While GUI agents have shown strong performance under explicit and completion instructions, real-world deployment requires aligning with users' more complex implicit intents.In this work, we highlight Hierarchical Implicit Intent Alignment for Personalized GUI Agent (PersonalAlign), a new agent task that requires agents to leverage long-term user records as persistent context to resolve omitted preferences in vague instructions and anticipate latent routines by user state for proactive assistance.To facilitate this study, we introduce AndroidIntent, a benchmark designed to evaluate agents' ability in resolving vague instructions and providing proactive suggestions through reasoning over long-term user records.We annotated 775 user-specific preferences and 215 routines from 20k long-term records across different users for evaluation.Furthermore, we introduce Hierarchical Intent Memory Agent (HIM-Agent), which maintains a continuously updating personal memory and hierarchically organizes user preferences and routines for personalization.Finally, we evaluate a range of GUI agents on AndroidIntent, including GPT-5, Qwen3-VL, and UI-TARS, further results show that HIM-Agent significantly improves both execution and proactive performance by 15.7% and 7.3%.

Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, Liqiang Nie. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Dataset distillation (DD) excels in synthesizing a small number of images per class (IPC) but struggles to maintain its effectiveness in high-IPC settings. Recent works on dataset distillation demonstrate that combining distilled and real data can mitigate the effectiveness decay. However, our analysis of the combination paradigm reveals that the current one-shot and independent selection mechanism induces an incompatibility issue between distilled and real images. To address this issue, we introduce a novel curriculum coarse-to-fine selection (CCFS) method for efficient high-IPC dataset distillation. CCFS employs a curriculum selection framework for real data selection, where we leverage a coarse-to-fine strategy to select appropriate real data based on the current synthetic dataset in each curriculum. Extensive experiments validate CCFS, surpassing the state-of-the-art by +6.6% on CIFAR-10, +5.8% on CIFAR-100, and +3.4% on Tiny-ImageNet under high-IPC settings. Notably, CCFS achieves 60.2% test accuracy on ResNet-18 with a 20% compression ratio of Tiny-ImageNet, closely matching full-dataset training with only 0.3% degradation. Code: https://github.com/CYDaaa30/CCFS.

Building an agent that can mimic human behavior patterns to accomplish various open-world tasks is a longterm goal. To enable agents to effectively learn behavioral patterns across diverse tasks, a key challenge lies in modeling the intricate relationships among observations, actions, and language. To this end, we propose Optimus-2, a novel Minecraft agent that incorporates a Multimodal Large Language Model (MLLM) for highlevel planning, alongside a Goal-Observation-Action Conditioned Policy (GOAP) for low-level control. GOAP contains (1) an Action-guided Behavior Encoder that models causal relationships between observations and actions at each timestep, then dynamically interacts with the historical observation-action sequence, consolidating it into fixedlength behavior tokens, and (2) an MLLM that aligns behavior tokens with open-ended language instructions to predict actions auto-regressively. Moreover, we introduce a high-quality Minecraft Goal-Observation-Action (MGOA) dataset, which contains 25,000 videos across 8 atomic tasks, providing about 30M goal-observation-action pairs. The automated construction method, along with the MGOA dataset, can contribute to the community’s efforts to train Minecraft agents. Extensive experimental results demonstrate that Optimus-2 exhibits superior performance across atomic tasks, long-horizon tasks, and open-ended instruction tasks in Minecraft. Please see the project page at https://cybertronagent.github.io/Optimus-2.github.io/.

The research focus of GUI agents is shifting from text-dependent to pure-vision-based approaches, which, though promising, prioritize comprehensive pre-training data collection while neglecting contextual modeling challenges. We probe the characteristics of element and history contextual modeling in GUI agent and summarize: 1) the high-density and loose-relation of element context highlight the existence of many unrelated elements and their negative influence; 2) the high redundancy of history context reveals the inefficient history modeling in current GUI agents. In this work, we propose a context-aware simplification framework for building an efficient and effective GUI Agent, termed SimpAgent. To mitigate potential interference from numerous unrelated elements, we introduce a masking-based element pruning method that circumvents the intractable relation modeling through an efficient masking mechanism. To reduce the redundancy in historical information, we devise a consistency-guided history compression module, which enhances implicit LLM-based compression through innovative explicit guidance, achieving an optimal balance between performance and efficiency. With the above components, SimpAgent reduces 27% FLOPs and achieves superior GUI navigation performances. Comprehensive navigation experiments across diverse web and mobile environments demonstrate the effectiveness and potential of our agent.

