We have published a series of LLM-based static analysis works that form the cornerstone of RepoAudit. Meanwhile, we continuously collect and categorize the latest research. The research projects and paper list below are provided for reference to researchers and practitioners in the field.
CodeLLMPaper: A Continuously Updated Collection of CodeLLM Papers
CodeLLMPaper is a curated collection of the latest research on LLM-for-Code published in top-tier venues in software engineering, programming languages, security, NLP, and machine learning. The collected research works cover diverse coding tasks, foundational principles of code models, empirical studies, and surveys.
BugScope: Learn to Find Bugs Like Human
BugScope is an intelligent bug detection agent that learns to identify diverse bugs from examples. It outperforms existing industrial tools like Cursor BugBot and CodeRabbit, detecting twice as many bugs while maintaining high precision.
An LLM Agent for Functional Bug Detection in Network Protocols
An LLM agent for statically detecting functional bugs in network protocol implementations against the natural language specifications in the RFC documents
Validating Network Protocol Parsers with Traceable RFC Document Interpretation
ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA 2025)
An LLM-driven test case generation technique that uncovers the inconsistencies between network protocol parsers and the packet format specifications in RFC documents.
LLMDFA: Analyzing Dataflow in Code with Large Language Models
The Thirty-Eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024)
An LLM-powered summary-based data-flow analysis framework, achieving comparable and even superior precision and recall to state-of-the-art symbolic static analysis tools.
LLMSAN: Sanitizing Large Language Models in Bug Detection with Data-Flow
The 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings 2024)
A sanitization technique that mitigates LLM hallucinations in bug detection, validating data-flow paths as verifiable chain-of-thoughts via divide-and-conquer.