Research
Peer-reviewed publications on AI safety, LLM behaviour, and software reliability.
Junjie Li, Fazle Rabbi, Cheng Cheng, Aseem Sangalay, Yuan Tian, Jinqiu Yang
Extended journal study examining whether fine-tuning pre-trained LLMs on vulnerability-fixing commits promotes secure code generation. Applied LoRA and IA3 parameter-efficient fine-tuning on multiple LLMs including CodeLlama across a dataset of 14,622 C/C++ files. Found that larger fine-tuning datasets produce more secure output without degrading correctness — CodeLlama showed a 2% PASS@1 improvement under HumanEval CPP with secure fine-tuning.
Fine Tuning Large Language Model for Secure Code Generation
Junjie Li, Aseem Sangalay, Cheng Cheng, Yuan Tian, Jinqiu Yang
Conference paper demonstrating that fine-tuning GPT-J on real-world vulnerability fixes steers LLM code generation away from insecure patterns. Achieved ~10% increase in vulnerability-free C code output, showing targeted fine-tuning on security-relevant data meaningfully shifts model behaviour without sacrificing generation quality.
Software Defect Prediction Using Abstract Syntax Trees Features and Object-Oriented Metrics
A. Sethi, A. Sangalay, R. Malhotra
Framed software bug prediction as a regression problem and compared LSTM and CNN models trained on Abstract Syntax Tree features and object-oriented code metrics against classical ML baselines. AST-based structural representations outperform flat OO metrics alone for predicting defect-prone modules.