Toward Evaluation Frameworks for Multi-Agent Scientific AI Systems

TL;DR

This paper discusses the challenges and strategies for benchmarking multi-agent scientific AI systems, emphasizing realistic evaluation methods and the importance of multi-turn interactions.

Contribution

It introduces approaches for constructing contamination-resistant problems and scalable task families, and presents an initial dataset for testing out-of-sample performance.

Findings

Constructed a dataset of novel research ideas for evaluation

Identified key challenges in benchmarking scientific AI systems

Gathered insights from quantum science researchers on AI interaction expectations

Abstract

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.