Linzhuang Sun, Hao Liang, Jingxuan Wei, Bihui Yu, Tianpeng Li, Fan Yang, Zenan Zhou, Wentao Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Language Models (LLMs) have exhibited significant potential in performing diverse tasks, including the ability to call functions or use external tools to enhance their performance. While current research on function calling by LLMs primarily focuses on single-turn interactions, this paper addresses the overlooked necessity for LLMs to engage in multi-turn function calling—critical for handling compositional, real-world queries that require planning with functions but not only use functions. To facilitate this, we introduce an approach, BUTTON, which generates synthetic compositional instruction tuning data via bottom-up instruction construction and top-down trajectory generation. In the bottom-up phase, we generate simple atomic tasks based on real-world scenarios and build compositional tasks using heuristic strategies based on atomic tasks. Corresponding function definitions are then synthesized for these compositional tasks. The top-down phase features a multi-agent environment where interactions among simulated humans, assistants, and tools are utilized to gather multi-turn function calling trajectories. This approach ensures task compositionality and allows for effective function and trajectory generation by examining atomic tasks within compositional tasks. We produce a dataset BUTTONInstruct comprising 8k data points and demonstrate its effectiveness through extensive experiments across various LLMs.

With the rapid progress of Multimodal LLMs, evaluating their mathematical reasoning capabilities has become an increasingly important research direction. In particular, visual-textual mathematical reasoning serves as a key indicator of an MLLM's ability to comprehend and solve complex, multi-step quantitative problems. While existing benchmarks such as MathVista and MathVerse have advanced the evaluation of multimodal math proficiency, they primarily rely on digitally rendered content and fall short in capturing the complexity of real-world scenarios. To bridge this gap, we introduce MathScape, a novel benchmark focused on assessing MLLMs' reasoning ability in realistic mathematical contexts. MathScape comprises 1,369 high-quality math problems paired with human-captured real-world images, closely reflecting the challenges encountered in practical educational settings. We conduct a thorough multi-dimensional evaluation across nine leading closed-source MLLMs, three open-source MLLMs with over 20 billion parameters, and seven smaller-scale MLLMs. Our results show that even SOTA models struggle with real-world math tasks, lagging behind human performance-highlighting critical limitations in current model capabilities. Moreover, we find that strong performance on synthetic or digitally rendered images does not guarantee similar effectiveness on real-world tasks. This underscores the necessity of MathScape in the next stage of multimodal mathematical reasoning.

Large Language Models (LLMs) have shown remarkable capabilities in reasoning, exemplified by the success of OpenAI-o1 and DeepSeek-R1. However, integrating reasoning with external search processes remains challenging, especially for complex multi-hop questions requiring multiple retrieval steps. We propose ReSearch, a novel framework that trains LLMs to Reason with Search via reinforcement learning without using any supervised data on reasoning steps. Our approach treats search operations as integral components of the reasoning chain, where when and how to perform searches is guided by text-based thinking, and search results subsequently influence further reasoning. We train ReSearch on Qwen2.5-7B(-Instruct) and Qwen2.5-32B(-Instruct) models and conduct extensive experiments. Despite being trained on only one dataset, our models demonstrate strong generalizability across various benchmarks. Analysis reveals that ReSearch naturally elicits advanced reasoning capabilities such as reflection and self-correction during the reinforcement learning process.

Although contrastive self-supervised learning (SSL) on dynamic graphs has made significant success, the issue of heavy reliance on data augmentation and training tricks has been a persistent pain point. Generative SSL, especially masked autoencoders (MAEs) have recently produced promising results and can avoid these issues. However, the research on MAE in dynamic graphs remains largely unexplored due to the following challenges: 1) how to design an effective masking strategy for dynamic graphs? and 2) how to design a decoder to retain temporal dependency when graphs are perturbed? In this article, we propose DVGMAE, a novel dynamic variational graph masked autoencoder model to solve these challenges. DVGMAE simultaneously captures the evolving behaviors and topological features via an innovative masking strategy and an elaborate decoder. Specifically, we first implement a temporal-aware masking strategy on the edges of each snapshot based on the updated probabilities derived from historical mask information. This strategy mitigates potential masking bias in dynamic graphs. We then design a globally enhanced decoder to recover the temporal and spatial information of each snapshot. Extensive experiments demonstrate that DVGMAE outperforms the existing state-of-the-art on various tasks across different datasets.

