Notes

• ZKP programming is hard and error-prone; LLMs lack systematic evaluation.
• ZK-Eval benchmark covers language, gadget, and end-to-end generation.
• LLMs handle syntax well but fail on gadgets and semantics; Circom is harder.
• ZK-Coder uses sketching, retrieval, and repair to improve reliability.
• Success rates on Circom/Noir rise from <30% to 80–90%.
• Repair loop is crucial; challenges remain in efficiency and data scarcity.

• ZKP 编程困难且易错，LLM 在该领域缺乏系统评估。
• ZK-Eval 基准涵盖语言知识、gadget 能力和端到端生成。
• LLM 语法掌握好，但 gadget 与语义正确性差，Circom 更难。
• ZK-Coder 结合约束 sketch、检索和交互修复提升生成质量。
• 在 Circom/Noir 上成功率从 <30% 提升至 80–90%。
• 修复循环最关键，挑战在电路效率与数据稀缺。