The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens.

Smartphone agents are increasingly important for helping users control devices efficiently, with (Multimodal) Large Language Model (MLLM)-based approaches emerging as key contenders. Fairly comparing these agents is essential but challenging, requiring a varied task scope, the integration of agents with different implementations, and a generalisable evaluation pipeline to assess their strengths and weaknesses. In this paper, we present SPA-Bench, a comprehensive SmartPhone Agent Benchmark designed to evaluate (M)LLM-based agents in an interactive environment that simulates real-world conditions. SPA-Bench offers three key contributions: (1) A diverse set of tasks covering system and third-party apps in both English and Chinese, focusing on features commonly used in daily routines; (2) A plug-and-play framework enabling real-time agent interaction with Android devices, integrating over ten agents with the flexibility to add more; (3) A novel evaluation pipeline that automatically assesses agent performance across multiple dimensions, encompassing seven metrics related to task completion and resource consumption. Our extensive experiments across tasks and agents reveal challenges like interpreting mobile user interfaces, action grounding, memory retention, and execution costs. We propose future research directions to ease these difficulties, moving closer to real-world smartphone agent applications.

Dataset distillation aims to encapsulate the rich information contained in dataset into a compact distilled dataset but it faces performance degradation as the image-per-class (IPC) setting or image resolution grows larger. Recent advancements demonstrate that integrating diffusion generative models can effectively facilitate the compression of large-scale datasets while maintaining efficiency due to their superiority in matching data distribution and summarizing representative patterns. However, images sampled from diffusion models are always blamed for lack of diversity which may lead to information redundancy when multiple independent sampled images are aggregated as a distilled dataset. To address this issue, we propose Adversary-guided Curriculum Sampling (ACS), which partitions the distilled dataset into multiple curricula. For generating each curriculum, ACS guides diffusion sampling process by an adversarial loss to challenge a discriminator trained on sampled images, thus mitigating information overlap between curricula and fostering a more diverse distilled dataset. Additionally, as the discriminator evolves with the progression of curricula, ACS generates images from simpler to more complex, ensuring efficient and systematic coverage of target data informational spectrum. Extensive experiments demonstrate the effectiveness of ACS, which achieves substantial improvements of 4.1% on Imagewoof and 2.1% on ImageNet-1k over the state-of-the-art.

Benefiting from the availability of extensive navigation trajectories, both manually and automatically annotated, current graphical user interface (GUI) agents have achieved remarkable advancements in performance. However, these annotated datasets often contain substantial noise, which impedes effective agent training and underscores the necessity for rigorous trajectory quality assessment. In contrast to existing prompting-based evaluators that rely on proprietary multimodal large language models (MLLMs), we propose an Uncertainty-aware Reinforced Self-Training (URST) framework to train lightweight MLLMs for efficient and reliable trajectory evaluation. URST iteratively fine-tunes MLLMs using their own generated thoughts and judgments to enable self-improvement, while its uncertainty-aware sampling strategy ensures the selection of the most informative training examples. To further enhance reasoning and judgment capabilities, we propose a simplified group policy optimization approach that effectively leverages diverse positive and negative samples for evaluator learning. Our evaluator demonstrates superior judgment performance across both in-domain and out-of-domain datasets. When used to filter navigation datasets, it consistently leads to performance improvements in training GUI agents.

Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multimodal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vIsual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between imagelevel and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imper-fect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multimodal benchmarks demonstrate the superiority of our model (e.g., improvement of 5% accuracy on VSR and 3% CIDEr on TextCaps over InstructBLIP, 5% accuracy on RefCOCOg over Kosmos-2).

