MPC-Patch-Bench: Security-Aware LLM Code Patch for Multi-Party Computation

The article introduces MPC-Patch-Bench, a novel benchmark designed for evaluating Large Language Model (LLM) code repair specifically for Secure Multi-Party Computation (MPC) software. It features a Data Curation Framework that uses a domain-specific agent to filter pull requests and create verified instances, along with an MPC Verifier that ensures security and numerical fidelity through dynamic testing and static analysis. This benchmark addresses the inadequacies of existing LLM evaluation methods in the context of MPC, highlighting the challenges LLMs face in generating cryptographically safe code, with the best-performing LLM achieving only a 22.9% functional resolution rate.