Requirements Engineering (RE) is an initial and critical phase in software development, with the aim of producing well-defined software requirements specifications (SRSs) from rough ideas of clients. It involves multiple tasks (e.g., elicitation, analysis) and roles (e.g., interviewer, analyst). With the rise of Large Language Models (LLMs), many studies have leveraged LLMs to support specific RE tasks. However, existing LLM-based agents often lack domain knowledge integration and fall short in simulating the complex collaboration of human experts across the full RE process. To address this gap, we propose KGMAF, a knowledge-guided multi-agent framework designed to assist requirements engineers in developing high-quality SRSs. KGMAF comprises six LLM-based agents and a shared artifact pool. Each agent is equipped with predefined actions, dedicated functions, and injected knowledge tailored to specific RE tasks. The artifact pool stores both intermediate and final artifacts, serving as a communication channel for inter-agent collaboration. A human-in-the-loop (HITL) mechanism is embedded to guide and validate agent outputs. We present the design of KGMAF, along with preliminary experiments and a case study to demonstrate its practicality. This work lays the foundation for future research on knowledge-driven multi-agent collaboration in RE and highlights key challenges in building trustworthy intelligent assistants for real-world RE tasks.

Binary Code Similarity Detection (BCSD), a technique for assessing the similarity between two given binary code snippets, holds significant value in searching for vulnerable functions within embedded device firmware, which is typically closed-source. However, existing BCSD approaches face two major challenges: the irreversible loss of semantic and structural information during the process of binary code compilation, which affects detection performance; and the inability to directly perform similarity detection between binary code and source code. In this paper, we present Bin2SrcSim, a novel BCSD approach that employs a Large Language Model (LLM) to convert binary code into source code representations. Bin2SrcSim fine-tunes an LLM at the function-level to transform assembly code and pseudocode into source code. Consequently, the similarity between any two binary code functions can be assessed by calculating the cosine similarity and Jaccard similarity of the transformed source code. The experimental results demonstrate that Bin2SrcSim outperforms all baselines, achieving Recall@1 scores of 0.82, 0.83, 0.93, and 0.81 across various scenarios involving cross-architecture, cross-compiler, and cross-optimization levels. Bin2SrcSim also demonstrates satisfactory performance in vulnerable function search within real-world IoT device firmware. Moreover, Bin2SrcSim supports similarity detection between binary code and source code, expanding the scope of detection applications.

The task of address matching involves linking unstructured addresses to standard ones in a database. The challenges presented by this task are manifold: misspellings, incomplete information, and variations in address content are some examples. While there have been previous studies on entity matching in natural language processing, for the address matching solution, existing approaches still rely on string-based similarity matching or manually-designed rules. In this paper, we propose StructAM, a novel method based on pre-trained language models (LMs) and graph neural networks to extract the textual and structured information of the addresses. The proposed method leverages the knowledge acquired by large language models during the pre-training phase, and refines it during the fine-tuning process on the address domain, to obtain address-specific semantic features. Meanwhile, it also applies an attribute attention mechanism based on Graph Sampling and Aggregation (GraphSAGE) module to capture internal hierarchy information of the address text. To further enhance the accuracy of our algorithm in dirty settings, we incorporate spatial coordinates and contextual information from the surrounding area as auxiliary guidance. We conduct extensive experiments on real-world datasets from four different countries and the results show that StructAM outperforms state-of-the-art baseline approaches for address matching.

Abstract Federated learning (FL) is a decentralized and privacy-preserving machine learning technique that protects data privacy by learning models locally and not sharing datasets. However, due to limited computing resources on devices and highly heterogeneous data in practical situations, the training efficiency and resource utilization of federated learning is low. In order to resolve these challenges, we introduce a blockchain-assisted dynamic adaptive and personalized federated learning framework (TV-FedAvg) in the presence of restricted computing power resources and data heterogeneity. After each round of local training, we utilize an improved scoring model based on VIKOR and TOPSIS to comprehensively score the devices. The scores are then utilized to choose devices for participation in global aggregation and to carry out model aggregation through blockchain consensus. Furthermore, resources are reallocated for the next round to enhance resource efficiency, model fairness, and performance. Finally, we demonstrate through experimentation that TV-FedAvg outperforms other models such as pFedMe, FedAvg, Per-FedAvg, and TOPSIS in terms of both efficiency and performance.

Despite myriad efforts in the literature designing neural representation system for emotion detection, very few works consider constructing effective model for apperceiving various emotion intensity on social media because of the informal expression and lack of context. In this paper, we proposed a character-aware convolutional recurrent networks with self-attention for emotion detection on user-generated content. The proposed model contains three parts: the character-level convolutional layer is designed to learn word representation based on character n-grams for capturing subword information and independent on pre-trained word embedding. The recurrent neural networks learn the sequential context information used both forward and backward recurrent neural network. And the self-attention module is used to extract different emotion aspects of the sentence into multiple vector representations. The attention module performs on top of recurrent networks which enables attention to be used in special domain or task when there are no extra inputs. We evaluate the proposed model on two public emotion datasets including both emotion intensity detection and emotion classification. We compare our model with the state-of-the-art methods on these datasets and the experimental results demonstrate that the proposed model outperforms several baselines on most emotion types detection and indicates the effectiveness of the designed model. In addition, the training of the proposed model for these tasks relies exclusively on initialized character vector which can be used for morphologically rich languages with long-tailed frequency distributions or domains with dynamic vocabuaries.