Zhen Wan, Chao-Han Huck Yang, Jinchuan Tian, Hanrong Ye, Ankita Pasad, Szu-Wei Fu, Arushi Goel, Ryo Hachiuma, Shizhe Diao, Kunal Dhawan, Sreyan Ghosh, Yusuke Hirota, Zhehuai Chen, Rafael Valle, Chenhui Chu, Shinji Watanabe, Boris Ginsburg, Yu-Chiang Frank Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Effective evaluation of Multimodal Large Language Models (MLLMs) is essential for understanding their capabilities and limitations. In this paper, we introduce MIA-Bench, a benchmark designed to assess MLLMs’ ability to strictly adhere to complex instructions. Our benchmark comprises a diverse set of 400 image-prompt pairs, each crafted to challenge the models’ compliance with layered instructions in generating accurate and contextually appropriate responses. Evaluation results from a wide array of state-of-the-art MLLMs reveal significant variations in performance, highlighting areas for improvement in instruction fidelity. Additionally, we create extra training data and explore supervised fine-tuning and direct preference optimization to enhance the models’ ability to strictly follow instructions without compromising performance on other tasks. We hope this benchmark not only serves as a tool for measuring MLLM adherence to instructions, but also guides future developments in MLLM training methods.

This research aims to comprehensively explore building a multimodal foundation model for egocentric video understanding. To achieve this goal, we work on three fronts. First, as there is a lack of QA data for egocentric video understanding, we automatically generate 7M high-quality QA samples for egocentric videos ranging from 30 seconds to one hour long in Ego4D based on human-annotated data. This is one of the largest egocentric QA datasets. Second, we contribute a challenging egocentric QA benchmark with 629 videos and 7,026 questions to evaluate the models' ability in recognizing and memorizing visual details across videos of varying lengths. We introduce a new de-biasing evaluation method to help mitigate the unavoidable language bias present in the models being evaluated. Third, we propose a specialized multimodal architecture featuring a novel ``Memory Pointer Prompting" mechanism. This design includes a global glimpse step to gain an overarching understanding of the entire video and identify key visual information, followed by a fallback step that utilizes the key visual information to generate responses. This enables the model to more effectively comprehend extended video content. With the data, benchmark, and model, we build MM-Ego, an egocentric multimodal LLM that shows powerful performance on egocentric video understanding.

This paper presents ThinkDiff, a novel alignment paradigm that empowers text-to-image diffusion models with multimodal in-context understanding and reasoning capabilities by integrating the strengths of vision-language models (VLMs). Existing multimodal diffusion finetuning methods largely focus on pixel-level reconstruction rather than in-context reasoning, and are constrained by the complexity and limited availability of reasoning-based datasets. ThinkDiff addresses these challenges by leveraging vision-language training as a proxy task, aligning VLMs with the decoder of an encoder-decoder large language model (LLM) instead of a diffusion decoder. This proxy task builds on the observation that the LLM decoder shares the same input feature space with diffusion decoders that use the corresponding LLM encoder for prompt embedding. As a result, aligning VLMs with diffusion decoders can be simplified through alignment with the LLM decoder. Without complex training and datasets, ThinkDiff effectively unleashes understanding, reasoning, and composing capabilities in diffusion models. Experiments demonstrate that ThinkDiff significantly improves accuracy from 19.2% to 46.3% on the challenging CoBSAT benchmark for multimodal in-context reasoning generation, with only 5 hours of training on 4 A100 GPUs. Additionally, ThinkDiff demonstrates exceptional performance in composing multiple images and texts into logically coherent images. Project page: https://mizhenxing.github.io/ThinkDiff.

In recent years, simultaneous learning of multiple dense prediction tasks with partially annotated label data has emerged as an important research area. Previous works primarily focus on leveraging cross-task relations or conducting adversarial training for extra regularization, which achieve promising performance improvements, while still suffering from the lack of direct pixel-wise supervision and extra training of heavy mapping networks. To effectively tackle this challenge, we propose a novel approach to optimize a set of compact learnable hierarchical task tokens, including global and fine-grained ones, to discover consistent pixel-wise supervision signals in both feature and prediction levels. Specifically, the global task tokens are designed for effective cross-task feature interactions in a global context. Then, a group of fine-grained task-specific spatial tokens for each task is learned from the corresponding global task tokens. It is embedded to have dense interactions with each task-specific feature map. The learned global and local fine-grained task tokens are further used to discover pseudo task-specific dense labels at different levels of granularity, and they can be utilized to directly supervise the learning of the multi-task dense prediction framework. Extensive experimental results on challenging NYUD-v2, Cityscapes, and PASCAL Context datasets demonstrate significant improvements over existing state-of-the-art methods for partially annotated multi-task dense prediction.

