Diffusion transformers have shown exceptional performance in visual generation but incur high computational costs. Token reduction techniques that compress models by sharing the denoising process among similar tokens have been introduced. However, existing approaches neglect the denoising priors of the diffusion models, leading to suboptimal acceleration and diminished image quality. This study proposes a novel concept: attend to prune feature redundancies in areas not attended by the diffusion process. We analyze the location and degree of feature redundancies based on the structure-then-detail denoising priors. Subsequently, we introduce SDTM, a structure-then-detail token merging approach that dynamically compresses feature redundancies. Specifically, we design dynamic visual token merging, compression ratio adjusting, and prompt reweighting for different stages. Served in a post-training way, the proposed method can be integrated seamlessly into any DiT architecture. Extensive experiments across various backbones, schedulers, and datasets showcase the superiority of our method, for example, it achieves 1.55 times acceleration with negligible impact on image quality. Project page: https://github.com/ICTMCG/SDTM.
Compute optimization using token reduction of LLM prompts is an emerging task in the fields of NLP and next generation, agentic AI. In this white paper, we introduce a novel (patent pending) text representation scheme and a first-of-its-kind word-level semantic compression of paragraphs that can lead to over 90% token reduction, while retaining high semantic similarity to the source text. We explain how this novel compression technique can be lossless and how the detail granularity is controllable. We discuss benchmark results over open source data (i.e. Bram Stoker's Dracula available through Project Gutenberg) and show how our results hold at the paragraph level, across multiple genres and models.
We present DyMU, an efficient, training-free framework that dynamically reduces the computational burden of vision-language models (VLMs) while maintaining high task performance. Our approach comprises two key components. First, Dynamic Token Merging (DToMe) reduces the number of visual token embeddings by merging similar tokens based on image complexity, addressing the inherent inefficiency of fixed-length outputs in vision transformers. Second, Virtual Token Unmerging (VTU) simulates the expected token sequence for large language models (LLMs) by efficiently reconstructing the attention dynamics of a full sequence, thus preserving the downstream performance without additional fine-tuning. Unlike previous approaches, our method dynamically adapts token compression to the content of the image and operates completely training-free, making it readily applicable to most state-of-the-art VLM architectures. Extensive experiments on image and video understanding tasks demonstrate that DyMU can reduce the average visual token count by 32%-85% while achieving comparable performance to full-length models across diverse VLM architectures, including the recently popularized AnyRes-based visual encoders. Furthermore, through qualitative analyses, we demonstrate that DToMe effectively adapts token reduction based on image complexity and, unlike existing systems, provides users more control over computational costs. Project page: https://mikewangwzhl.github.io/dymu/.
Large Language Models (LLMs) have shown outstanding performance across a variety of tasks, partly due to advanced prompting techniques. However, these techniques often require lengthy prompts, which increase computational costs and can hinder performance because of the limited context windows of LLMs. While prompt compression is a straightforward solution, existing methods confront the challenges of retaining essential information, adapting to context changes, and remaining effective across different tasks. To tackle these issues, we propose a task-agnostic method called Dynamic Compressing Prompts (LLM-DCP). Our method reduces the number of prompt tokens while aiming to preserve the performance as much as possible. We model prompt compression as a Markov Decision Process (MDP), enabling the DCP-Agent to sequentially remove redundant tokens by adapting to dynamic contexts and retaining crucial content. We develop a reward function for training the DCP-Agent that balances the compression rate, the quality of the LLM output, and the retention of key information. This allows for prompt token reduction without needing an external black-box LLM. Inspired by the progressive difficulty adjustment in curriculum learning, we introduce a Hierarchical Prompt Compression (HPC) training strategy that gradually increases the compression difficulty, enabling the DCP-Agent to learn an effective compression method that maintains information integrity. Experiments demonstrate that our method outperforms state-of-the-art techniques, especially at higher compression rates. The code for our approach will be available at https://github.com/Fhujinwu/DCP.
