Modern multimodal large language models (MLLMs) can reason over hour-long video, yet their key-value (KV) cache grows linearly with time--quickly exceeding the fixed memory of phones, AR glasses, and edge robots. Prior compression schemes either assume the whole video and user query are available offline or must first build the full cache, so memory still scales with stream length. InfiniPot-V is the first training-free, query-agnostic framework that enforces a hard, lengthindependent memory cap for streaming video understanding. During video encoding it monitors the cache and, once a user-set threshold is reached, runs a lightweight compression pass that (i) removes temporally redundant tokens via Temporal-axis Redundancy (TaR) metric and (ii) keeps semantically significant tokens via Value-Norm (VaN) ranking. Across four open-source MLLMs and four long-video and streaming-video benchmarks, InfiniPot-V cuts peak GPU memory by up to 94%, sustains real-time generation, and matches or surpasses full-cache accuracy--even in multi-turn dialogues. By dissolving the KV cache bottleneck without retraining or query knowledge, InfiniPot-V closes the gap for on-device streaming video assistants.

Additionally, we introduce memory bank sampling, a frame sampling strategy designed to extract informative frames from historical images, further improving effectiveness and efficiency. Experimental results demonstrate that our pretraining method and architectural components substantially enhance model performance. In both simulated and real-world experiments, our model achieves a significant increase in success rate over OpenVLA [1]. To further assess spatial perception and generalization to novel views, we introduce MV-Bench, a multi-view simulation benchmark. Our model consistently outperforms existing methods, demonstrating stronger spatial understanding and adaptability.

This study identifies that visual-related functional decoding is distributed across different decoder layers in Multimodal Large Language Models (MLLMs). Typically, each function, such as counting, grounding, or OCR recognition, narrows down to two or three layers, which we define as Vision Function Layers (VFL). Additionally, the depth and its order of different VFLs exhibits a consistent pattern across different MLLMs, which is well-aligned with human behaviors (e.g., recognition occurs first, followed by counting, and then grounding). These findings are derived from Visual Token Swapping, our novel analytical framework that modifies targeted KV cache entries to precisely elucidate layer-specific functions during decoding. Furthermore, these insights offer substantial utility in tailoring MLLMs for real-world downstream applications. For instance, when LoRA training is selectively applied to VFLs whose functions align with the training data, VFLLoRA not only outperform full-LoRA but also prevent out-of-domain function forgetting. Moreover, by analyzing the performance differential on training data when particular VFLs are ablated, VFL-select automatically classifies data by function, enabling highly efficient data selection to directly bolster corresponding capabilities. Consequently, VFL-select surpasses human experts in data selection, and achieves 98% of full-data performance with only 20% of the original dataset. This study delivers deeper comprehension of MLLM visual processing, fostering the creation of more efficient, interpretable, and robust models.

A well-known dilemma in large vision-language models (e.g., GPT-4, LLaVA) is that while increasing the number of vision tokens generally enhances visual understanding, it also significantly raises memory and computational costs, especially in long-term, dense video frame streaming scenarios. Although learnable approaches like Q-Former and Perceiver Resampler have been developed to reduce the vision token burden, they overlook the context causally modeled by LLMs (i.e., key-value cache), potentially leading to missed visual cues when addressing user queries. In this paper, we introduce a novel approach to reduce vision compute by leveraging redundant vision tokens ``skipping layers'' rather than decreasing the number of vision tokens. Our method, VideoLLM-MoD, is inspired by mixture-of-depths LLMs and addresses the challenge of numerous vision tokens in long-term or streaming video. Specifically, for certain transformer layer, we learn to skip the computation for a high proportion (e.g., 80\%) of vision tokens, passing them directly to the next layer. This approach significantly enhances model efficiency, achieving approximately 42% time and 30% memory savings for the entire training. Moreover, our method reduces the computation in the context and avoid decreasing the vision tokens, thus preserving or even improving performance compared to the vanilla model. We conduct extensive experiments to demonstrate the effectiveness of VideoLLM-MoD, showing its state-of-the-art results on multiple benchmarks, including narration, forecasting, and summarization tasks in COIN, Ego4D, and Ego-Exo4D datasets.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in multimodal tasks. Despite their impressive performance, MLLMs suffer from the modality imbalance issue, where visual information is often underutilized compared to textual representations in deeper layers, leading to degraded visual performance or hallucinations. This issue stems from the predominant reliance on next-text-token-prediction during training, which fails to provide direct visual supervisory signals, resulting in progressive homogenization of visual representations throughout the layers. To this end, we propose Latent Visual Reconstruction (LaVer), a novel training framework that facilitates MLLMs in learning more discriminative visual representations via masked image modeling in the joint latent semantic space of LLM. Our method offers direct visual activation to MLLMs, which exhibit increased visual attention allocation, indicating enhanced utilization of visual information. Extensive experiments across diverse benchmarks prove the superiority of our approach in various scenarios, especially those requiring dense visual capabilities. Code of LaVer is available at https://github.com/Fir-lat/LaVer.

