Multi-Turn and Agentic AI Workflows for Scientific Investigation

Why this matters

Current benchmarks and systems predominantly evaluate single-turn interactions, but real scientific inquiry requires sustained, multi-step reasoning, iterative hypothesis refinement, and tool-using agents operating over extended horizons. The field's inability to evaluate and build multi-turn scientific AI agents represents a fundamental gap between current capabilities and the complex workflows researchers actually need.

Suggested approaches

• Design multi-turn evaluation benchmarks that simulate realistic scientific investigation sessions, including literature search, hypothesis refinement, experimental design, and peer feedback cycles
• Build and evaluate agentic systems that orchestrate multiple specialized tools (search, code execution, statistical analysis) in long-horizon scientific tasks
• Study how errors, misconduct, and reasoning failures emerge and propagate across extended multi-turn interactions in scientific AI systems

Expected impact

Advancing multi-turn agentic capabilities would enable AI systems to serve as genuine research collaborators rather than single-query tools, supporting complex experimental design, iterative analysis, and sustained scientific reasoning.

A question to explore

I want to investigate how AI systems perform across multi-turn scientific investigation workflows. What does the evidence tell us about the unique challenges of extended, agentic scientific reasoning compared to single-turn tasks, and what would a rigorous benchmark design look like to evaluate multi-turn scientific AI agents?