An AI4AI Framework for Visual Token Pruning

Hugging Face Daily Papers Papers

Summary

AutoPrune is a training-free framework that uses LLMs to automatically design visual-token pruning policies for multimodal LLMs via a domain-specific language and residual search, achieving high efficiency with minimal performance loss (99% performance retained while removing 94.4% of visual tokens).

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.
Original Article
View Cached Full Text

Cached at: 08/14/26, 07:26 AM

Paper page - An AI4AI Framework for Visual Token Pruning

Source: https://huggingface.co/papers/2608.07193

Abstract

AutoPrune uses large language models to automatically design visual-token pruning policies for multimodal models via a domain-specific language and residual search formulation, achieving high efficiency with minimal performance loss.

Visual-token pruningcan substantially reduce the inference cost ofmultimodal 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 forvisual-token pruningby 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 ofvisual-token pruning. Based on this insight, we proposeAutoPrune, a training-free framework for LLM-drivenvisual-token pruningpolicy design. At its core,AutoPruneintroduces aToken Pruning Domain-Specific Language(TPDSL) comprising 131 reusable atoms for budget control, token scoring, selection constraints, and token reassembly. A key property ofTPDSLis that it represents each search state as aresidual modificationof a strongbase 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 ofAutoPrune. Even when removing 94.4% of visual tokens,AutoPrunepreserves more than 99% of full-token performance while reducingFLOPsby 9.9x andprefill latencyby 6.4x.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2608\.07193

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.07193 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.07193 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.07193 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Optimizing Korean-Centric LLMs via Token Pruning

arXiv cs.CL

This paper presents a systematic benchmark of token pruning—a compression technique that removes tokens and embeddings for irrelevant languages—applied to Korean-centric LLM tasks. The study evaluates popular multilingual models (Qwen3, Gemma-3, Llama-3, Aya) across different vocabulary configurations and finds that token pruning significantly improves generation stability and reduces memory footprint for domain-specific deployments.

Fine-tuning Multi-modal LLMs with ART: Art-based Reinforcement Training

Hugging Face Daily Papers

ART (Art-based Reinforcement Training) enables parameter-efficient fine-tuning of frozen multimodal LLMs by optimizing raw visual input via gradient backpropagation, achieving performance comparable to LoRA while supporting pre-compiled computational graphs for high-throughput engines like vLLM.

Small LLMs: Pruning vs. Training from Scratch

arXiv cs.LG

This paper empirically compares pruning vs. training small language models from scratch, finding that pruning provides a strong advantage under limited token budgets but that the advantage diminishes as training scales, especially with coarse pruning.