A multi-agent system for automated MCNP input review using open-weight LLMs

Large language models (LLMs) have demonstrated promising performance in code generation, but their reliability remains limited for domain-specific simulation codes, such as Monte Carlo N-Particle (MCNP), which requires a strict card-based syntax and often iterative debugging. While commercial LLMs can assist in debugging such scripts, their effectiveness is limited due to insufficient domain knowledge of MCNP. In nuclear engineering, the use of commercial LLMs is constrained by security requirements, motivating local review systems for MCNP input scripts. Although open-weight LLMs are a practical choice, their performance remains lower than that of commercial LLMs, requiring mechanisms to compensate for this limitation. In this paper, we propose a multi-agent system for MCNP input script review. The system consists of Supervisor, Fixer, and Validator agents and supporting tools. A benchmark dataset of 360 MCNP input scripts representing nuclear engineering problems was constructed for evaluation. Results show that the proposed system achieves an execution success rate of 75.83% and a functional success rate of 47.22%, which are higher than those of the single-agent baseline. These results demonstrate the effectiveness of multi-agent orchestration for consistent and improved validation and correction of MCNP code in local environments, while framing the gap between execution and physical correctness.