Speech-to-Speech (S2S) Large Language Models (LLMs) are foundational to natural human-computer interaction, enabling end-to-end spoken dialogue systems. However, evaluating these models remains a fundamental challenge. We propose SageLM, an end-to-end, multi-aspect, and explainable speech LLM for comprehensive S2S LLMs evaluation. First, unlike cascaded approaches that disregard acoustic features, SageLM jointly assesses both semantic and acoustic dimensions. Second, it leverages rationale-based supervision to enhance explainability and guide model learning, achieving superior alignment with evaluation outcomes compared to rule-based reinforcement learning methods. Third, we introduce SpeechFeedback, a synthetic preference dataset, and employ a two-stage training paradigm to mitigate the scarcity of speech preference data. Trained on both semantic and acoustic dimensions, SageLM achieves an 82.79% agreement rate with human evaluators, outperforming cascaded and SLM-based baselines by at least 7.42% and 26.20%, respectively.

Yuan Ge, Haishu Zhao, AoKai Hao, Junxiang Zhang, Bei Li, Xiaoqian Liu, Chenglong Wang, Jianjin Wang, Bingsen Zhou, Bingyu Liu, JingBo Zhu, Zhengtao Yu, Tong Xiao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Demand for precise evaluation and in-depth analysis of patent technology value has become increasingly urgent with the accelerated iteration of innovation, while existing patent platforms primarily focus on surface-level bibliographic information, offering limited support for understanding the technological value and evolutionary trajectories of patents. Hence, we develop PatentMerit, a comprehensive system for in-depth analysis of the value of patents, based on large language models, network analysis, and data mining techniques. PatentMerit features five core functionalities, including multidimensional patent classification retrieval, disruptive technology identification, citation network and technology evolution visualization, topic semantic analysis, and automated comprehensive report generation. Its core strength lies in transcending single data element analysis to conduct in-depth exploration, such as accurate technology assessment of individual patents and their technological clusters, which reduces the cognitive and operational burden for R&D personnel. PatentMerit serves as an intelligent decision-making tool for technology trend forecasting, R&D decision-making, and patent strategy planning. The PatentMerit system is accessible via the following link: https://patentmerit.com/.

Effective representations of complex web semantic graphs are essential for various web applications, including link prediction, recommendation systems, and social network analysis. However, existing methods assume that multi-interactions (or multi-relationships) between two connected nodes are independent, while these relationships inherently exhibit characteristics of mutual promotion or mutual inhibition. Moreover, these semantic characteristics across different relationships cannot be easily captured by a simple linear combination. To tackle this challenge, we propose an Adaptive Multi-Interaction (AMI) web semantic graph representation method. Specifically, AMI consists of three modules, including a multi-interaction aggregation module, a global pattern aggregation module, and an adaptive relation-specific decoder module. Firstly, we construct a learnable multi-interaction behavior pattern matrix that captures the mutual promotion and mutual inhibition effects between two connected nodes. Secondly, the global pattern aggregation module is designed to efficiently capture global homogeneous interaction patterns through graph convolution networks. Finally, the adaptive relation-specific decoder module employs a hybrid scoring strategy to adaptively decode node embeddings based on their distinct relationships. Extensive experiments on benchmark web datasets for link prediction tasks demonstrate that AMI outperforms state-of-the-art baselines. Our codes are available at https://github.com/AI-stronger123/AMI.

Design patterns are powerful tools that provide standardized solutions to common problems in the software design process. Detecting design patterns in software systems can greatly facilitate software maintenance and code understanding, but manual detection is a complex and challenging task. In recent years, with the development of deep learning techniques, researchers have proposed automatic design pattern detection methods that utilize code features and machine learning classifiers. However, these approaches typically focus on either code semantic features or structural features, lacking the integration of both. To address this issue, in this paper, we propose a novel design pattern detection method with multi-feature fusion, DPDMF2, which effectively fuses code semantic features and software network structural features for design pattern detection. Specifically, we first represent software as a CSN (Class-level Software Network) to capture its topology, including classes, their relationships, and the types of these relationships. Concurrently, we utilize code representation learning techniques to obtain semantic features of the code, which are then used as node attribute features within the network. Finally, we utilize graph neural networks (GNNs) to fuse these two types of features, generating a vector representation that encapsulates both semantic and structural information for design pattern detection task. Experimental results on public datasets show that DPDMF2achieves 81.7% precision and 82.1% recall, validating the overall effectiveness of our method.

Enhancing model performance on low-capacity devices remains a significant challenge in the Internet of Models (IoM). Inspired by the efficiency of disentangled representation learning, this study proposes a model decomposition approach to improve response time on low-capacity devices while maintaining accuracy, through collaborative learning with models deployed on higher-capacity devices. First, we introduce a divide-and-conquer strategy that decomposes and learns knowledge within each data sample. This enables lightweight sub-models, tailored to specific knowledge components, to respond efficiently on low-capacity devices in the IoM system. Second, we design a progressive crossfusion mechanism to promote mutual enhancement among these knowledge-specific sub-models. Third, we optimize the learning process by aligning the updates of these sub-models with those of the models on higher-capacity devices. This collaborative optimization is guided via knowledge distillation during model aggregation. Experimental results on image classification tasks demonstrate that our approach reduces response time by 9.87 % to 60.31 % without compromising accuracy.

In modern society, software systems are integral to various aspects of life. Finding an efficient vulnerability identification approach is crucial for ensuring security and preventing malicious attacks. In recent years, many deep learning-based methods have shown outstanding performance in the vulnerability detection task. However, these methods still have limitations. Some methods consider code input as token sequences and apply architectures typically used in natural language processing. They fail to utilize the structural information from various code components' interactions, which limits these models' performance. Other methods based on graph neural networks, although better at learning structural information, treat each node equally and fail to emphasize key elements. To overcome these limitations, we propose CTGGSG, an effective vulnerability detection approach over Code Token Graph (CTG) based on GCN with score gate. In our model, we use PLE-CG-SE module to represent the source code samples as the CTGs, effectively utilizing the high-quality feature representation of PL-PLM (Pre-trained Language Model based on Program Languages) and retaining structural information from the source code. During the graph learning process, we combine GCN convolution and score gate mechanism to make the model focus more on the key nodes within the graph and increase the receptive field of the nodes. To comprehensively evaluate the performance and scalability of our model, we conducted experiments on two real-world datasets: CodeX Glue,which contains balanced sample labels, and Re-veal, which contains imbalanced sample labels. These datasets contain 27,318 and 22,734 function-level samples, respectively, derived from large-scale, popular real-world projects. Compared to existing advanced vulnerability detection methods, our model achieved state-of-the-art performance overall.

Deep reinforcement learning has attracted wide interest because of its extraordinary capabilities in multiple fields. However, morphology optimization by using evolutionary computation techniques has not been intensively investigated. In this paper, we explore the use of genetic algorithms (GA) to automatically design the morphology of an agent. Evaluating the performance of an agent is very time-consuming because it needs to be trained from scratch. Moreover, it is computationally infeasible to train separate controllers for all possible different morphologies of agents to identify the optimal ones and is difficult to obtain the accurate cumulative reward of an agent to estimate the performance of the morphologies. To address these issues, we use a morphology comparator as a surrogate model to estimate the probability of one morphology being better than the other, instead of directly predicting the performance of each morphology. A set of surrogate models based on a radial basis function network are developed before evolution to make full use of the data to guide the search. Experimental results indicate that the proposed method is able to efficiently find out optimal morphologies to achieve better performance than the default morphology.