New Tools developed to Check Artificial Intelligence Agents in Scientific Code Writing

The use of artificial intelligence in scientific calculations is rapidly common, but it creates a big challenge to control the accuracy of the code it produces. A new research released in arXiv offers a practical solution to this problem: Rules to Tools (R2T) makes written requirements automatically controlled with system, equation and limit conditions.

The system works as follows:

• Scientific coding agents receive equation, limit conditions and output requirements in writing
• R2T creates** control mechanisms in accordance with these rules and shares with SciCode repair groups
• Each task group receives written controls, starting code, model and budget, offers a executable application in return
• In tests on two cohort, full repair success was achieved in the version of the text using 30 tasks in 26, while in other version 29

The main contribution of the research is the ability to convert abstract rules into concrete and can work immediately. The scientific code written by artificial intelligence can be automatically compared to the requirements after being sent. The system saves time in scientific researches, especially for multiple models and parameters.

This development is seen as an important step** to improve the reliability of the sociopay intelligence agents in scientific calculations. Similar methods in the future can also be used in code verification processes in other areas.