Artificial-intelligence (AI) technologies and software have become deeply embedded in the information society, serving as the core infrastructure that powers efficient operations across every sector. Their trustworthiness has therefore emerged as one of the most critical challenges and research frontiers. Assistant Professor Jia Li is dedicated to constructing trustworthy AI and software systems that can operate safely, robustly, and responsibly in complex and ever-changing environments. Specifically, her research is organized around two complementary thrusts:

1. Trustworthy AI

Through technical innovations—ranging from novel model architectures to advanced alignment techniques—she systematically elevates the intrinsic trustworthiness of AI systems. Key research themes include:

➢ Security: ensuring AI systems withstand malicious attacks and remain stable and reliable under uncertainty.

➢ Privacy Preservation: developing technologies that rigorously protect user data and prevent the leakage of sensitive information.

2. Trustworthy Software

She leverages AI to enhance the trustworthiness of traditional software and to address real-world challenges. Key research themes include:

➢ Reliability: using AI to generate test cases, perform intelligent fuzzing, and detect and repair source-level bugs, thereby improving software reliability.

➢ Security: employing AI to identify and fix defects and vulnerabilities in code, reducing attack surfaces and strengthening system defenses.

➢ Advanced the convergence of AI and software engineering, accelerating large-model-based code generation. Led or co-led the training of multiple code-oriented large language models that set new international benchmarks on downstream tasks and provide the community with robust foundational models. Introduced deep-reasoning code-generation techniques that unlock the inferential power of large models, enabling them to tackle complex development demands. Established evaluation benchmarks grounded in real-world software projects, driving the adoption of large models in practical development.

➢ Over the past five years, published more than 20 papers in CCF-A top-tier conferences and journals (NeurIPS, ACL, ICSE, ASE, FSE, etc.), including several Oral presentations. These works have been cited over 1,000 times by researchers from institutions such as MIT, Stanford, NTU, and CUHK.

➢ Served on program committees of premier international conferences (e.g., ASE) and has been repeatedly invited to give oral presentations. Research outcomes have been featured by mainstream media including China Science Daily, China Daily, and Synced. Honors include Beijing Outstanding Graduate and the “Excellent Ph.D. Student Award” at the ChinaSoft Conference.

➢ Code-oriented Large Language Models

– Co-trained aiXcoder-7B (7 B params), the first LLM to explicitly inject code-structural priors—syntax, dependencies, etc.—into pre-training. Novelties span pre-training objectives, data sampling, and cleansing. aiXcoder-7B outperforms same-scale international baselines (e.g., Meta’s Code Llama-7B, DeepSeek-Coder-6.7B) on eight mainstream benchmarks, earned 2,271 GitHub stars, and ranked in Hugging Face’s Global Trending Top-30 (May 2024).

– Led the training of aiXcoder-7B-v2, which further boosts long-context capability via reinforcement learning and sets new records on repository-level code completion.

➢ Deep-Reasoning Code Generation

Introduced a four-stage reasoning pipeline—requirement understanding, planning, implementation, and optimization—mirroring real-world developer cognition. Each stage refines the previous output, yielding up to 88.4 % relative Pass@1 improvement. The technique has sparked follow-up studies from MIT, Peking University, and other leading groups.

➢ Real-Project-Aligned Evaluation Benchmarks

Proposed and open-sourced DevEval (static) and EvoCodeBench (dynamic), curated from high-quality open-source projects whose distribution mirrors real software. EvoCodeBench auto-updates to prevent data leakage. These benchmarks have been adopted by researchers from ByteDance, Baidu, NTU, and more, highlighting current limitations and guiding future development.