Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control

Researchers developed a self-evolving scientific agent powered by large language models that autonomously discovers interpretable control policies for complex physical systems. The system successfully solved an underactuated fluid-dynamics problem (dogfish swimmer navigation) by iteratively testing strategies, diagnosing behaviors, and refining source code—achieving generalization to unseen targets without retraining.

This research represents a meaningful intersection of AI autonomy and scientific discovery, demonstrating that LLM-driven agents can move beyond black-box optimization toward explainable, physically-grounded solutions. The key innovation lies in the agent's ability to maintain interpretability throughout discovery—rather than optimizing neural network weights, it generates and refines human-readable control code based on empirical evidence from simulations. This approach bridges a critical gap between deep reinforcement learning's effectiveness and science's requirement for traceable reasoning.

The breakthrough emerges from a long-standing challenge in AI-assisted science: how to leverage LLM reasoning without sacrificing the interpretability that domain experts demand. Previous approaches either sacrifice explainability for performance or fail to discover sophisticated policies. By closing this loop—testing policies, observing outcomes, generating diagnostic insights, and refining code iteratively—the agent produced a controller encoding principles like traveling-wave propulsion and adaptive cadence relief that human engineers might recognize.

For the AI research community, this demonstrates LLMs' emerging value in automated scientific reasoning beyond language tasks. The generalization to unseen targets and dynamic trajectories without modification suggests the agent discovered robust underlying principles rather than overfitting to training scenarios. For applied domains like robotics, autonomous systems, and fluid dynamics, this workflow could accelerate controller design by automating hypothesis generation and testing cycles that typically require human expertise and months of iteration.

• →LLM-driven agents can autonomously discover interpretable control policies by iterating code generation rather than weight optimization
• →The framework successfully solved a complex underactuated fluid-dynamics problem and generalized to unseen targets without retraining
• →Synthesized policies remain mathematically readable and scientifically auditable, maintaining full transparency in the discovery process
• →This approach bridges AI performance with scientific rigor, addressing the interpretability demand that pure deep learning cannot satisfy
• →The workflow could accelerate scientific discovery and controller design across robotics, autonomous systems, and physical simulations