Introduces stage-aware sparsity via Visual Token Compressor for modality alignment and Layer Dynamic Skipper for instruction tuning to improve MLLM training efficiency.
A multi- level framework for accelerating training transformer models,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Improving MLLM Training Efficiency via Stage-Aware Sparsity
Introduces stage-aware sparsity via Visual Token Compressor for modality alignment and Layer Dynamic Skipper for instruction tuning to improve MLLM training efficiency.