We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames). Code and models are available at https://github.com/NVlabs/Long-RL

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints.While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth–width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth–width ratios, with the key finding that although deep–thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy–latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy–latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. This technique can serve as a generalizable component for future SLMs. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy–efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5\% average accuracy, 1.3$\times$/1.9$\times$ lower latency, and 18.7$\times$/45.6$\times$ higher throughput compared to Qwen3-1.7B/0.6B, respectively.

Efficient vision transformer remains a bottleneck for high-resolution images and long-video related real-world applications. Generalized Spatial Propagation Network (GSPN) \cite{wang2025parallel} addresses this by replacing quadratic self-attention with a line-scan propagation scheme, bringing the cost close to linear in the number of rows or columns, while retaining accuracy. Despite this advancement, the existing GSPN implementation still suffers from (i) heavy overhead due to repeatedly launching GPU kernels, (ii) excessive data transfers from global GPU memory, and (iii) redundant computations caused by maintaining separate propagation weights for each channel. We introduce GSPN-2, a joint algorithm–system redesign. In particular, we eliminate thousands of micro-launches from the previous implementation into one single 2D kernel, explicitly pin one warp to each channel slice, and stage the previous column's activations in shared memory. On the model side, we introduce a set of channel-shared propagation weights that replace per-channel matrices, trimming parameters, and align naturally with the affinity map used in transformer attention. Experiments demonstrate GSPN-2's effectiveness across image classification and text-to-image synthesis tasks, matching transformer-level accuracy with significantly lower computational cost. GSPN-2 establishes a new efficiency frontier for modeling global spatial context in vision applications through its unique combination of structured matrix transformations and GPU-optimized implementation.

Recently, there has been an increased interest in the practical problem of learning multiple dense scene understanding tasks from partially annotated data, where each training sample is only labeled for a subset of the tasks. The missing of task labels in training leads to low-quality and noisy predictions, as can be observed from state-of-the-art methods. To tackle this issue, we reformulate the partially-labeled multi-task dense prediction as a pixel-level denoising problem, and propose a novel multi-task denoising diffusion framework coined as DiffusionMTL. It designs a joint diffusion and denoising paradigm to model a potential noisy distribution in the task prediction or feature maps and generate rectified outputs for different tasks. To exploit multi-task consistency in denoising, we further introduce a Multi-Task Conditioning strategy, which can implicitly utilize the complementary nature of the tasks to help learn the unlabeled tasks, leading to an improvement in the denoising performance of the different tasks. Extensive quantitative and qualitative experiments demonstrate that the proposed multi-task denoising diffusion model can significantly improve multi-task prediction maps, and out-perform the state-of-the-art methods on three challenging multi-task benchmarks, under two different partial-labeling evaluation settings. The code is available at https://prismformore.github.io/diffusionmtl/.

Multi-task scene understanding aims to design models that can simultaneously predict several scene understanding tasks with one versatile model. Previous studies typically process multi-task features in a more local way, and thus cannot effectively learn spatially global and cross-task interactions, which hampers the models' ability to fully leverage the consistency of various tasks in multi-task learning. To tackle this problem, we propose an Inverted Pyramid multi-task Transformer, capable of modeling cross-task interaction among spatial features of different tasks in a global context. Specifically, we first utilize a transformer encoder to capture task-generic features for all tasks. And then, we design a transformer decoder to establish spatial and cross-task interaction globally, and a novel UP-Transformer block is devised to increase the resolutions of multi-task features gradually and establish cross-task interaction at different scales. Furthermore, two types of Cross-Scale Self-Attention modules, i.e., Fusion Attention and Selective Attention, are proposed to efficiently facilitate cross-task interaction across different feature scales. An Encoder Feature Aggregation strategy is further introduced to better model multi-scale information in the decoder. Comprehensive experiments on several 2D/3D multi-task benchmarks clearly demonstrate our proposal's effectiveness, establishing significant state-of-the-art performances.

