Scientific inquiry consists of iterative cycles of variable manipulation, observation collection, and reasoning over those observations to uncover causal relationships in complex systems. Large language models (LLMs) excel at components such as inductive reasoning, hypothesis generation, and data analysis, yet prior work typically evaluates these abilities in isolation, leaving the integration across a full inquiry cycle unexamined. To address this gap, we propose a unified interactive evaluation framework that extends the Countdown task into a mathematical environment where an agent probes the system by adjusting variables, gathering observations, and inferring the relationships among them. This setup enables analysis of exploration strategies, causal discovery, and hypothesis refinement throughout the inquiry process. Experiments across seven LLMs and multiple prompting strategies reveal limited inherent inquiry capability and low success rates in uncovering the underlying expressions. However, providing structured guidance inspired by educational scaffolding markedly improves performance. The findings highlight current limitations of LLMs in strategic exploration and iterative hypothesis refinement, while demonstrating that simple guidance techniques developed for human learners can substantially enhance inquiry behavior.
Review