Building a general-purpose agent is a long-standing vision in the field of artificial intelligence. Existing agents have made remarkable progress in many domains, yet they still struggle to complete long-horizon tasks in an open world. We attribute this to the lack of necessary world knowledge and multimodal experience that can guide agents through a variety of long-horizon tasks. In this paper, we propose a Hybrid Multimodal Memory module to address the above challenges. It 1) transforms knowledge into Hierarchical Directed Knowledge Graph that allows agents to explicitly represent and learn world knowledge, and 2) summarises historical information into Abstracted Multimodal Experience Pool that provide agents with rich references for in-context learning. On top of the Hybrid Multimodal Memory module, a multimodal agent, Optimus-1, is constructed with dedicated Knowledge-guided Planner and Experience-Driven Reflector, contributing to a better planning and reflection in the face of long-horizon tasks in Minecraft. Extensive experimental results show that Optimus-1 significantly outperforms all existing agents on challenging long-horizon task benchmarks, and exhibits near human-level performance on many tasks. In addition, we introduce various Multimodal Large Language Models (MLLMs) as the backbone of Optimus-1. Experimental results show that Optimus-1 exhibits strong generalization with the help of the Hybrid Multimodal Memory module, outperforming the GPT-4V baseline on many tasks.

While the conditional sequence modeling with the transformer architecture has demonstrated its effectiveness in dealing with offline reinforcement learning (RL) tasks, it is struggle to handle out-of-distribution states and actions. Existing work attempts to address this issue by data augmentation with the learned policy or adding extra constraints with the value-based RL algorithm. However, these studies still fail to overcome the following challenges: (1) insufficiently utilizing the historical temporal information among inter-steps, (2) overlooking the local intra-step relationships among return-to-gos (RTGs), states, and actions, (3) overfitting suboptimal trajectories with noisy labels. To address these challenges, we propose $\textbf{D}$ecision $\textbf{M}$amba ($\textbf{DM}$), a novel multi-grained state space model (SSM) with a self-evolving policy learning strategy. DM explicitly models the historical hidden state to extract the temporal information by using the mamba architecture. To capture the relationship among RTG-state-action triplets, a fine-grained SSM module is designed and integrated into the original coarse-grained SSM in mamba, resulting in a novel mamba architecture tailored for offline RL. Finally, to mitigate the overfitting issue on noisy trajectories, a self-evolving policy is proposed by using progressive regularization. The policy evolves by using its own past knowledge to refine the suboptimal actions, thus enhancing its robustness on noisy demonstrations. Extensive experiments on various tasks show that DM outperforms other baselines substantially.

Multimodal large language models (MLLMs) have demonstrated impressive capabilities across various vision-language tasks. However, a generalist MLLM typically underperforms compared with a specialist MLLM on most VL tasks, which can be attributed to task interference. In this paper, we propose a mixture of multimodal experts (MoME) to mitigate task interference and obtain a generalist MLLM. Our MoME is composed of two key components, a mixture of vision experts (MoVE) and a mixture of language experts (MoLE). MoVE can adaptively modulate the features transformed from various vision encoders, and has a strong compatibility in transformation architecture. MoLE incorporates sparsely gated experts into LLMs to achieve painless improvements with roughly unchanged inference costs. In response to task interference, our MoME specializes in both vision and language modality to adapt to task discrepancies. Extensive experiments show that MoME significantly improves the performance of generalist MLLMs across various VL tasks.

The goal of few-shot image recognition is to classify different categories with only one or a few training samples. Previous works of few-shot learning mainly focus on simple images, such as object or character images. Those works usually use a convolutional neural network (CNN) to learn the global image representations from training tasks, which are then adapted to novel tasks. However, there are many more abstract and complex images in real world, such as scene images, consisting of many object entities with flexible spatial relations among them. In such cases, global features can hardly obtain satisfactory generalization ability due to the large diversity of object relations in the scenes, which may hinder the adaptability to novel scenes. This paper proposes a composite object relation modeling method for few-shot scene recognition, capturing the spatial structural characteristic of scene images to enhance adaptability on novel scenes, considering that objects commonly co- occurred in different scenes. In different few-shot scene recognition tasks, the objects in the same images usually play different roles. Thus we propose a task-aware region selection module (TRSM) to further select the detected regions in different few-shot tasks. In addition to detecting object regions, we mainly focus on exploiting the relations between objects, which are more consistent to the scenes and can be used to cleave apart different scenes. Objects and relations are used to construct a graph in each image, which is then modeled with graph convolutional neural network. The graph modeling is jointly optimized with few-shot recognition, where the loss of few-shot learning is also capable of adjusting graph based representations. Typically, the proposed graph based representations can be plugged in different types of few-shot architectures, such as metric-based and meta-learning methods. Experimental results of few-shot scene recognition show the effectiveness of the proposed method.