Generative self-supervised learning (SSL), especially masked autoencoders (MAE), has greatly succeeded and garnered substantial research interest in graph machine learning. However, the research of MAE in dynamic graphs is still scant. This gap is primarily due to the dynamic graph not only possessing topological structure information but also encapsulating temporal evolution dependency. Applying a random masking strategy which most MAE methods adopt to dynamic graphs will remove the crucial subgraph that guides the evolution of dynamic graphs, resulting in the loss of crucial spatio-temporal information in node representations. To bridge this gap, in this paper, we propose a novel Informative Subgraphs Aware Masked Auto-Encoder in Dynamic Graph, namely DyGIS. Specifically, we introduce a constrained probabilistic generative model to generate informative subgraphs that guide the evolution of dynamic graphs, successfully alleviating the issue of missing dynamic evolution subgraphs. The informative subgraph identified by DyGIS will serve as the input of dynamic graph masked autoencoder (DGMAE), effectively ensuring the integrity of the evolutionary spatio-temporal information within dynamic graphs. Extensive experiments on eleven datasets demonstrate that DyGIS achieves state-of-the-art performance across multiple tasks.

Change point detection in dynamic networks aims to detect the points of sudden change or abnormal events within the network. It has garnered substantial interest from researchers due to its potential to enhance the stability and reliability of real-world networks. Most change point detection methods are based on statistical characteristics and phased training, and some methods are required to set the percent of change points. Meanwhile, existing methods for change point detection suffer from two limitations. On one hand, they struggle to extract snapshot features that are crucial for accurate change point detection, thereby limiting their overall effectiveness. On the other hand, they are typically tailored for specific network types and lack the versatility to adapt to networks of varying scales. To solve these issues, we propose a novel unified end-to-end framework called Variational Graph Gaussian Mixture model (VGGM) for change point detection in dynamic networks. Specifically, VGGM combines Variational Graph Auto-Encoder (VGAE) and Gaussian Mixture Model (GMM) through joint training, incorporating a Mixture-of-Gaussians prior to model dynamic networks. This approach yields highly effective snapshot embeddings via VGAE and a dedicated readout function, while automating change point detection through GMM. The experimental results, conducted on both real-world and synthetic datasets, clearly demonstrate the superiority of our model in comparison to the current state-of-the-art methods for change point detection.

Temporal community detection is helpful to discover and analyze significant groups or clusters hidden in dynamic networks in the real world. A variety of methods, such as modularity optimization, spectral method, and statistical network model, has been developed from diversified perspectives. Recently, network embedding-based technologies have made significant progress, and one can exploit deep learning superiority to network tasks. Although some methods for static networks have shown promising results in boosting community detection by integrating community embedding, they are not suitable for temporal networks and unable to capture their dynamics. Furthermore, the dynamic embedding methods only model network varying without considering community structures. Hence, in this article, we propose a novel unsupervised dynamic community detection model, which is based on network embedding and can effectively discover temporal communities and model dynamic networks. More specifically, we propose the community prior by introducing the Gaussian mixture model (GMM) in the variational autoencoder, which can obtain community information and better model the evolutionary characteristics of community structure and node embedding by utilizing the variant of gated recurrent unit (GRU). Extensive experiments conducted in real-world and artificial networks demonstrate that our proposed model has a better effect on improving the accuracy of dynamic community detection.

