Jul 2026· IEEE Transactions on Pattern Analysis and Machine Intelligence· Vol PP, pp. 1-13· 0 citations
Medicine
TL;DR
This paper introduces TinyMem, a model built upon compact multimodal memory for long-form video action detection that outperforms a range of state-of-the-art models on AVA v2.2 while using 5 times fewer memory tokens than the baseline with dense visual memory embeddings.
Abstract
Despite the great advances in video understanding with deep neural networks, current solutions still struggle with input videos that last for minutes, if not hours. To mitigate this issue, existing approaches typically build memory caches of dense visual embeddings for video transformers to model long-range spatio-temporal dependencies. However, even with hundreds of extended memory tokens, their performance remains limited. In this paper, we argue that compact yet informative memory representations can effectively improve performance. To this end, we introduce TinyMem, a model built upon compact multimodal memory for long-form video action detection. In particular, we condense redundant video content into succinct descriptions to extract abstract textual semantics. Subsequently, we integrate region-condensed visual embeddings with text embeddings. TinyMem outperforms a range of state-of-the-art models on AVA v2.2, Epic-Kitchens-100 and Breakfast with highly condensed memory, e.g., 37.4 mAP with TinyMem-24-12 on AVA v2.2 while using 5 times fewer memory tokens than the baseline with dense visual memory embeddings.
The success of large language models (LLMs) has inspired the development of foundation-level multimodal systems that integrate vision and language. However, current video-language models—such as Video-LLaMA and VideoChat—struggle with fine-grained human motion understanding and fail to summarize long videos effectively. Meanwhile, motion-focused models are limited to short clips and lack mechanisms to capture long-range spatiotemporal context. We introduce ActionLMM, a memory-augmented vision-language model for long-video action summarization. It aligns visual and motion modalities through joint representation learning and leverages a novel dual-memory mechanism to retain both local motion details and global temporal structure. To support evaluation, we propose a large-scale benchmark dataset with 33,887 longform action videos and 169,435 caption annotations across 1920 action categories. Experiments show that ActionLMM significantly outperforms prior methods, offering a robust and scalable solution for fine-grained human action understanding.
Ruirui Li, Dari Abdullah Alrwoaily, Turgut Sofuyev et al.· International Conference on...· 0 citations
Large vision-language models (LVLMs) have shown significant progress in video understanding, but they struggle to scale to long videos due to limited context windows. Existing methods reduce input dimensionality via frame sampling and feature compression, yet discard details and incur high computational cost for post-training. In contrast, retrieval-augmented generation (RAG) that indexes long videos for query retrieval and memory-based methods that maintain evolving long-term stores, offer a lighter and deployment-friendly solution. Nevertheless, they rely on shallow retrieval that selects only top-ranked segments and fails to integrate information across multiple relevant video episodes. Inspired by Multiple-Trace Theory in cognitive psychology, we revisit long video understanding from a probe-echo perspective, in which human episodic memories are activated and integrated in parallel. Building on this insight, we propose ProEchoMem, a cognitive-inspired framework that simulates the probe-echo mechanism: (1) Incremental Episodic Memory Construction builds structured knowledge graphs from video streams; (2) Probe-Driven Memory Activation generates probe signals from user queries to activate all stored traces simultaneously; (3) Memory Echo Synthesis integrates activated traces into a coherent and structured memory echo. Experiments on LongerVideos, LVBench, and cross-domain settings demonstrate the effectiveness of ProEchoMem, with multi-trace probing achieving up to 14.2% higher relevance and ablation studies validating the contribution of each module. The code is available at https://github.com/Applied-Machine-Learning-Lab/SIGIR26_ProEchoMem
Derong Xu, Yanxin Chen, Wanyu Wang et al.· Annual International ACM SIG...· 0 citations
Accurate video forecasting enables autonomous vehicles to anticipate hazards, robotics and surveillance systems to predict human intent, and environmental models to issue timely warnings for extreme weather events. However, existing methods remain limited: transformers rely on global attention with quadratic complexity, making them impractical for high-resolution, long-horizon video prediction, while convolutional and recurrent networks suffer from short-range receptive fields and vanishing gradients, losing key information over extended sequences. To overcome these challenges, we introduce VideoTitans , the first architecture to adapt the gradient-driven Titans memory—originally designed for language modelling to video prediction. VideoTitans integrates three core ideas: (i) a sliding-window attention core that scales linearly with sequence length and spatial resolution, (ii) an episodic memory that dynamically retains only informative tokens based on a gradient-based surprise signal, and (iii) a small set of persistent tokens encoding task-specific priors that stabilize training and enhance generalization. Extensive experiments on Moving-MNIST, Human3.6M, TrafficBJ and WeatherBench benchmarks show that VideoTitans consistently reduces computation (FLOPs) and achieves competitive visual fidelity compared to state-of-the-art recurrent, convolutional, and efficient-transformer methods. Comprehensive ablations confirm that each proposed component contributes significantly.
