# Saliency-driven Dynamic Token Pruning for Large Language Models

arXiv: https://arxiv.org/abs/2504.04514 · HTML: https://mlsanity.com/paper/2504.04514/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Visual token pruning in multimodal models, one of 12 directions in Efficient LLM inference and serving.
- Builds on 35 of its 40 references that are in the map.
- First posted on arXiv 6 Apr 2025.

## Placed in

- Efficient LLM inference and serving / [Visual token pruning in multimodal models](https://mlsanity.com/inference/visual-token-pruning-in-multimodal-models/)

## Builds on (most cited first, 30 of 35)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [PIQA: Reasoning about Physical Commonsense in Natural Language](https://mlsanity.com/paper/1911.11641/) (2019)
- [Efficient Streaming Language Models with Attention Sinks](https://mlsanity.com/paper/2309.17453/) (2023)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [BLOOM: A 176B-Parameter Open-Access Multilingual Language Model](https://mlsanity.com/paper/2211.05100/) (2022)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding](https://mlsanity.com/paper/2308.14508/) (2023)
- [SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models](https://mlsanity.com/paper/2211.10438/) (2022)
- [TinyBERT: Distilling BERT for Natural Language Understanding](https://mlsanity.com/paper/1909.10351/) (2019)
- [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot](https://mlsanity.com/paper/2301.00774/) (2023)
- [DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification](https://mlsanity.com/paper/2106.02034/) (2021)
- [H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models](https://mlsanity.com/paper/2306.14048/) (2023)
- [Are Sixteen Heads Really Better than One?](https://mlsanity.com/paper/1905.10650/) (2019)
- [How Much Knowledge Can You Pack Into the Parameters of a Language Model?](https://mlsanity.com/paper/2002.08910/) (2020)
- [MathQA: Towards Interpretable Math Word Problem Solving with Operation-Based Formalisms](https://mlsanity.com/paper/1905.13319/) (2019)
- [LLM-Pruner: On the Structural Pruning of Large Language Models](https://mlsanity.com/paper/2305.11627/) (2023)
- [Reducing Transformer Depth on Demand with Structured Dropout](https://mlsanity.com/paper/1909.11556/) (2019)
- [Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time](https://mlsanity.com/paper/2305.17118/) (2023)
- [WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations](https://mlsanity.com/paper/1808.09121/) (2018)
- [LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression](https://mlsanity.com/paper/2403.12968/) (2024)
- [The (Un)reliability of saliency methods](https://mlsanity.com/paper/1711.00867/) (2017)
- [LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models](https://mlsanity.com/paper/2310.05736/) (2023)
- [Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time](https://mlsanity.com/paper/2310.17157/) (2023)
- [SpAtten: Efficient Sparse Attention Architecture with Cascade Token and Head Pruning](https://mlsanity.com/paper/2012.09852/) (2020)
- [Learned Token Pruning for Transformers](https://mlsanity.com/paper/2107.00910/) (2021)
- [Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Search](https://mlsanity.com/paper/2010.07003/) (2020)

## Built on by (most cited first, 5 of 5)

- [A Survey of Context Engineering for Large Language Models](https://mlsanity.com/paper/2507.13334/) (2025)
- [Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning](https://mlsanity.com/paper/2506.02867/) (2025)
- [Token Reduction Should Go Beyond Efficiency in Generative Models -- From Vision, Language to Multimodality](https://mlsanity.com/paper/2505.18227/) (2025)
- [HEART-VIT: Hessian-Guided Efficient Dynamic Attention and Token Pruning in Vision Transformer](https://mlsanity.com/paper/2512.20120/) (2025)
- [HasMem: Hard-Origin Adaptively Softened Memory for Long-Term LLM Agents](https://mlsanity.com/paper/2609.30797/) (2026)

## Abstract (arXiv)

Despite the recent success of large language models (LLMs), LLMs are particularly challenging in long-sequence inference scenarios due to the quadratic computational complexity of the attention mechanism. Inspired by the interpretability theory of feature attribution in neural network models, we observe that not all tokens have the same contribution. Based on this observation, we propose a novel token pruning framework, namely Saliency-driven Dynamic Token Pruning (SDTP), to gradually and dynamically prune redundant tokens based on the input context. Specifically, a lightweight saliency-driven prediction module is designed to estimate the importance score of each token with its hidden state, which is added to different layers of the LLM to hierarchically prune redundant tokens. Furthermore, a ranking-based optimization strategy is proposed to minimize the ranking divergence of the saliency score and the predicted importance score. Extensive experiments have shown that our framework is generalizable to various models and datasets. By hierarchically pruning 65% of the input tokens, our method greatly reduces 33% ~ 47% FLOPs and achieves speedup up to 1.75× during inference, while maintaining comparable performance. We further demonstrate that SDTP can be combined with KV cache compression method for further compression.