Anomaly detection in dynamic networks aims to find network elements (e.g., nodes, edges, subgraphs, change points) with significantly different behaviors from the vast majority, it can also devote to community detection and evolution and prediction tasks. Most existing methods focus on one specific task, that is, only detect anomalies of one type of element isolated, so they lose the ability to model the correlation and driving mechanism between different abnormal behavior. Considering that the anomaly detection of one type of element is helpful to other types of elements, i.e., the temporal evolution hidden the dynamic networks are driven by indivisible behavior patterns. So in this paper, we propose a unified Generation model to analyze the dynamic network for Exploring the Abnormal Behaviors of different Scales (GEABS). It can model the relation and catch different levels (node, community and network) of anomaly with a joint statistical network model and detect the community structure and its evolution. Specifically, we denote the parameters of node popularity, community membership to generate the dynamic network with stochastic block model (SBM), we also describe the varying of node and community by dynamic process. With a well-designed generative mechanism, it can detect the change point on network level, temporal evolution on community level and abnormal behavior on node level synchronously, besides, it also detects the community structure effectively. We also propose an effective optimization algorithm with variational inference. Experimental results show that the GEABS achieves better performance on abnormal behavior and community structure compared with baselines.

A variety of methods have been proposed for modeling and mining dynamic complex networks, in which the topological structure varies with time. As the most popular and successful network model, the stochastic block model (SBM) has been extended and applied to community detection, link prediction, anomaly detection, and evolution analysis of dynamic networks. However, all current models based on the SBM for modeling dynamic networks are designed at the community level, assuming that nodes in each community have the same dynamic behavior, which usually results in poor performance on temporal community detection and loses the modeling of node abnormal behavior. To solve the above-mentioned problem, this article proposes a hierarchical Bayesian dynamic SBM (HB-DSBM) for modeling the node-level and community-level dynamic behavior in a dynamic network synchronously. Based on the SBM, we introduce a hierarchical Dirichlet generative mechanism to associate the global community evolution with the microscopic transition behavior of nodes near-perfectly and generate the observed links across the dynamic networks. Meanwhile, an effective variational inference algorithm is developed and we can easy to infer the communities and dynamic behaviors of the nodes. Furthermore, with the two-level evolution behaviors, it can identify nodes or communities with abnormal behavior. Experiments on simulated and real-world networks demonstrate that HB-DSBM has achieved state-of-the-art performance on community detection and evolution. In addition, abnormal evolutionary behavior and events on dynamic networks can be effectively identified by our model.

In convolutional neural networks (CNNs), we propose to estimate the importance of a feature vector at a spatial location in the feature maps by the network's uncertainty on its class prediction, which can be quantified using the information entropy. Based on this idea, we propose the entropy-based feature weighting method for semantics-aware feature pooling which can be readily integrated into various CNN architectures for both training and inference. We demonstrate that such a location-adaptive feature weighting mechanism helps the network to concentrate on semantically important image regions, leading to improvements in the large-scale classification and weakly-supervised semantic segmentation tasks. Furthermore, the generated feature weights can be utilized in visual tasks such as weakly-supervised object localization. We conduct extensive experiments on different datasets and CNN architectures, outperforming recently proposed pooling methods and attention mechanisms in ImageNet classification as well as achieving state-of-the-arts in weakly-supervised semantic segmentation on PASCAL VOC 2012 dataset.

We present a deep-learning architecture for multi-view stereo with conditional random fields (MVSCRF). Given an arbitrary number of input images, we first use a U-shape neural network to extract deep features incorporating both global and local information, and then build a 3D cost volume for the reference camera. Unlike previous learning based methods, we explicitly constraint the smoothness of depth maps by using conditional random fields (CRFs) after the stage of cost volume regularization. The CRFs module is implemented as recurrent neural networks so that the whole pipeline can be trained end-to-end. Our results show that the proposed pipeline outperforms previous state-of-the-arts on large-scale DTU dataset. We also achieve comparable results with state-of-the-art learning based methods on outdoor Tanks and Temples dataset without fine-tuning, which demonstrates our method's generalization ability.

