Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem?

Multimodal large language models (MLLMs) have shown remarkable performance for cross-modal understanding and generation, yet still suffer from severe inference costs. Recently, abundant works have been proposed to solve this problem with token pruning, which identifies the redundant tokens in MLLMs and then prunes them to reduce the computation and KV storage costs, leading to significant acceleration without training. While these methods claim efficiency gains, critical questions about their fundamental design and evaluation remain unanswered: Why do many existing approaches underperform even compared to naive random token selection? Are attention-based scoring sufficient for reliably identifying redundant tokens? Is language information really helpful during token pruning? What makes a good trade-off between token importance and duplication? Are current evaluation protocols comprehensive and unbiased? The ignorance of previous research on these problems hinders the long-term development of token pruning. In this paper, we answer these questions one by one, providing insights into the design of future token pruning methods.

Stop Looking for Important Tokens in Multimodal Language Models: Duplication Matters More

Vision tokens in multimodal large language models often dominate huge computational overhead due to their excessive length compared to linguistic modality. Abundant recent methods aim to solve this problem with token pruning, which first defines an importance criterion for tokens and then prunes the unimportant vision tokens during inference. However, in this paper, we show that the importance is not an ideal indicator to decide whether a token should be pruned. Surprisingly, it usually results in inferior performance than random token pruning and leading to incompatibility to efficient attention computation operators.Instead, we propose DART (Duplication-Aware Reduction of Tokens), which prunes tokens based on its duplication with other tokens, leading to significant and training-free acceleration. Concretely, DART selects a small subset of pivot tokens and then retains the tokens with low duplication to the pivots, ensuring minimal information loss during token pruning. Experiments demonstrate that DART can prune 88.9% vision tokens while maintaining comparable performance, leading to a 1.99$\times$ and 2.99$\times$ speed-up in total time and prefilling stage, respectively, with good compatibility to efficient attention operators. Our codes are available at https://github.com/ZichenWen1/DART.

RealRAG: Retrieval-augmented Realistic Image Generation via Self-reflective Contrastive Learning

Recent text-to-image generative models, e.g., Stable Diffusion V3 and Flux, have achieved notable progress. However, these models are strongly restricted to their limited knowledge, a.k.a., their own fixed parameters, that are trained with closed datasets. This leads to significant hallucinations or distortions when facing fine-grained and unseen novel real-world objects, e.g., the appearance of the Tesla Cybertruck. To this end, we present the first real-object-based retrieval-augmented generation framework (RealRAG), which augments fine-grained and unseen novel object generation by learning and retrieving real-world images to overcome the knowledge gaps of generative models. Specifically, to integrate missing memory for unseen novel object generation, we train a reflective retriever by self-reflective contrastive learning, which injects the generator's knowledge into the sef-reflective negatives, ensuring that the retrieved augmented images compensate for the model's missing knowledge. Furthermore, the real-object-based framework integrates fine-grained visual knowledge for the generative models, tackling the distortion problem and improving the realism for fine-grained object generation. Our Real-RAG is superior in its modular application to all types of state-of-the-art text-to-image generative models and also delivers remarkable performance boosts with all of them, such as a gain of 16.18% FID score with the auto-regressive model on the Stanford Car benchmark.

Compression with Global Guidance: Towards Training-free High-Resolution MLLMs Acceleration

Multimodal large language models (MLLMs) have attracted considerable attention due to their exceptional performance in visual content understanding and reasoning. However, their inference efficiency has been a notable concern, as the increasing length of multimodal contexts leads to quadratic complexity. Token compression techniques, which reduce the number of visual tokens, have demonstrated their effectiveness in reducing computational costs. Yet, these approaches have struggled to keep pace with the rapid advancements in MLLMs, especially the AnyRes strategy in the context of high-resolution image understanding. In this paper, we propose a novel token compression method, GlobalCom$^2$, tailored for high-resolution MLLMs that receive both the thumbnail and multiple crops. GlobalCom$^2$ treats the tokens derived from the thumbnail as the ``commander'' of the entire token compression process, directing the allocation of retention ratios and the specific compression for each crop. In this way, redundant tokens are eliminated while important local details are adaptively preserved to the highest extent feasible. Empirical results across 10 benchmarks reveal that GlobalCom$^2$ achieves an optimal balance between performance and efficiency, and consistently outperforms state-of-the-art token compression methods with LLaVA-NeXT-7B/13B models. Our code is released at \url{https://github.com/xuyang-liu16/GlobalCom2}.

Token Pruning for Caching Better: 9 Times Acceleration on Stable Diffusion for Free

Stable Diffusion has achieved remarkable success in the field of text-to-image generation, with its powerful generative capabilities and diverse generation results making a lasting impact. However, its iterative denoising introduces high computational costs and slows generation speed, limiting broader adoption. The community has made numerous efforts to reduce this computational burden, with methods like feature caching attracting attention due to their effectiveness and simplicity. Nonetheless, simply reusing features computed at previous timesteps causes the features across adjacent timesteps to become similar, reducing the dynamics of features over time and ultimately compromising the quality of generated images. In this paper, we introduce a dynamics-aware token pruning (DaTo) approach that addresses the limitations of feature caching. DaTo selectively prunes tokens with lower dynamics, allowing only high-dynamic tokens to participate in self-attention layers, thereby extending feature dynamics across timesteps. DaTo combines feature caching with token pruning in a training-free manner, achieving both temporal and token-wise information reuse. Applied to Stable Diffusion on the ImageNet, our approach delivered a 9$\times$ speedup while reducing FID by 0.33, indicating enhanced image quality. On the COCO-30k, we observed a 7$\times$ acceleration coupled with a notable FID reduction of 2.17.

