An AI4AI Framework for Visual Token Pruning
AutoPrune has an LLM design visual-token pruning policies instead of relying on handcrafted rules.
The paper introduces a training-free framework built around TPDSL, a pruning-specific language with 131 reusable policy atoms. Its search states are framed as residual changes to a strong base policy, narrowing what the LLM has to explore. In tests across 14 multimodal benchmarks and three MLLM backbones, the authors report that AutoPrune kept more than 99% of full-token performance after removing 94.4% of visual tokens, with 9.9x lower FLOPs and 6.4x lower prefill latency. HF Daily Papers' note
The paper introduces a training-free framework built around TPDSL, a pruning-specific language with 131 reusable policy atoms. Its search states are framed as residual changes to a strong base policy, narrowing what the LLM has to explore. In tests across 14 multimodal benchmarks and three MLLM backbones, the authors report that AutoPrune kept more than 99% of full-token performance after removing 94.4% of visual tokens, with 9.9x lower FLOPs and 6.4x lower prefill latency. HF Daily Papers' note
score 4