Bingchen Zhao, Sara Beery, Oisin Mac Aodha
An LLM-powered framework for autonomous open-ended scientific discovery that performs second-order reflection on its own accumulated findings.
Existing autonomous scientific discovery systems are limited by constrained search spaces or predefined research questions, hindering true open-ended inquiry. They also lack the ability to explicitly synthesize their own accumulated findings to uncover complex, interconnected phenomena.
DiscoPER dynamically generates and executes code to explore datasets without pre-specified research objectives, using an LLM. Every proposed discovery must pass statistical testing for scientific validity. The core innovation is a second-order reasoning mechanism that periodically analyzes prior discoveries as empirical data to identify structural patterns, confounds, and epistemic gaps, actively redirecting exploration. The search space is further expanded by incorporating tool use to process multimodal sources like images.
Evaluated on iNatDisco, a new multimodal ecological knowledge benchmark, DiscoPER recovers 8 of 9 known patterns with a 72.7% hypothesis support rate, outperforming classical causal discovery and LLM-guided baselines. Ablations confirm that DiscoPER scales with more data and validate the benefits of second-order meta-reflection.