Semantic attention has long been adopted to image captioning models to enhance the image captioning performances. The models pre-trained for attribute recognition are utilized to generate image attributes in image captioning. Generally, these models are not jointly trained with image captioning models. In this paper, we propose attribute refinement network, which incorporates attribute recognition with image captioning to boost the performance on both tasks. We model the correlation between attributes with the semantic information from image captioning to improve the recognition accuracy. In turn, better attribute recognition results effectively enhance image captioning performance. Our model achieves CIDEr-D/SPICE scores of 115.1 and 20.9 respectively on the MS COCO test set, comprehensively yields improvement over all compared methods.

Human parsing has recently attracted considerable interests due to its wide application potentials. However, developing an accurate human parsing system is still a challenge for researchers. In this paper, we demonstrate that global information are critical for accurate prediction by applying a non-local operation for effectively extracting global information. Meanwhile several training data refinement methodologies are proposed to further boost the performance. Benefiting from all the approaches, the proposed single human parsing model NLGINet achieves the state-of-the-art segmentation accuracy on two human parsing benchmark datasets LIP and Pascal-Person-Parts.

We propose CFS, a distributed file system for large scale container platforms. CFS supports both sequential and random file accesses with optimized storage for both large files and small files, and adopts different replication protocols for different write scenarios to improve the replication performance. It employs a metadata subsystem to store and distribute the file metadata across different storage nodes based on the memory usage. This metadata placement strategy avoids the need of data rebalancing during capacity expansion. CFS also provides POSIX-compliant APIs with relaxed semantics and metadata atomicity to improve the system performance. We performed a comprehensive comparison with Ceph, a widely-used distributed file system on container platforms. Our experimental results show that, in testing 7 commonly used metadata operations, CFS gives around 3 times performance boost on average. In addition, CFS exhibits better random-read/write performance in highly concurrent environments with multiple clients and processes.

We propose a large-margin Gaussian Mixture (L-GM) loss for deep neural networks in classification tasks. Different from the softmax cross-entropy loss, our proposal is established on the assumption that the deep features of the training set follow a Gaussian Mixture distribution. By involving a classification margin and a likelihood regularization, the L-GM loss facilitates both a high classification performance and an accurate modeling of the training feature distribution. As such, the L-GM loss is superior to the softmax loss and its major variants in the sense that besides classification, it can be readily used to distinguish abnormal inputs, such as the adversarial examples, based on their features' likelihood to the training feature distribution. Extensive experiments on various recognition benchmarks like MNIST, CIFAR, ImageNet and LFW, as well as on adversarial examples demonstrate the effectiveness of our proposal.

Subpixel translation estimation using phase correlation is a fundamental task for numerous applications in the remote sensing community. The major drawback of the existing subpixel phase correlation methods lies in their sensitivity to corruption, including aliasing and noise, as well as the poor performance in the case of practical remote sensing data. This paper presents a novel subpixel phase correlation method using singular value decomposition (SVD) and the unified random sample consensus (RANSAC) algorithm. In the proposed method, SVD theoretically converts the translation estimation problem to one dimensions for simplicity and efficiency, and the unified RANSAC algorithm acts as a robust estimator for the line fitting, in this case for the high accuracy, stability, and robustness. The proposed method integrates the advantages of Hoge's method and the RANSAC algorithm and avoids the corresponding shortfalls of the original phase correlation method based only on SVD. A pixel-to-pixel dense matching scheme on the basis of the proposed method is also developed for practical image registration. Experiments with both simulated and real data were carried out to test the proposed method. In the simulated case, the comparative results estimated from the generated synthetic image pairs indicate that the proposed method outperforms the other existing methods in the presence of both aliasing and noise, in both accuracy and robustness. Moreover, the pixel locking effect that commonly occurs in subpixel matching was also investigated. The degree of pixel locking effect was found to be significantly weakened by the proposed method, as compared with the original Hoge's method. In the real data case, experiments using different bands of ZY-3 multispectral sensor-corrected images demonstrate the promising performance and feasibility of the proposed method, which is able to identify seams of the image stitching between sub-charge-coupled device units.