Title: Principle-Evolvable Scientific Discovery via Uncertainty Minimization

Original: arXiv:2602.06448v2 Announce Type: replace-cross

Abstract: While Large Language Model (LLM)-driven scientific agents have significantly expedited the pace of discovery, their efficiency is often compromised by a rigid reliance on initial priors. Most current methods are confined to a static hypothesis space, which hampers the identification of new phenomena and leads to computational redundancy when foundational theories prove inadequate. To overcome these limitations, we advocate for a paradigm shift from merely searching for hypotheses to actively evolving the core scientific principles themselves. We introduce PiEvo, a framework designed for principle evolution that conceptualizes scientific discovery as Bayesian optimization within a dynamically expanding principle space. PiEvo empowers agents to autonomously update their theoretical perspectives by combining an anomaly-driven augmentation mechanism with Information-Directed Hypothesis Selection utilizing Gaussian Processes. Our evaluations across four distinct benchmarks reveal that PiEvo delivers superior results: it secures an average solution quality between 90.81% and 93.15%, marking a 29.7% to 31.1% gain over existing state-of-the-art methods. Furthermore, by optimizing a compact principle space to reduce sample complexity, PiEvo accelerates convergence steps by 83.3%. The framework also demonstrates consistent robustness across various scientific fields and different LLM backbones. The source code is publicly accessible at \hyperlink{https://github.com/amair-lab/PiEvo}{github.com/amair-lab/PiEvo}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC