Multi-agent systems (MAS) built on large language models (LLMs) provide a promising framework for solving complex tasks through role specialization and structured interaction. However, their performance is often limited by miscoordination and, more fundamentally, the lack of fine-grained credit assignment across agents. Existing approaches typically rely on coarse-grained feedback, making it difficult to identify which agents or interaction steps are responsible for errors. We propose Gradient-Based Connections (GBC), an approach for fine-grained attribution and optimization of multi-agent systems. GBC models a MAS as a computational graph and introduces gradient-based connection weights to quantify the influence of each agent’s output on downstream agents at the token level. By constructing an attribution graph and propagating task-specific loss signals backward, our method enables precise identification of error sources and targeted prompt optimization. We further develop AgentChord, an efficient implementation that leverages prefix-based gradient computation. Experiments on MultiWOZ and τ-bench show that GBC improves multi-agent performance and outperforms strong single-agent and multi-agent baselines, and higher attribution quality is associated with greater optimization effectiveness. Code is available at: https://github.com/yxc-cyber/AgentChord.

Abstract User simulators are essential to conversational AI, enabling scalable agent development and evaluation through simulated interactions. While current Large Language Models (LLMs) have advanced user simulation capabilities, we reveal that they struggle to consistently demonstrate goal-oriented behavior across multi-turn conversations, which is a critical limitation that compromises their reliability in downstream applications. We introduce User Goal State Tracking (UGST), a novel framework that tracks user goal progression throughout conversations. Leveraging UGST, we present a three-stage methodology for developing user simulators that can autonomously track goal progression and reason to generate goal-aligned responses. Moreover, we establish comprehensive evaluation metrics for measuring goal alignment in user simulators, and demonstrate that our approach yields substantial improvements across two benchmarks (MultiWOZ 2.4 and τ-Bench). Our contributions address a critical gap in conversational AI and establish UGST as an essential framework for developing goal-aligned user simulators. All code and data is released to facilitate future research 1.

Cheng Qian, Peixuan Han, Qinyu Luo, Bingxiang He, Xiusi Chen, Yuji Zhang, Hongyi Du, Jiarui Yao, Xiaocheng Yang, Denghui Zhang, Yunzhu Li, Heng Ji. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Language Models (LLMs) have shown remarkable capabilities as autonomous agents; yet existing benchmarks either focus on single-agent tasks or are confined to narrow domains, failing to capture the dynamics of multi-agent coordination and competition. In this paper, we introduce MultiAgentBench, a comprehensive benchmark designed to evaluate LLM-based multi-agent systems across diverse, interactive scenarios. Our framework measures not only task completion but also the quality of collaboration and competition using novel, milestone-based key performance indicators. Moreover, we evaluate various coordination protocols (including star, chain, tree, and graph topologies) and innovative strategies such as group discussion and cognitive planning. Notably, gpt-4o-mini reaches the average highest task score, graph structure performs the best among coordination protocols in the research scenario,and cognitive planning improves milestone achievement rates by 3%. Code and datasets are publicavailable at https://github.com/ulab-uiuc/MARBLE.

Reconstructing the brightness temperature map from the visibilities in synthetic aperture interferometric radiometers (SAIRs) has been proved to be an ill-posed inverse problem. The regularization methods are crucial to effectively overcome the ill-condition of the inverse problem. This letter presents a maximum a posteriori (MAP) approach for retrieving the accurate brightness temperature map in SAIRs. Furthermore, edge-preserving regularization using a Huber–Markov random field (Huber-MRF) model is used to provide a stable solution and eliminate the oscillation artifacts. Numerical experiment results and quantitative analyses verify the effectiveness of the proposed method.

Heterogeneous graph neural networks (HGNNs) have emerged as powerful algorithms for processing heterogeneous graphs (HetGs), widely used in many critical fields. To capture both structural and semantic information in HetGs, HGNNs first aggregate the neighboring feature vectors for each vertex in each semantic graph and then fuse the aggregated results across all semantic graphs for each vertex. Unfortunately, existing graph neural network accelerators are ill-suited to accelerate HGNNs. This is because they fail to efficiently tackle the specific execution patterns and exploit the high-degree parallelism as well as data reusability inside and across the processing of semantic graphs in HGNNs. In this work, we first quantitatively characterize a set of representative HGNN models on GPU to disclose the execution bound of each stage, inter-semantic-graph parallelism, and inter-semantic-graph data reusability in HGNNs. Guided by our findings, we propose a high-performance HGNN accelerator, HiHGNN, to alleviate the execution bound and exploit the newfound parallelism and data reusability in HGNNs. Specifically, we first propose a bound-aware stage-fusion methodology that tailors to HGNN acceleration, to fuse and pipeline the execution stages being aware of their execution bounds. Second, we design an independency-aware parallel execution design to exploit the inter-semantic-graph parallelism. Finally, we present a similarity-aware execution scheduling to exploit the inter-semantic-graph data reusability. Compared to the state-of-the-art software framework running on NVIDIA GPU T4 and GPU A100, HiHGNN respectively achieves an average 40.0× and 8.3× speedup as well as 99.59% and 99.74% energy reduction with quintile the memory bandwidth of GPU A100.

Heterogeneous graph neural networks (HGNNs) have the powerful capability to embed rich structural and semantic information of a heterogeneous graph into node representations. Existing HGNNs inherit many mechanisms from graph neural networks (GNNs) designed for homogeneous graphs, especially the attention mechanism and the multi-layer structure. These mechanisms bring excessive complexity, but seldom work studies whether they are really effective on heterogeneous graphs. In this paper, we conduct an in-depth and detailed study of these mechanisms and propose the Simple and Efficient Heterogeneous Graph Neural Network (SeHGNN). To easily capture structural information, SeHGNN pre-computes the neighbor aggregation using a light-weight mean aggregator, which reduces complexity by removing overused neighbor attention and avoiding repeated neighbor aggregation in every training epoch. To better utilize semantic information, SeHGNN adopts the single-layer structure with long metapaths to extend the receptive field, as well as a transformer-based semantic fusion module to fuse features from different metapaths. As a result, SeHGNN exhibits the characteristics of a simple network structure, high prediction accuracy, and fast training speed. Extensive experiments on five real-world heterogeneous graphs demonstrate the superiority of SeHGNN over the state-of-the-arts on both accuracy and training speed.

Synthetic aperture imaging radiometers (SAIRs) are a powerful tool for high-resolution imaging. The reconstruction process from the visibility function to the brightness temperature map in SAIRs has been demonstrated to be an ill-posed inverse problem. Although the existing regularization methods can effectively overcome the ill-condition, there are still large residual errors and oscillation ripples, especially at the edges of the reconstructed map. In this letter, a reweighted total variation (RTV) method is presented to reconstruct the brightness temperature map in SAIRs and preserve the edge information of the map without windowing. A split Bregman iteration algorithm is used to optimize the RTV regularization model and increase the calculation speed. Numerical simulation experiments are carried out to verify the effectiveness and show the performance of the proposed method.

Multi-class object detection always involves the tasks of accurate target localization which is mainly related to bounding box regression. Smooth L1 loss is the most popular bounding box regression loss used in the current state-of-the-art object detection systems. However, such loss for regressing the parameters of a bounding box can't accurately and consistently regress the bounding box to the associated ground truth well. We instead propose the near-duplicated loss, a loss that better evaluate the disparity between the bounding box and the ground truth consistently. We present an approximate algorithm associated with a kernel function that not only considers the absolute distance but also involves the relative overlap area between the two bounding boxes. The new loss doesn't need additional supervision and is easy to embed into existing networks. Our final result, by incorporating the near-duplicated loss into the state-of-the-art object detection detectors (Faster RCNN, RetinaNet), shows consistent and significant improvements on popular object detection benchmarks (MS COCO and Pascal VOC).