Accelerating Diffusion Transformers with Dual Feature Caching

Diffusion Transformers (DiT) have become the dominant methods in image and video generation yet still suffer substantial computational costs. As an effective approach for DiT acceleration, feature caching methods are designed to cache the features of DiT in previous timesteps and reuse them in the next timesteps, allowing us to skip the computation in the next timesteps. However, on the one hand, aggressively reusing all the features cached in previous timesteps leads to a severe drop in generation quality. On the other hand, conservatively caching only the features in the redundant layers or tokens but still computing the important ones successfully preserves the generation quality but results in reductions in acceleration ratios. Observing such a tradeoff between generation quality and acceleration performance, this paper begins by quantitatively studying the accumulated error from cached features. Surprisingly, we find that aggressive caching does not introduce significantly more caching errors in the caching step, and the conservative feature caching can fix the error introduced by aggressive caching. Thereby, we propose a dual caching strategy that adopts aggressive and conservative caching iteratively, leading to significant acceleration and high generation quality at the same time. Besides, we further introduce a V-caching strategy for token-wise conservative caching, which is compatible with flash attention and requires no training and calibration data. Our codes have been released in Github: \textbf{Code: \href{https://github.com/Shenyi-Z/DuCa}{\texttt{\textcolor{cyan}{https://github.com/Shenyi-Z/DuCa}}}}

Intelligent System for Automated Molecular Patent Infringement Assessment

Automated drug discovery offers significant potential for accelerating the development of novel therapeutics by substituting labor-intensive human workflows with machine-driven processes. However, a critical bottleneck persists in the inability of current automated frameworks to assess whether newly designed molecules infringe upon existing patents, posing significant legal and financial risks. We introduce PatentFinder, a novel tool-enhanced and multi-agent framework that accurately and comprehensively evaluates small molecules for patent infringement. It incorporates both heuristic and model-based tools tailored for decomposed subtasks, featuring: MarkushParser, which is capable of optical chemical structure recognition of molecular and Markush structures, and MarkushMatcher, which enhances large language models' ability to extract substituent groups from molecules accurately. On our benchmark dataset MolPatent-240, PatentFinder outperforms baseline approaches that rely solely on large language models, demonstrating a 13.8\% increase in F1-score and a 12\% rise in accuracy. Experimental results demonstrate that PatentFinder mitigates label bias to produce balanced predictions and autonomously generates detailed, interpretable patent infringement reports. This work not only addresses a pivotal challenge in automated drug discovery but also demonstrates the potential of decomposing complex scientific tasks into manageable subtasks for specialized, tool-augmented agents.

Learning Robust Anymodal Segmentor with Unimodal and Cross-modal Distillation

Simultaneously using multimodal inputs from multiple sensors to train segmentors is intuitively advantageous but practically challenging. A key challenge is unimodal bias, where multimodal segmentors over rely on certain modalities, causing performance drops when others are missing, common in real world applications. To this end, we develop the first framework for learning robust segmentor that can handle any combinations of visual modalities. Specifically, we first introduce a parallel multimodal learning strategy for learning a strong teacher. The cross-modal and unimodal distillation is then achieved in the multi scale representation space by transferring the feature level knowledge from multimodal to anymodal segmentors, aiming at addressing the unimodal bias and avoiding over-reliance on specific modalities. Moreover, a prediction level modality agnostic semantic distillation is proposed to achieve semantic knowledge transferring for segmentation. Extensive experiments on both synthetic and real-world multi-sensor benchmarks demonstrate that our method achieves superior performance.

SAVEn-Vid: Synergistic Audio-Visual Integration for Enhanced Understanding in Long Video Context

Endeavors have been made to explore Large Language Models for video analysis (Video-LLMs), particularly in understanding and interpreting long videos. However, existing Video-LLMs still face challenges in effectively integrating the rich and diverse audio-visual information inherent in long videos, which is crucial for comprehensive understanding. This raises the question: how can we leverage embedded audio-visual information to enhance long video understanding? Therefore, (i) we introduce SAVEn-Vid, the first-ever long audio-visual video dataset comprising over 58k audio-visual instructions. (ii) From the model perspective, we propose a time-aware Audio-Visual Large Language Model (AV-LLM), SAVEnVideo, fine-tuned on SAVEn-Vid. (iii) Besides, we present AVBench, a benchmark containing 2,500 QAs designed to evaluate models on enhanced audio-visual comprehension tasks within long video, challenging their ability to handle intricate audio-visual interactions. Experiments on AVBench reveal the limitations of current AV-LLMs. Experiments also demonstrate that SAVEnVideo outperforms the best Video-LLM by 3.61% on the zero-shot long video task (Video-MME) and surpasses the leading audio-visual LLM by 1.29% on the zero-shot audio-visual task (Music-AVQA). Consequently, at the 7B parameter scale, SAVEnVideo can achieve state-of-the-art performance. Our dataset and code will be released at https://ljungang.github.io/SAVEn-Vid/ upon acceptance.

ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models

Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model's focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our code will be public.