Reproduce “AICrypto: Evaluating Cryptography Capabilities of Large Language Models” 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
Yu Wang, Yijian Liu, Liheng Ji, Han Luo, Wenjie Li, Xiaofei Zhou, Chiyun Feng, Puji Wang and 6 more
Researchers introduced AICrypto, a comprehensive benchmark to evaluate Large Language Models' (LLMs) cryptographic capabilities across knowledge, practical exploitation, and formal reasoning. The benchmark revealed that top LLMs excel in factual knowledge (e.g., o3 achieved 97.8% accuracy on MCQs) and routine proof generation, but significantly lag human experts in complex Capture-the-Flag (CTF) challenges, struggling with precise mathematical computation and deep abstract reasoning.
Key insights
- Leading LLMs exhibit a tiered performance, demonstrating strong foundational cryptographic knowledge and proficiency in routine proof generation.
- A significant performance gap exists between even the best LLMs and human experts in practical cryptographic problem-solving (CTF challenges) and advanced mathematical reasoning.
- LLMs consistently struggle with precise numerical computation and often rely on pattern matching over deep analytical or abstract mathematical comprehension, leading to logical gaps in proofs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.