AI-for-Science Low-code Platform with Bayesian Adversarial Multi-Agent Framework

Multi-agent systems leveraging Large Language Models (LLMs) show immense potential for solving complex scientific problems. However, their reliability is undermined by the probabilistic nature of LLMs, which can produce hallucinations in both generated code and its corresponding test cases. In a multi-agent architecture, these errors can propagate and compound, leading to flawed final outputs. To overcome these core limitations, we introduce a novel Bayesian Adversarial Multi-agent Framework for AI for Science (AI4S). Delivered as a Low-code Platform (LCP), our framework enhances the coding capability for scientific tasks across a wide range of base models, from 1.7B open-source LLMs to up-to-date commercial ones. Our framework employs three agents in a recursive loop that adversarially co-optimizes the generated solutions, the test cases used for evaluation, and the prompts driving generation. This process is governed by a non-LLM-based Bayesian updating rule, which systematically reduces evaluation uncertainty and mitigates the system's dependence on any single LLM's reliability. Furthermore, the LCP empowers domain experts by translating high-level natural language prompts into executable, domain-specific requirements, eliminating the need for intricate prompt engineering. Extensive experiments confirm that our framework generates robust solutions while effectively minimizing error propagation. On a complex, cross-disciplinary Earth Science benchmark, our platform demonstrates superior reliability and outperforms state-of-the-art models, where a 32B opensource model can beat the performance of a 235B model in the ScienceCode benchmark with our framework.