Improving MLLM Training Efficiency via Stage-Aware Sparsity

In three linesSparse Training Scheme (STS) improves MLLM training efficiency through stage-aware sparsity: Visual Token Compressor reduces visual token load during modality alignment, Layer Dynamic Skipper skips unnecessary layers during instruction tuning. Framework adapts to varying redundancy across training stages.

Summary generated by Claude — human-verified