PersonaScienceBench introduces a suite of 200 tasks drawn from five benchmarks and spanning four scientific domains. The task objectives and grading remain fixed while simulated researchers, defined by persona profiles, supply context and feedback; agents only perceive researcher preferences through conversation.
Across eighteen model configurations we evaluate task completion, fact‑recovery, and self‑repetition, revealing distinct strengths per metric. A fine‑grained GPT‑6 Astra analysis of interaction traces shows how the same model behaves differently under various researcher personas, with targeted clarification and workflow revisions appearing in the dialogue even when final scores match. Thus PersonaScienceBench assesses both an agent's ability to finish a task and its capability to collaborate with diverse researchers.
Review