heavy inputs, where large numbers of visual tokens induce broader and less predictable expert accesses. We present VisMMoE, a VL-MoE offloading system built on a single systems insight: pruning redundant visual tokens can improve offloading not only by reducing computation, but also by reshaping expert demand. We refer to this effect as visual-expert affinity: token pruning makes expert accesses more concentrated within layers and more stable across layers, producing a smaller and more predictable expert working set. Guided by this insight, VisMMoE combines affinity-aware token compression, lookahead expert prediction, and cache/pipeline orchestration to improve expert locality and prefetch effectiveness under tight memory budgets. We implement VisMMoE on multiple frameworks and evaluate it on representative VL-MoE models and benchmarks. VisMMoE improves end-to-end inference performance by up to 2.68x and 1.61x, respectively, over strong baselines for today's VL-MoE deployments while maintaining competitive accuracy.

Unifying the complementary strengths of diverse Vision Foundation Models (VFMs) into a single efficient model is highly desirable but challenged by the negative transfer inherent in monolithic distillation. To address these feature conflicts, we introduce PRISM, a novel dual-stream Mixture-of-Experts (MoE) framework that synergizes VFMs via modular specialization. We propose a two-stage paradigm: (1) expertise deconstruction, where a teacher-conditional router guides experts to specialize in distinct representational subspaces to mitigate interference, followed by (2) dynamic recomposition, where the router learns to assemble these experts into tailored computational pathways for downstream tasks. Experiments on PASCAL-Context and NYUD-v2 show that PRISM establishes a new state of the art, validating that sparse, emergent specialization is a scalable approach for integrating diverse visual knowledge.

Mixture-of-Experts (MoE) architectures enable conditional computation by activating only a subset of model parameters for each input. Although sparse routing has been highly effective in language models and has also shown promise in vision, most vision MoE methods operate at the image or patch level. This granularity is poorly aligned with object detection, where the fundamental unit of reasoning is an object query corresponding to a candidate instance. We propose Hierarchical Instance-Conditioned Mixture-of-Experts (HI-MoE), a DETR-style detection architecture that performs routing in two stages: a lightweight scene router first selects a scene-consistent expert subset, and an instance router then assigns each object query to a small number of experts within that subset. This design aims to preserve sparse computation while better matching the heterogeneous, instance-centric structure of detection. In the current draft, experiments are concentrated on COCO with preliminary specialization analysis on LVIS. Under these settings, HI-MoE improves over a dense DINO baseline and over simpler token-level or instance-only routing variants, with especially strong gains on small objects. We also provide an initial visualization of expert specialization patterns. We present the method, ablations, and current limitations in a form intended to support further experimental validation.