This paper targets the problem of multi-task dense prediction which aims to achieve simultaneous learning and inference on a bunch of multiple dense prediction tasks in a single framework. A core objective in design is how to effectively model cross-task interactions to achieve a comprehensive improvement on different tasks based on their inherent complementarity and consistency. Existing works typically design extra expensive distillation modules to perform explicit interaction computations among different task-specific features in both training and inference, bringing difficulty in adaptation for different task sets, and reducing efficiency due to clearly increased size of multi-task models. In contrast, we introduce feature-wise contrastive consistency into modeling the cross-task interactions for multi-task dense prediction. We propose a novel multi-task contrastive regularization method based on the consistency to effectively boost the representation learning of the different sub-tasks, which can also be easily generalized to different multi-task dense prediction frameworks, and costs no additional computation in the inference. Extensive experiments on two challenging datasets (i.e. NYUD-v2 and Pascal-Context) clearly demonstrate the superiority of the proposed multi-task contrastive learning approach for dense predictions, establishing new state-of-the-art performances.

Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multitask feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available.

RGB-infrared (IR) person re-identification is a challenging task due to the large modality gap between RGB and IR images. Many existing methods bridge the modality gap by style conversion, requiring high-similarity images exchanged by complex CNN structures, like GAN. In this paper, we propose a highly compact modality-aware style adaptation (MSA) framework, which aims to explore more potential relations between RGB and IR modalities by introducing new related modalities. Therefore, the attention is shifted from bridging to filling the modality gap with no requirement on high-quality generated images. To this end, we firstly propose a concise feature-free image generation structure to adapt the original modalities to two new styles that are compatible with both inputs by patch-based pixel redistribution. Secondly, we devise two image style quantification metrics to discriminate styles in image space using luminance and contrast. Thirdly, we design two image-level losses based on the quantified results to guide the style adaptation during an end-to-end four-modality collaborative learning process. Experimental results on two datasets SYSU-MM01 and RegDB show that MSA achieves significant improvements with little extra computation cost and outperforms the state-of-the-art methods.

RGB-Infrared person re-identification (RGB-IR Re-ID) is a cross-modality matching problem, where the modality discrepancy is a big challenge. Most existing works use Euclidean metric based constraints to resolve the discrepancy between features of images from different modalities. However, these methods are incapable of learning angularly discriminative feature embedding because Euclidean distance cannot measure the included angle between embedding vectors effectively. As an angularly discriminative feature space is important for classifying the human images based on their embedding vectors, in this paper, we propose a novel ranking loss function, named Bi-directional Exponential Angular Triplet Loss, to help learn an angularly separable common feature space by explicitly constraining the included angles between embedding vectors. Moreover, to help stabilize and learn the magnitudes of embedding vectors, we adopt a common space batch normalization layer. The quantitative and qualitative experiments on the SYSU-MM01 and RegDB dataset support our analysis. On SYSU-MM01 dataset, the performance is improved from 7.40% / 11.46% to 38.57% / 38.61% for rank-1 accuracy / mAP compared with the baseline. The proposed method can be generalized to the task of single-modality Re-ID and improves the rank-1 accuracy / mAP from 92.0% / 81.7% to 94.7% / 86.6% on the Market-1501 dataset, from 82.6% / 70.6% to 87.6% / 77.1% on the DukeMTMC-reID dataset.

Estimation of the frequency and duration of logos in videos is important and challenging in the advertisement industry as a way of estimating the impact of ad purchases. Since logos occupy only a small area in the videos, the popular methods of image retrieval could fail. This paper develops an algorithm called Video Logo Retrieval (VLR), which is an image-to-video retrieval algorithm based on the spatial distribution of local image descriptors that measure the distance between the query image (the logo) and a collection of video images. VLR uses local features to overcome the weakness of global feature based models such as convolutional neural networks (CNN). Meanwhile, VLR is flexible and does not require training after setting some hyper-parameters. The performance of VLR is evaluated on two challenging open benchmark tasks (SoccerNet and Standford I2V), and compared with other state-of-the-art logo retrieval or detection algorithms. Overall, VLR shows significantly higher accuracy compared with the existing methods.

Rain removal aims to remove the rain streaks on rain images. The state-of-the-art methods are mostly based on Convolutional Neural Network (CNN). However, as CNN is not equivariant to object rotation, these methods are unsuitable for dealing with the tilted rain streaks. To tackle this problem, we propose Deep Symmetry Enhanced Network (DSEN) that is able to explicitly extract the rotation equivariant features from rain images. In addition, we design a self-refining mechanism to remove the accumulated rain streaks in a coarse-to-fine manner. This mechanism reuses DSEN with a novel information link which passes the gradient flow to the higher stages. Extensive experiments on both synthetic and real-world rain images show that our self-refining DSEN yields the top performance.