To effectively reduce the visual tokens in Visual Large Language Models (VLLMs), we propose a novel approach called Window Token Concatenation (WiCo). Specifically, we employ a sliding window to concatenate spatially adjacent visual tokens. However, directly concatenating these tokens may group diverse tokens into one, and thus obscure some fine details. To address this challenge, we propose fine-tuning the last few layers of the vision encoder to adaptively adjust the visual tokens, encouraging that those within the same window exhibit similar features. To further enhance the performance on fine-grained visual understanding tasks, we introduce WiCo+, which decomposes the visual tokens in later layers of the LLM. Such a design enjoys the merits of the large perception field of the LLM for fine-grained visual understanding while keeping a small number of visual tokens for efficient inference. We perform extensive experiments on both coarse- and fine-grained visual understanding tasks based on LLaVA-1.5 and Shikra, showing better performance compared with existing token reduction projectors. The code is available: https://github.com/JackYFL/WiCo.
Grounded Conversation Generation (GCG) is an emerging vision-language task that requires models to generate natural language responses seamlessly intertwined with corresponding object segmentation masks. Recent models, such as GLaMM and OMG-LLaVA, achieve pixel-level grounding but incur significant computational costs due to processing a large number of visual tokens. Existing token pruning methods, like FastV and PyramidDrop, fail to preserve the local visual features critical for accurate grounding, leading to substantial performance drops in GCG tasks. To address this, we propose Adaptive Local-Aware Token Pruning (ALTP), a simple yet effective framework that accelerates GCG models by prioritizing local object information. ALTP introduces two key components: (1) Detail Density Capture (DDC), which uses superpixel segmentation to retain tokens in object-centric regions, preserving fine-grained details, and (2) Dynamic Density Formation (DDF), which dynamically allocates tokens based on information density, ensuring higher retention in semantically rich areas. Extensive experiments on the GranDf dataset demonstrate that ALTP significantly outperforms existing token pruning methods, such as FastV and PyramidDrop, on both GLaMM and OMG-LLaVA models. Notably, when applied to GLaMM, ALTP achieves a 90% reduction in visual tokens with a 4.9% improvement in AP50 and a 5.0% improvement in Recall compared to PyramidDrop. Similarly, on OMG-LLaVA, ALTP improves AP by 2.1% and mIOU by 3.0% at a 90% token reduction compared with PDrop.
Token-based video representation has emerged as a promising approach for enabling LLMs to interpret video content. However, existing token reduction, such as token pruning and token merging, often disrupt essential spatial-temporal positional embeddings, failing to adequately balance computational efficiency with fewer tokens. Consequently, these methods result in lengthy token sequences, limiting their applicability in scenarios requiring extreme token compression, such as video large language models. In this paper, we introduce the novel task of extreme short token reduction, aiming to represent extensive video sequences with a minimal number of tokens. To address this challenge, we propose Token Dynamics, a new video representation framework that dynamically reduces token count while preserving spatial-temporal coherence. Specifically, we disentangle video representations by separating visual embeddings from grid-level motion information, structuring them into: 1. a concise token hash table, created by clustering tokens that describe object-level content; 2. a token indices key map, capturing detailed spatial-temporal motion patterns across grids; 3. a token hash function, which vector-quantizes the token hash table to reconstruct the token sequence from the key map. Furthermore, we introduce a cross-dynamics attention mechanism that integrates motion features into the token base without increasing token length, thereby maintaining compactness and spatial-temporal integrity. The experiments demonstrate a reduction of token count to merely 0.07% of the original tokens, with only a minor performance drop of 1.13%. Additionally, we propose two novel subtasks within extreme token reduction (fixed-length and adaptive-length compression). Our method offers significantly lower theoretical complexity, fewer tokens, and enhanced throughput, thus providing an efficient solution for video LLMs.