Scene images are usually composed of foreground and background regional contents. Some existing methods propose to extract regional contents with dense grids or objectness region proposals. However, dense grids may split the object into several discrete parts, learning semantic ambiguity for the patches. The objectness methods may focus on particular objects but only pay attention to the foreground contents and do not exploit the background that is key to scene recognition. In contrast, we propose a novel scene recognition framework with amorphous region detection and context modeling. In the proposed framework, discriminative regions are first detected with amorphous contours that can tightly surround the targets through semantic segmentation techniques. In addition, both foreground and background regions are jointly embedded to obtain the scene representations with the graph model. Based on the graph modeling module, we explore the contextual relations between the regions in geometric and morphology aspects, and generate the discriminative representations for scene recognition. Experimental results on MIT67 and SUN397 demonstrate the effectiveness and generality of the proposed method.

As the data size in Machine Learning fields grows exponentially, it is inevitable to accelerate the computation by utilizing the ever-growing large number of available cores provided by high-performance computing hardware. However, existing parallel methods for clustering or regression often suffer from problems of low accuracy, slow convergence, and complex hyperparameter-tuning. Furthermore, the parallel efficiency is usually difficult to improve while striking a balance between preserving model properties and partitioning computing workloads on distributed systems. In this paper, we propose a novel and simple data structure capturing the most important information among data samples. It has several advantageous properties supporting a hierarchical clustering strategy that contains well-defined metrics for determining optimal hierarchy, balanced partition for maintaining the clustering property, and efficient parallelization for accelerating computation phases. Then we combine the clustering with regression techniques as a parallel library and utilize a hybrid structure of data and model parallelism to make predictions. Experiments illustrate that our library obtains remarkable performance on convergence, accuracy, and scalability.

Scene classification is a valuable classification subtask and has its own characteristics which still needs more in-depth studies. Basically, scene characteristics are distributed over the whole image, which cause the need of “seeing” comprehensive and informative regions. Previous works mainly focus on region discovery and aggregation, while rarely involves the inherent properties of CNN along with its potential ability to satisfy the requirements of scene classification. In this paper, we propose to understand scene images and the scene classification CNN models in terms of the focus area. From this new perspective, we find that large focus area is preferred in scene classification CNN models as a consequence of learning scene characteristics. Meanwhile, the analysis about existing training schemes helps us to understand the effects of focus area, and also raises the question about optimal training method for scene classification. Pursuing the better usage of scene characteristics, we propose a new learning scheme with a tailored loss in the goal of activating larger focus area on scene images. Since the supervision of the target regions to be enlarged is usually lacked, our alternative learning scheme is to erase already activated area, and allow the CNN models to activate more area during training. The proposed scheme is implemented by keeping the pairwise consistency between the output of the erased image and its original one. In particular, a tailored loss is proposed to keep such pairwise consistency by leveraging category-relevance information. Experiments on Places365 show the significant improvements of our method with various CNNs. Our method shows an inferior result on the object-centric dataset, ImageNet, which experimentally indicates that it captures the unique characteristics of scenes.

Exploiting the spatial structure in scene images is a key research direction for scene recognition. Due to the large intra-class structural diversity, building and modeling flexible structural layout to adapt various image characteristics is a challenge. Existing structural modeling methods in scene recognition either focus on predefined grids or rely on learned prototypes, which all have limited representative ability. In this paper, we propose Prototype-agnostic Scene Layout (PaSL) construction method to build the spatial structure for each image without conforming to any prototype. Our PaSL can flexibly capture the diverse spatial characteristic of scene images and have considerable generalization capability. Given a PaSL, we build Layout Graph Network (LGN) where regions in PaSL are defined as nodes and two kinds of independent relations between regions are encoded as edges. The LGN aims to incorporate two topological structures (formed in spatial and semantic similarity dimensions) into image representations through graph convolution. Extensive experiments show that our approach achieves state-of-the-art results on widely recognized MIT67 and SUN397 datasets without multi-model or multi-scale fusion. Moreover, we also conduct the experiments on one of the largest scale datasets, Places365. The results demonstrate the proposed method can be well generalized and obtains competitive performance.