Hyperspectral image (HSI) is usually composed of hundreds of bands which contain very rich spatial and spectral information. However, the high-dimensional data may lead to the curse of dimensionality phenomenon when it is used for land use classification or other applications, making it difficult to be utilized effectively. In this paper, we developed a deep learning classification framework based on the spectral and spatial information of hyperspectral image. Firstly, the deep learning features in different layers could be extracted automatically. Secondly, based on the learned deep learning features, we could obtain the classification of hyperspectral image with logistic regression (LR) classifier. Finally, we compared our approach with other methods including quadratic discriminant analysis with the multilevel logistic spatial prior (QDAMLL), logistic discriminant analysis with the multilevel logistic spatial prior (logDAMLL), linear discriminant analysis with the multilevel logistic spatial prior (LDAMLL), subspace multiclass logistic regression with the multilevel logistic spatial prior (MLRsub MLL), support vector machine on extended morphological profiles (SVM/EMP), support vector machine on expectation maximization and post-regularization (SVM-EM-PR). The experimental results showed that our method obtained the optimum accuracy, which was better than the other six approaches. And the OA was up to 99.39%. Therefore, the deep convolutional neural networks (DCNNs) is a robust method for land use classification with hyperspectral image.

A scheme of feature analysis in synthetic aperture radar (SAR) imaging based on compressed sensing (CS) is presented. First, a conventional imaging algorithm, such as range-Doppler algorithm or chirp-scaling algorithm, is used to acquire the complex image of the target scene. Then, a greedy algorithm is applied to this complex image to locate the scattering coefficients of the target, where the research of the peak and the removal of the point spread effect are involved. Third, the point set of each target is determined by a clutter method and the range of the target is further fitted by a geometric shape. They are, finally, used to extract the feature parameter of the target. A series of experimental results using real field data demonstrates the effectiveness of our method.

A novel despeckling algorithm for synthetic aperture radar (SAR) image is presented. This algorithm focuses on despeckling a local region for each target pixel. Thus, the calculations of similarities and weights are located within a limited region rather than the entire image. Also, the similarity is measured by using the Euclidean distance of features between the target pixel and neighborhood pixel. In addition, a cosine integral image technique is applied to this scheme to improve the computation efficiency. The weight function is first decomposed into a linear combination of cosine functions, and each convolution with a cosine function requires a constant number of operations per pixel. Then, all the summation operations are carried out by the summed image method. A series of experiment results demonstrates the effectiveness of our method.

The conventional range gated lidar systems require long laser pulse width and range bins to maintain the far detection range, which limits the ranging resolution of the system. A range gated lidar system with phase coded method is proposed with high ranging resolution. The system utilizes a kN+1 bits of the laser pulse sequence and controls the accumulation of the detector in the range bins to get the echo signals. The correlation calculation of the accumulated signals is conducted in the computer to acquire the targe range with the ranging resolution defined by the sequence bit width. The indoor demonstration experiment achieved 2 us resolution with the sampling period of 32 us by employing a 15-bit m-sequence. The result shows the ability of achieving high ranging resolution of range gated lidar.

Based on sparsity or compressibility of microwave radiation images, Compressive Sensing in this paper is adopted to achieve microwave radiation imaging in order to reduce the complexity and hardware cost of the imaging system and get the image of high spatial resolution. Compared with the common wavelet basis, differential matrix is proposed to sparsely represent microwave radiation images. OMP algorithm is adopted to reconstruct the microwave radiation image. The simulation results show that the differential matrix to sparsely represent microwave radiation images is better than the wavelet.

Full Polarization Interferometric Radiometer (FPIR) is concept of one dimensional synthetic radiometer. This paper introduces the design of the FPIR antenna. A novel linear array antenna is described with emphasis on the simulation of the antenna pattern, design of the stripline feeding network and simulation of the polarimetric isolation.

Design limitations for synthetic aperture radiometer are quantitatively analyzed. Firstly, the image quality of a small scaled system with a few baselines is analyzed. The image quality is getting poorer due to the truncation in the spatial frequency domain while the number of baselines is getting fewer. This is called the lower limit for the design. Secondly, the design limitation for a large scaled system, where many baselines exist, is analyzed related with system bandwidth. Once the bandwidth of a system is determined, the physical sampling intervals will behavior differently at the upper and lower frequencies which may generate confusion in the retrieved image. This will limit the increasing of spatial resolution while the number of baselines of a system is increasing. This is called the upper limit for the design.