Parameter-efficient tuning (PET) aims to transfer pre-trained foundation models to downstream tasks by learning a small number of parameters. Compared to traditional fine-tuning, which updates the entire model, PET significantly reduces storage and transfer costs for each task regardless of exponentially increasing pre-trained model capacity. However, most PET methods inherit the inference latency of their large backbone models and often introduce additional computational overhead due to additional modules (e.g. adapters), limiting their practicality for compute-intensive applications. In this paper, we propose Faster Parameter-Efficient Tuning (FPET), a novel approach that enhances inference speed and training efficiency while maintaining high storage efficiency. Specifically, we introduce a plug-and-play token redundancy reduction module delicately designed for PET. This module refines tokens from the self-attention layer using an adapter to learn the accurate similarity between tokens and cuts off the tokens through a fully-differentiable token merging strategy, which uses a straight-through estimator for optimal token reduction. Experimental results prove that our FPET achieves faster inference and higher memory efficiency than the pre-trained backbone while keeping competitive performance on par with state-of-the-art PET methods.
Hallucination has been a long-standing and inevitable problem that hinders the application of Large Vision-Language Models (LVLMs) in domains that require high reliability. Various methods focus on improvement depending on data annotations or training strategies, yet place less emphasis on LLM's inherent problems. To fill this gap, we delve into the attention mechanism of the decoding process in the LVLM. Intriguingly, our investigation uncovers the prevalent attention redundancy within the hierarchical architecture of the LVLM, manifesting as overextended image processing in deep layers and an overabundance of non-essential image tokens. Stemming from the observation, we thus propose MINT, a novel training-free decoding strategy, MItigating hallucinations via tokeN reducTion. Specifically, we dynamically intensify the LVLM's local perception capability by masking its attention to irrelevant image tokens. In addition, we use contrastive decoding that pushes the model to focus more on those key image regions. Our full method aims to guide the model in concentrating more on key visual elements during generation. Extensive experimental results on several popular public benchmarks show that our approach achieves a 4% improvement in mitigating hallucinations caused by distracted perception compared to original models. Meanwhile, our approach is demonstrated to make the model perceive 5% more visual points even though we reduce a suite of image tokens.
State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames (640$\times$360) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least 50% reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by 4.3% on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.
Image tokenization has significantly advanced visual generation and multimodal modeling, particularly when paired with autoregressive models. However, current methods face challenges in balancing efficiency and fidelity: high-resolution image reconstruction either requires an excessive number of tokens or compromises critical details through token reduction. To resolve this, we propose Latent Consistency Tokenizer (Layton) that bridges discrete visual tokens with the compact latent space of pre-trained Latent Diffusion Models (LDMs), enabling efficient representation of 1024x1024 images using only 256 tokens-a 16 times compression over VQGAN. Layton integrates a transformer encoder, a quantized codebook, and a latent consistency decoder. Direct application of LDM as the decoder results in color and brightness discrepancies. Thus, we convert it to latent consistency decoder, reducing multi-step sampling to 1-2 steps for direct pixel-level supervision. Experiments demonstrate Layton's superiority in high-fidelity reconstruction, with 10.8 reconstruction Frechet Inception Distance on MSCOCO-2017 5K benchmark for 1024x1024 image reconstruction. We also extend Layton to a text-to-image generation model, LaytonGen, working in autoregression. It achieves 0.73 score on GenEval benchmark, surpassing current state-of-the-art methods. Project homepage: https://github.com/OPPO-Mente-Lab/Layton
Efficient vision-language understanding of large Remote Sensing Images (RSIs) is meaningful but challenging. Current Large Vision-Language Models (LVLMs) typically employ limited pre-defined grids to process images, leading to information loss when handling gigapixel RSIs. Conversely, using unlimited grids significantly increases computational costs. To preserve image details while reducing computational complexity, we propose a text-guided token pruning method with Dynamic Image Pyramid (DIP) integration. Our method introduces: (i) a Region Focus Module (RFM) that leverages text-aware region localization capability to identify critical vision tokens, and (ii) a coarse-to-fine image tile selection and vision token pruning strategy based on DIP, which is guided by RFM outputs and avoids directly processing the entire large imagery. Additionally, existing benchmarks for evaluating LVLMs' perception ability on large RSI suffer from limited question diversity and constrained image sizes. We construct a new benchmark named LRS-VQA, which contains 7,333 QA pairs across 8 categories, with image length up to 27,328 pixels. Our method outperforms existing high-resolution strategies on four datasets using the same data. Moreover, compared to existing token reduction methods, our approach demonstrates higher efficiency under high-resolution settings. Dataset and code are in https://github.com/VisionXLab/LRS-VQA.