Scene recognition is challenging due to the intra-class diversity and inter-class similarity. Previous works recognize scenes either with global representations or with the intermediate representations of objects. In contrast, we investigate more discriminative image representations of object-to-object relations for scene recognition, which are based on the triplets of obtained with detection techniques. Particularly, two types of representations, including co-occurring frequency of object-to-object relation (denoted as COOR) and sequential representation of object-to-object relation (denoted as SOOR), are proposed to describe objects and their relative relations in different forms. COOR is represented as the intermediate representation of co-occurring frequency of objects and their relations, with a three order tensor that can be fed to scene classifier without further embedding. SOOR is represented in a more explicit and freer form that sequentially describe image contents with local captions. And a sequence encoding model (e.g., recurrent neural network (RNN)) is implemented to encode SOOR to the features for feeding the classifiers. In order to better capture the spatial information, the proposed COOR and SOOR are adapted to RGB-D data, where a RGB-D proposal fusion method is proposed for RGB-D object detection. With the proposed approaches COOR and SOOR, we obtain the state-of-the-art results of RGB-D scene recognition on SUN RGB-D and NYUD2 datasets.

Scene recognition has been a challenging task in the field of computer vision and multimedia for a long time. The current scene recognition works often extract object features and scene features through CNN, and combine these two types of features to obtain complementary and discriminative scene representations. However, when the scene categories are visually similar, the object features might lack of discriminations. Therefore, it may be debatable to consider only object features. In contrast to the existing works, in this paper, we discuss the discrimination of scene attributes in local regions and utilize scene attributes as the complementary features of object and scene features. We extract these visual features from two individual CNN branches, one extracting the global features of the image while the other extracting the features of local regions. Through contextual modeling framework, we aggregate these features and generate more discriminative scene representations, which achieve better performance than the feature aggregation of object and scene. Moreover, we achieve the new state-of-the-art performance on both standard scene recognition benchmarks by aggregating more complementary visual features: MIT67 (88.06%) and SUN397 (74.12%).

Due to the abstraction of scenes, comprehensive scene understanding requires semantic modeling in both global and local aspects. Scene recognition is usually researched from a global point of view, while dense captioning is typically studied for local regions. Previous works separately research on the modeling of scene recognition and dense captioning. In contrast, we propose a joint learning framework that benefits from the mutual coupling of scene recognition and dense captioning models. Generally, these two tasks are coupled through two steps, 1) fusing the supervision by considering the contexts between scene labels and local captions, and 2) jointly optimizing semantically symmetric LSTM models. Particularly, in order to balance bias between dense captioning and scene recognition, a scene adaptive non-maximum suppression (NMS) method is proposed to emphasize the scene related regions in region proposal procedure, and a region-wise and category-wise weighted pooling method is proposed to avoid over attention on particular regions in local to global pooling procedure. For the model training and evaluation, scene labels are manually annotated for Visual Genome database. The experimental results on Visual Genome show the effectiveness of the proposed method. Moreover, the proposed method also can improve previous CNN based works on public scene databases, such as MIT67 and SUN397.

Scene classification is a challenging problem. Compared with object images, scene images are more abstract, as they are composed of objects. Object and scene images have different characteristics with different scales and composition structures. How to effectively integrate the local mid-level semantic representations including both object and scene concepts needs to be investigated, which is an important aspect for scene classification. In this article, the idea of a sharing codebook is introduced by organically integrating deep learning, concept feature, and local feature encoding techniques. More specifically, the shared local feature codebook is generated from the combined ImageNet1K and Places365 concepts (Mixed1365) using convolutional neural networks. As the Mixed1365 features cover all the semantic information including both object and scene concepts, we can extract a shared codebook from the Mixed1365 features, which only contain a subset of the whole 1,365 concepts with the same codebook size. The shared codebook can not only provide complementary representations without additional codebook training but also be adaptively extracted toward different scene classification tasks. A method of fusing the encoded features with both the original codebook and the shared codebook is proposed for scene classification. In this way, more comprehensive and representative image features can be generated for classification. Extensive experimentations conducted on two public datasets validate the effectiveness of the proposed method. Besides, some useful observations are also revealed to show the advantage of shared codebook.