Toward Evaluation Frameworks for Multi-Agent Scientific AI Systems
This addresses evaluation frameworks for researchers and engineers developing AI systems in scientific domains, but it appears incremental as it builds on existing benchmarking challenges with specific domain focus.
The paper tackles the challenge of benchmarking multi-agent scientific AI systems by analyzing issues like distinguishing reasoning from retrieval, data contamination, and lack of ground truth for novel problems, and proposes strategies such as contamination-resistant tasks and multi-turn interactions. As a feasibility test, it demonstrates constructing a dataset of novel research ideas and discusses insights from interviews with quantum science researchers on AI interaction expectations.
We analyze the challenges of benchmarking scientific (multi)-agentic systems, including the difficulty of distinguishing reasoning from retrieval, the risks of data/model contamination, the lack of reliable ground truth for novel research problems, the complications introduced by tool use, and the replication challenges due to the continuously changing/updating knowledge base. We discuss strategies for constructing contamination-resistant problems, generating scalable families of tasks, and the need for evaluating systems through multi-turn interactions that better reflect real scientific practice. As an early feasibility test, we demonstrate how to construct a dataset of novel research ideas to test the out-of-sample performance of our system. We also discuss the results of interviews with several researchers and engineers working in quantum science. Through those interviews, we examine how scientists expect to interact with AI systems and how these expectations should shape evaluation methods.