Vision-Language Models (VLMs) are powerful yet computationally intensive for widespread practical deployments. To address such challenge without costly re-training, post-training acceleration techniques like quantization and token reduction are extensively explored. However, current acceleration evaluations primarily target minimal overall performance degradation, overlooking a crucial question: does the accelerated model still give the same answers to the same questions as it did before acceleration? This is vital for stability-centered industrial applications where consistently correct answers for specific, known situations are paramount, such as in AI-based disease diagnosis. We systematically investigate this for accelerated VLMs, testing four leading models (LLaVA-1.5, LLaVA-Next, Qwen2-VL, Qwen2.5-VL) with eight acceleration methods on ten multi-modal benchmarks. Our findings are stark: despite minimal aggregate performance drops, accelerated models changed original answers up to 20% of the time. Critically, up to 6.5% of these changes converted correct answers to incorrect. Input perturbations magnified these inconsistencies, and the trend is confirmed by case studies with the medical VLM LLaVA-Med. This research reveals a significant oversight in VLM acceleration, stressing an urgent need for instance-level stability checks to ensure trustworthy real-world deployment.
Recent advancements in Large Language Model (LLM)-based Natural Language Generation evaluation have largely focused on single-example prompting, resulting in significant token overhead and computational inefficiencies. In this work, we introduce BatchGEMBA-MQM, a framework that integrates batched prompting with the GEMBA-MQM metric for machine translation evaluation. Our approach aggregates multiple translation examples into a single prompt, reducing token usage by 2-4 times (depending on the batch size) relative to single-example prompting. Furthermore, we propose a batching-aware prompt compression model that achieves an additional token reduction of 13-15% on average while also showing ability to help mitigate batching-induced quality degradation. Evaluations across several LLMs (GPT-4o, GPT-4o-mini, Mistral Small, Phi4, and CommandR7B) and varying batch sizes reveal that while batching generally negatively affects quality (but sometimes not substantially), prompt compression does not degrade further, and in some cases, recovers quality loss. For instance, GPT-4o retains over 90% of its baseline performance at a batch size of 4 when compression is applied, compared to a 44.6% drop without compression. We plan to release our code and trained models at https://github.com/NL2G/batchgemba to support future research in this domain.
Large language models (LLMs) process entire input contexts indiscriminately, which is inefficient in cases where the information required to answer a query is localized within the context. We present dynamic context cutoff, a human-inspired method enabling LLMs to self-terminate processing upon acquiring sufficient task-relevant information. Through analysis of model internals, we discover that specific attention heads inherently encode "sufficiency signals" - detectable through lightweight classifiers - that predict when critical information has been processed. This reveals a new efficiency paradigm: models' internal understanding naturally dictates processing needs rather than external compression heuristics. Comprehensive experiments across six QA datasets (up to 40K tokens) with three model families (LLaMA/Qwen/Mistral, 1B0-70B) demonstrate 1.33x average token reduction while improving accuracy by 1.3%. Furthermore, our method demonstrates better performance with the same rate of token reduction compared to other context efficiency methods. Additionally, we observe an emergent scaling phenomenon: while smaller models require require probing for sufficiency detection, larger models exhibit intrinsic self-assessment capabilities through prompting.
