Search papers, labs, and topics across Lattice.
This paper introduces AutoPrune, an AI4AI framework that leverages large language models (LLMs) to automate the design of visual-token pruning algorithms, addressing the limitations of existing heuristic-based methods. By developing a Token Pruning Domain-Specific Language (TPDSL) with 131 reusable components, the framework effectively narrows the search space and enhances the LLM's ability to generate optimal pruning policies. Experimental results across 14 multimodal benchmarks reveal that AutoPrune can eliminate up to 94.4% of visual tokens while maintaining over 99% of full-token performance, achieving significant reductions in FLOPs and prefill latency.
AutoPrune enables LLMs to autonomously design visual-token pruning strategies, achieving a remarkable 9.9x reduction in FLOPs without sacrificing performance.
Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. As pruning objectives, budgets, and model architectures diversify, manually navigating the expanding design space becomes increasingly difficult. This paper aims to build an AI4AI framework for visual-token pruning by addressing a natural question: Can large language models automatically design effective visual-token reduction algorithms? Although LLMs possess broad algorithmic knowledge and strong reasoning capabilities, translating such general knowledge into effective solutions for a specialized task remains nontrivial. We argue that the key lies in designing an appropriate search-state representation that connects the internal knowledge of LLMs with the structural requirements and constraints of visual-token pruning. Based on this insight, we propose AutoPrune, a training-free framework for LLM-driven visual-token pruning policy design. At its core, AutoPrune introduces a Token Pruning Domain-Specific Language (TPDSL) comprising 131 reusable atoms for budget control, token scoring, selection constraints, and token reassembly. A key property of TPDSL is that it represents each search state as a residual modification of a strong base policy. This residual formulation narrows the search space and directs the LLM's attention toward the policy components that are most consequential for performance. Experiments on 14 multimodal benchmarks and three MLLM backbones demonstrate the effectiveness, efficiency, and transferability of AutoPrune. Even when removing 94.4% of visual tokens, AutoPrune preserves more than 99% of full-token performance while reducing FLOPs by 9.9x and prefill latency by 6.4x.
