IEEE CS Macau Chapter ForumAPSEC 2025

Zhi Jin

Title: ClarifySTL: LLM-powered transformation of natural language descriptions into signal temporal logic specifications

Abstract: Behavior constraints are key concerns when modeling Cyber Physical Systems (CPSs), which are becoming the important intelligent infrastructure in the digitalization era. Signal Temporal Logic (STL), a formal description language that can precisely express temporal logical constraints between continuous-time signals, has been widely used in the system modeling of CPSs. However, in most cases, CPSs behavior constraints are usually recorded and communicated in natural language provided by domain experts. At present, CPSs modeling practices mainly rely on manually writing STL formulas by system modelers. This process requires rich logical modeling experience and is highly prone to semantic deviations or logical errors. With the rapid development of Natural Language Processing (NLP), many attempts are conducted to use NLP techniques including LLM-enabled techniques to realize the automatic conversion from natural language to STL formulas. However, challenges such as the lack of corpora, the inherent ambiguity and incompleteness of natural language, and the syntactic complexity of STL pose great obstacles to the end-to-end mapping. This talk presents the exploration about data construction, constraints clarification, and reinforcement learning on the task of STL specification transformation using LLMs. It also verifies the significant improvements of the proposed approach in accuracy and interpretability across multiple benchmark tasks,demonstrating the feasibility in the automated extraction and specification of temporal constraints for modeling the real-world CPSs.

Bio: Zhi Jin is professor of computer science at Peking University and Hongyi Visiting Professor at Wuhan University. Her main research interest is AI for SE, with a long-term focus on domain knowledge-led requirements engineering. She has published over 300 scientific articles in refereed international journals, such as IEEE T-KDE, T-SE, ACM T-OSEM, and T-CHI, and high rank conferences, such as ICSE, FSE, ASE, ACL and RE. She has co-authored five books and has held more than 30 approved invention patents. She is five times recipient of ACM SIGSOFT Distinguished Paper Awards.

Shing-Chi Cheung

Speaker: Prof. Shing-Chi Cheung（The Hong Kong University of Science and Technology）

Title: Can LLMs assist with software engineering tasks? The road ahead

Abstract: Large Language Models (LLMs) are said to offer promises for automating a wide range of software engineering tasks. GitHub’s leader forecasted LLMs would craft most of our code just months ago, and early adopters are already embracing the shift with confidence. The talk shares some preliminary thoughts on mapping a viable path to harness LLMs for assuring software quality. It will discuss the deployment of LLMs in two scenarios: test generation and formal verification. In the test generation scenario, it explores an intention-first strategy that leverages LLMs’ outstanding language capabilities. In the formal verification scenario, it investigates whether LLMs can transform formal verification by automating the translation of natural language program descriptions into formal specifications and generating verifiable proofs.

Biography: Professor Shing-chi Cheung from the Hong Kong University of Science and Technology (HKUST) specializes in leveraging advanced testing methodologies, artificial intelligence technologies, and empirical research techniques to identify, diagnose, and repair faults in reliable and intelligent software systems. He is a Chair Professor of Computer Science and Engineering and an IEEE Fellow. Professor Cheung and his research team have been working on ensuring the quality of software systems using methodologies adopted by both academia and industry. In 1998, Professor Cheung introduced Metamorphic Testing, which has since emerged as a leading testing methodology for artificial intelligence systems. Recently, his team has been exploring the application of generative AI for software development and maintenance.

Hongyu Zhang

Title: Agentic Software Engineering

Abstract: In the era of big data and artificial intelligence, our goal is to achieve intelligent software engineering. In our earlier work, we built machine learning-based models by mining software-related data to automate tasks such as programming, testing, maintenance and operation, thereby improving software development productivity and reducing maintenance work. In recent years, LLM-based agents have been widely applied to intelligent software engineering, changing the way software is developed and maintained. This “Agentic Software Engineering” approach is a new “AI-native” approach to software engineering. However, the development of agentic software engineering still faces challenges. This report will review the development history of intelligent software engineering, introduce agentic software engineering, and discuss the challenges and possible solutions.

Bio: Hongyu Zhang is Distinguished Professor and Dean of School of Big Data and Software Engineering, Chongqing University, China. He is also an honorary professor at the University of Newcastle, Australia. He received his PhD degree from National University of Singapore in 2003. His research is in the area of Software Engineering, in particular, intelligent software engineering, software analytics, maintenance, and quality assurance. He has published more than 250 research papers in reputable international journals and conferences, and received 12 ACM/IEEE Distinguished Paper awards. He is a Distinguished Member of ACM, a Distinguished Member of CCF, a Senior Member of IEEE, and a Fellow of Engineers Australia (FIEAust). He is a recipient of David Parnas Fellowship in 2024.