Vision-Language Models (VLMs) have achieved remarkable success across a range of multimodal tasks; however, their practical deployment is often constrained by high computational costs and prolonged inference times. Since the vision modality typically carries more information than the text modality, compressing visual prompts offers a promising solution to alleviate these challenges. Existing approaches predominantly focus on refining model architectures or directly reducing the number of visual tokens. However, these methods often compromise inference performance due to a lack of consideration for the unique spatial and temporal characteristics of visual data. In this work, we propose a token compression paradigm that operates on both spatial and temporal dimensions. Our approach includes a learning-free, plug-and-play compression pipeline that can be seamlessly integrated into most Multimodal Large Language Model (MLLM) frameworks. By leveraging this method, we enhance the model inference capability while simultaneously reducing its computational cost. Experimental results on the Video-QA task demonstrate the effectiveness of the proposed approach, showcasing significant improvements in efficiency without sacrificing performance.
Vision-Language Models (VLMs) have achieved notable success in multimodal tasks but face practical limitations due to the quadratic complexity of decoder attention mechanisms and autoregressive generation. Existing methods like FASTV and VTW have achieved notable results in reducing redundant visual tokens, but these approaches focus on pruning tokens in a single forward pass without systematically analyzing the redundancy of visual tokens throughout the entire generation process. In this paper, we introduce a dynamic pruning strategy tailored for VLMs, namedDynamic Rate (DyRate), which progressively adjusts the compression rate during generation. Our analysis of the distribution of attention reveals that the importance of visual tokens decreases throughout the generation process, inspiring us to adopt a more aggressive compression rate. By integrating a lightweight predictor based on attention distribution, our approach enables flexible adjustment of pruning rates based on the attention distribution. Our experimental results demonstrate that our method not only reduces computational demands but also maintains the quality of responses.
The success of VLMs often relies on the dynamic high-resolution schema that adaptively augments the input images to multiple crops, so that the details of the images can be retained. However, such approaches result in a large number of redundant visual tokens, thus significantly reducing the efficiency of the VLMs. To improve the VLMs' efficiency without introducing extra training costs, many research works are proposed to reduce the visual tokens by filtering the uninformative visual tokens or aggregating their information. Some approaches propose to reduce the visual tokens according to the self-attention of VLMs, which are biased, to result in inaccurate responses. The token reduction approaches solely rely on visual cues are text-agnostic, and fail to focus on the areas that are most relevant to the question, especially when the queried objects are non-salient to the image. In this work, we first conduct experiments to show that the original text embeddings are aligned with the visual tokens, without bias on the tailed visual tokens. We then propose a self-adaptive cross-modality attention mixture mechanism that dynamically leverages the effectiveness of visual saliency and text-to-image similarity in the pre-LLM layers to select the visual tokens that are informative. Extensive experiments demonstrate that the proposed approach achieves state-of-the-art training-free VLM acceleration performance, especially when the reduction rate is sufficiently large.
The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily focus on importance-based token pruning, which overlooks the redundancy caused by frame resemblance and repetitive visual elements. In this paper, we analyze the high vision token similarities in LVLMs. We reveal that token similarity distribution condenses as layers deepen while maintaining ranking consistency. Leveraging the unique properties of similarity over importance, we introduce FrameFusion, a novel approach that combines similarity-based merging with importance-based pruning for better token reduction in LVLMs. FrameFusion identifies and merges similar tokens before pruning, opening up a new perspective for token reduction. We evaluate FrameFusion on diverse LVLMs, including Llava-Video-{7B,32B,72B}, and MiniCPM-V-8B, on video understanding, question-answering, and retrieval benchmarks. Experiments show that FrameFusion reduces vision tokens by 70$\%$, achieving 3.4-4.4x LLM speedups and 1.6-1.9x end-to-end speedups, with an average performance impact of less than 3$\%$. Our code is available at https://github.com/thu-nics/FrameFusion.
Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their abilities to generate and understand cross-modal data. However, processing long sequences of visual tokens extracted from visual backbones poses a challenge for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating both computational and memory demands during training and inference. Through a comprehensive analysis of the token reduction process, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We showcase the effectiveness of FOLDER by integrating it into the visual backbone of several MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. In both contexts, FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to 70% of visual tokens.