Quality Characterization and Adaptive Testing

When a system has countless possible inputs, which tests are most likely to reveal problems?

This topic studies how quality measures and testing evidence can guide test generation, selection, and prioritization. DeepGini (ISSTA 2020) investigates test prioritization for deep neural networks, while Adaptive Test Selection (ICSE 2022) explores how to adapt test selection for these systems. For question-answering systems, QATest (ASE 2022) provides a uniform fuzzing framework.

Multi-Source Defect Diagnosis and Robustness Enhancement

When an AI system makes a mistake, does the problem come from its data, its model, or the conditions in which it operates?

This topic investigates different sources of failure and develops techniques for diagnosis and robustness testing. At the data level, Datactive (ISSTA 2024) localizes data faults in object detection systems. At the model level, AtPatch (ICSE 2026) diagnoses and corrects anomalous attention in Transformer models. Research on operating conditions includes Drivence (TSE 2026), which synthesizes realistic driving sequences to test multi-sensor fusion perception systems.

Trustworthiness and Security of Large Language Models and Agents

When should we trust a model’s answer or an agent’s action?

This topic investigates how to evaluate models and agents in the contexts where they are used. TACO (ICSE 2026) studies the trustworthiness of large language models in coding assistance, while VADA (ICSE 2026) supports value-aware data generation and alignment evaluation across cultural contexts. The team is also investigating the behavioral trustworthiness and security of agents, including their responses to unsafe or adversarial interactions.

System Testing and Fault Diagnosis for AI Infrastructure

How can we find and diagnose failures in the software that trains and runs AI models?

This topic studies defects in the libraries, frameworks, and distributed systems supporting AI applications. An empirical study of deep learning framework bugs (Information and Software Technology 2022) examines their characteristics, providing evidence for understanding infrastructure failures. Building testing techniques for these systems, MoCo (TSE 2025) generates tests for deep learning libraries through code assembly. For distributed-system debugging, RediI (ICSE 2025) enables deterministic failure reproduction. The team is currently extending its research to testing, fault diagnosis, and failure reproduction in inference services and distributed training systems.

Type Analysis and Compilation Efficiency

How can tools infer useful information about a program, and how can its compilation process become more efficient?

This topic focuses on type analysis and the organization of compilation workflows. The team is currently working on type analysis for dynamically typed programs, together with techniques for reusing compilation results and coordinating distributed compilation tasks. These efforts aim to improve the information available to development tools and reduce the overhead of building software.

Semantic Preservation and Constraint Adaptation in Cross-Language Translation

How can we migrate C/C++ software to Rust while preserving its intended behavior?

Moving between languages requires addressing differences in types, memory management, ownership, and interfaces. The team is currently working on trustworthy C/C++-to-Rust translation, investigating how to preserve program semantics while adapting to Rust’s language constraints. A central concern is how to assess the correctness of translated programs and support practical migration with explicit attention to semantic preservation.

Language Constraints for Defect Detection and Repair

How can language rules and API contracts help us generate better tests and find bugs?

This topic uses language properties, program context, and API information to guide testing and analysis, with Rust as a central focus. FRIES (ISSTA 2024) tests Rust library interactions through ecosystem-guided target generation. PALM (ASE 2025) combines program analysis and language models to improve unit-test coverage, while RustGTC (ASE 2026, accepted) investigates constraint-guided API test generation using comments and code context.

The research also addresses consistency checking, repair, and compiler diagnostics. RustC4++ (TOSEM 2026) detects inconsistencies between Rust code and comments, and PanicFI (TOSEM 2026) provides infrastructure for fixing panic bugs. At the toolchain level, SugBreaker (FSE 2026) tests the reliability of Rust compiler fix suggestions through constraint-violation-guided mutation.

Distributed File Systems and Heterogeneous Engines for AI

How can distributed storage and heterogeneous engines better support AI workloads?

The team is currently working on the design and implementation of distributed file systems and heterogeneous engines for AI. The file-system work addresses the storage and data-access needs of AI workloads, while the engine work concerns execution support in heterogeneous environments. This research brings together storage-system design and execution-engine development to provide efficient and reliable infrastructure for AI applications.