Principle-Evolvable Scientific Discovery via Uncertainty Minimization

Large Language Model (LLM)-based scientific agents have accelerated scientific discovery, yet they often suffer from significant inefficiencies due to adherence to fixed initial priors. Existing approaches predominantly operate within a static hypothesis space, which restricts the discovery of novel phenomena, resulting in computational waste when baseline theories fail. To address this, we propose shifting the focus from searching hypotheses to evolving the underlying scientific principles. We present $\textbf{PiEvo}$, a principle-evolvable framework that treats scientific discovery as Bayesian optimization over an expanding principle space. By integrating Information-Directed Hypothesis Selection via Gaussian Process and an anomaly-driven augmentation mechanism, PiEvo enables agents to autonomously refine their theoretical worldview. Evaluation across four benchmarks demonstrates that PiEvo (1) achieves an average solution quality of up to 90.81\%$\sim$93.15\%, representing a 29.7\%$\sim$31.1\% improvement over the state-of-the-art, (2) attains an 83.3\% speedup in convergence step via significantly reduced sample complexity by optimizing the compact principle space, and (3) maintains robust performance across diverse scientific domains and LLM backbones. Code is publicly available at github.com/amair-lab/PiEvo.

Machine learning systems are increasingly being used to help scientists search for new materials, molecules, and physical mechanisms. However, most current AI research agents work within a fixed set of starting assumptions, so when those assumptions are wrong, they can waste time exploring the wrong ideas and may miss genuinely new discoveries. Our work introduces a new framework called PiEvo that allows an AI scientist to revise its guiding scientific principles as it gathers evidence. Instead of only asking “Which hypothesis should I test next?”, the system also asks “Do my current scientific beliefs still explain what I am observing?” When surprising results appear, the system treats them as clues that its current worldview may be incomplete and proposes new principles to better explain the data. Across four discovery benchmarks in chemistry, biology, physics, and materials science, this approach found better solutions than previous methods while using far fewer experimental steps. In one case study, it also helped uncover a plausible physical mechanism behind optical chirality in nanohelix. This suggests that future AI scientists may become more useful not just by searching faster, but by learning to rethink their own assumptions.