Papers with token pruning
Length-Adaptive Distillation: Customizing Small Language Model for Dynamic Token Pruning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to accelerate inference speed are model compression and dynamic computation (e.g., dynamic token pruning). |
| Approach: | They propose a two-stage knowledge distillation framework that produces a customized small language model for dynamic token pruning. |
| Outcome: | The proposed framework can make the small language model more customized for dynamic token pruning and achieve better speed-performance trade-off. |
Reducing Token Redundancy in LVLMs: A Systematic Review of Token Pruning Methods (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) excel at visual understanding but face severe computational bottlenecks when processing high-resolution images and long videos due to massive visual token counts. |
| Approach: | They propose a taxonomy categorizing methods into vision-side, LLM-side and hybrid paradigms and analyze token selection mechanisms and pruning strategy. |
| Outcome: | The proposed method selectively removes less informative tokens while maintaining performance. |
Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More (2025.emnlp-main)
Copied to clipboard
Zichen Wen, Yifeng Gao, Shaobo Wang, Junyuan Zhang, Qintong Zhang, Weijia Li, Conghui He, Linfeng Zhang
| Challenge: | Vision tokens in multimodal large language models often dominate computational overhead due to excessive length compared to linguistic modality. |
| Approach: | They propose a token pruning method which defines an importance criterion for vision tokens and prunes the unimportant vision token during inference. |
| Outcome: | The proposed method can prune 88.9% of vision tokens while maintaining comparable performance. |
From Inheritance to Saturation: Disentangling the Evolution of Visual Redundancy for Architecture-Aware MLLM Inference Acceleration (2026.acl-long)
Copied to clipboard
| Challenge: | Existing acceleration strategies suffer from severe "backbone dependency" Existing strategies such as token pruning or layer sparsity suffer from this . |
| Approach: | They propose a framework that decouples visual redundancy into IVR and architecture-dependent secondary saturation redundancies. |
| Outcome: | The proposed framework outperforms existing frameworks on Qwen25-VL and Qwa25-LL. |
Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem? (2025.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models have shown remarkable performance for cross-modal understanding and generation, yet suffer from severe inference costs. |
| Approach: | They propose to prune redundant tokens in MLLMs to reduce computation and storage costs. |
| Outcome: | The proposed method reduces the computational and storage costs of MLLMs by identifying redundant tokens and pruning them. |
Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring (2026.acl-long)
Copied to clipboard
| Challenge: | Existing efforts to mitigate this via token compression fail due to its autoregressive nature . linguistically redundant tokens are erroneously pruned, leading to hallucinations . |
| Approach: | They propose a method that reformulates token pruning as a Visual-Anchored Information Bottleneck (VA-IB) optimization problem. |
| Outcome: | Experiments on Qwen2-VL and Llama-3.2 families show that the proposed model achieves a speedup with negligible accuracy loss. |
Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained transformers suffer from a computationally expensive self-attention mechanism that interacts with all tokens, including those unfavorable to classification performance. |
| Approach: | They propose to integrate token pruning and token combining strategies to improve model performance and reduce computational demands. |
| Outcome: | Experiments with various datasets show that the proposed model performs better than baseline models, with the best improvement over the existing model. |
Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings (2025.findings-acl)
Copied to clipboard
Yubo Ma, Jinsong Li, Yuhang Zang, Xiaobao Wu, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Haodong Duan, Jiaqi Wang, Yixin Cao, Aixin Sun
| Challenge: | Visualized Document Retrieval (VDR) uses large vision-language models to encode document pages into embeddings. |
| Approach: | They evaluate methods to reduce patch embeddings per page while minimizing performance degradation. |
| Outcome: | The proposed method maintains 98.2% of retrieval performance with only 11.8% of original memory usage and preserves 94.6% effectiveness at 2% memory footprint. |
Attention Score is not All You Need for Token Importance Indicator in KV Cache Reduction: Value Also Matters (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have focused on scaling the context size of large language models (LLMs) however, the enormous inference costs of LLMs limit their applications. |
| Approach: | They propose a method which uses attention scores and the l 1 norm to evaluate token importance. |
| Outcome: | Extensive experiments on LLaMA2-7B-chat and Vicuna-v1.5-7B show that the proposed method outperforms attention-score-only baselines in over 12 tasks. |
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to reduce sequence length rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. |
| Approach: | They propose a method that selectively halts stabilized tokens by monitoring layer-wise update dynamics of the self-attention mechanism. |
| Outcome: | The proposed method can reduce prefill complexity while preserving model accuracy and hardware efficiency. |