Young-Jae Park, Minseok Seo, Hae-Gon Jeon· Neural Information Processin...· 3 citations
Streaming video understanding requires multimodal large language models (MLLMs) to preserve relevant evidence from continuously evolving streams under strict causality and bounded memory. Yet existing paradigms remain limited: model-based methods require intrusive backbone updates, while memory-based methods expend substantial visual-encoding computation on temporally redundant content and rely on rigid access to visual history. To address these limitations, we introduce StreamFlow, an efficient visual memory framework that enables dynamic, on-demand access to historical visual information. StreamFlow combines a lightweight, dynamics-aware mid-term memory that filters temporal redundancy before visual encoding with a latent long-term memory that consolidates historical video content into visual latents accessible to subsequent reasoning. During generation, an attention-guided retrieval mechanism injects relevant visual latents when the model's reliance on visual evidence weakens. StreamFlow achieves state-of-the-art streaming video understanding performance, reaching 67.73% overall accuracy on StreamingBench, while also delivering strong performance on offline long-video benchmarks. Relative to the vanilla setting, it improves the visual attention score (VAS) by 59.1% while reducing end-to-end latency and peak memory by 50.4% and 21.1%, respectively, enabling more visually grounded and efficient reasoning.
Muxin Fu, Yifan Zhang, Wentao Zhang et al.· 0 citations
Large Multimodal Models (LMMs) for video reasoning have long been hindered by the high computational cost of processing vast amounts of visual information. This dilemma motivates the transfer of the reasoning capabilities of large models to smaller, more efficient ones. On-Policy Distillation (OPD) offers a promising solution by matching output-token distributions along student-generated trajectories. However, video reasoning often depends on evidence accumulated across multiple frames. In this context, output-level supervision only captures information expressed through token predictions and does not directly constrain the latent representations formed during reasoning. To address this limitation, we propose Latent-OPD, which augments OPD with trajectory-level latent distillation. Specifically, our method focuses on the position at the end of each trajectory, where hidden states effectively summarize the accumulated visual evidence and reasoning context. Furthermore, we introduce a progressive teacher-lookahead strategy, which aligns middle-to-late student layers with increasingly deeper teacher layers. Experiments on six video reasoning benchmarks show that Latent-OPD consistently outperforms output-only OPD. Notably, the improvements are particularly pronounced in scenarios with limited frames, long videos, or tasks requiring complex evidence aggregation. These results establish Latent-OPD as a highly effective approach to frame-efficient video reasoning.
Aoni Shen, Yongheng Zhang, Yinghui Li et al.· 1 citation
Though Multimodal Large Language Models (MLLMs) have shown impressive potential in video understanding, long video understanding (LVU) remains challenging since distracting noise in complex and lengthy contexts can obscure localized details, misleading MLLMs to produce incorrect answers. Recent works mitigate these issues by incentivizing deep reasoning to include relevant evidence. However, these methods have two main problems: First, the reinforcement fine-tuning framework (RFT) they leveraged incurs substantial training overheads, including high annotation costs and complicated reward designs. Second, the self-reflective and iterative-perception mechanism in some methods causes lengthy outputs and high inference latency. To alleviate these problems, we propose a novel Segment-to-Video Supervision} method (S2V) to efficiently enhance fine-grained reasoning in LVU. Specifically, we generate question answer pairs (VQA) based on localized segments, and then transfer these segment-based VQA back to the whole video for training. Due to focusing on short segments, segment-based VQA can naturally notice details which tend to be overlooked from a whole-video perspective. Training on such data can enforce MLLMs to correctly associate fine-grained details with QA while avoiding distracting noise in the whole video. The S2V training involves just reinforcement learning (RL) with a simple accuracy reward based on only 10K VQA samples and the resulting S2V model predicts answer using a single forward pass with limited output tokens. Experimental results demonstrate that S2V can consistently improve LVU performance across multiple LVU benchmarks, outperforming both general MLLMs and reasoning-based methods not only in LVU accuracy but also in training and inference efficiency.
Beibei Zhang, Chao Xu, Jun Lan et al.· 0 citations