Reproduce “Natural Language based Specification and Verification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaorui Li, Chengyu Song
Researchers at the University of California, Riverside introduce NLForge, a compositional framework that leverages large language models (LLMs) to perform memory-safety verification in C/C++ programs using natural-language summaries for interprocedural reasoning. This approach improved total true positives by 2.5% and reduced false negatives by 28% across benchmarks compared to non-compositional LLM baselines, particularly excelling in complex data structure and control-flow categories.
Key insights
- LLMs can function as high-coverage, high-recall verifiers for bug-finding, especially where traditional tools report 'Unknown', but are not yet sound replacements for formal verification due to notable false positives and negatives.
- Free-form natural language is sufficiently expressive to enable bottom-up compositional verification, effectively propagating complex memory safety facts across function boundaries.
- Compositional reasoning with natural-language summaries provides substantial improvements in recall and reduction of false negatives, particularly for smaller LLMs and challenging real-world programs with complex control flow.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.