We introduce TimeViper, a hybrid vision-language model designed to tackle challenges of long video understanding. Processing long videos demands both an efficient model architecture and an effective mechanism for handling extended temporal contexts. To this end, TimeViper adopts a hybrid Mamba-Transformer backbone that combines the efficiency of state-space models with the expressivity of attention mechanisms. Through this hybrid design, we reveal the vision-to-text information aggregation phenomenon, where information progressively flows from vision tokens to text tokens across increasing LLM depth, resulting in severe vision token redundancy. Motivated by this observation, we propose TransV, a token information transfer module that transfers and compresses vision tokens into instruction tokens while maintaining multimodal understanding capabilities. This design enables TimeViper to process hour-long videos exceeding 10,000 frames. Extensive experiments across multiple benchmarks demonstrate that TimeViper competes with state-of-the-art models while extending frame numbers. We further analyze attention behaviors of both Mamba and Transformer layers, offering new insights into hybrid model interpretability. This work represents an initial step towards developing, interpreting, and compressing hybrid Mamba-Transformer architectures.

In Vision Language Models (VLMs), vision tokens are quantity-heavy yet information-dispersed compared with language tokens, thus consume too much unnecessary computation. Pruning redundant vision tokens for high VLM inference efficiency has been continuously studied but all existing methods resort to indirect and non-guaranteed ways. W e propose OC-VTP, a direct and guaranteed approach to select the most representative vision tokens for high-efficiency yet accuracy-preserving VLM inference. Our OC-VTP requires merely light-weight pre-training of a small object-centric vision token pruner, which can then be inserted into existing VLMs, without fine-tuning of any models on any datasets. It is gauranteed that the most representative vision tokens are kept by minimizing the error in reconstructing the original unpruned tokens from the selected ones. Across any vision pruning ratios, i.e., inference efficiency, our OC-VTP consistently helps mainstream VLMs to preserve the highest inference accuracy. Our pruning also demonstrates interesting interpretability.

Multimodal large language models (MLLMs) have achieved impressive performance, but high-resolution visual inputs result in long sequences of visual tokens and substantial inference latency. Reducing redundant visual tokens is critical to ease computational/memory burdens while preserving performance, enabling MLLM deployment in resource-constrained or latency-sensitive scenarios. Current visual token pruning methods mainly rely on attention-based redundancy analysis and are tailored to dense architectures. We propose Fast Multimodal Mixture-of-Experts (FastMMoE), a training-free acceleration framework for mixture-of-experts (MoE) based MLLMs, developed from a routing analysis perspective. FastMMoE combines two complementary strategies: (i) expert activation reduction for visual tokens to minimize unnecessary expert computation; and (ii) routing-aware token pruning that leverages similarity in routing probability distributions to identify and remove highly redundant visual tokens. Experiments on large-scale MoE-MLLMs such as DeepSeek-VL2 and InternVL3.5 demonstrate that FastMMoE can reduce FLOPs by up to 55.0% while retaining approximately 95.5% of the original performance, consistently outperforming dense-model pruning baselines including FastV and SparseVLM across multiple retention rates.

Recently, augmenting vision-language-action models (VLAs) with world-models has shown promise in robotic policy learning. However, it remains challenging to jointly predict next-state observations and action sequences because of the inherent difference between the two modalities. To address this, we propose DUal-STream diffusion (DUST), a world-model augmented VLA framework that handles the modality conflict and enhances the performance of VLAs across diverse tasks. Specifically, we propose a multimodal diffusion transformer architecture that explicitly maintains separate modality streams while enabling cross-modal knowledge sharing. In addition, we propose training techniques such as independent noise perturbations for each modality and a decoupled flow matching loss, which enables the model to learn the joint distribution in a bidirectional manner while avoiding the need for a unified latent space. Furthermore, based on the decoupled training framework, we introduce a sampling method where we sample action and vision tokens asynchronously at different rates, which shows improvement through inference-time scaling. Through experiments on simulated benchmarks such as RoboCasa and GR-1, DUST achieves up to 6% gains over a standard VLA baseline and implicit world-modeling methods, with our inference-time scaling approach providing an additional 2-5% gain on success rate. On real-world tasks with the Franka Research 3, DUST outperforms baselines in success rate by 13%, confirming its effectiveness beyond simulation. Lastly, we demonstrate the effectiveness of DUST in large-scale pretraining with action-free videos from BridgeV2, where DUST leads to significant gain when transferred to the RoboCasa benchmark.