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

A self-evolving scientific agent driven by large language models has produced a generalizable fluid-control policy for an underactuated robotic swimmer, according to a preprint submitted to arXiv in 2026 [1]. The workflow automates controller design while preserving a fully traceable chain of physical reasoning [1]. The framework, described in a paper submitted on 7 June 2026, targets a long-standing tension in scientific machine learning: deep reinforcement learning can optimize control policies, but it rarely yields the interpretable, physically grounded reasoning that scientific discovery demands [1][2]. The authors instead deploy a self-evolving agent that writes and revises source code rather than adjusting neural-network weights [2]. Candidate strategies are tested inside physical simulations, and the agent diagnoses dynamic behaviors from multimodal evidence before translating observations into code refinements [2]. The testbed is a highly nonlinear fluid-structure interaction problem involving an underactuated, two-joint dogfish swimmer that must reach spatial targets using only joint angular accelerations [1][2]. The initial propulsive seed policy exhibited a one-sided steering bias [2]. From that starting point, the agent autonomously discovered and refined a unified controller that captured all canonical targets [2]. Without any retraining or target-specific branching, the synthesized policy generalized to unseen static targets and dynamically curved pursuit trajectories [1][2]. An auditable evolve log showed that the emergent control architecture rests on traveling-wave propulsion, body-frame target guidance, yaw-rate feedback, signed mean-tail curvature, and adaptive cadence relief [2]. The resulting policy is mathematically readable and the discovery process remains fully traceable [2]. Large language models form the engine of the agent's reasoning loop [2][11]. LLMs are neural networks trained on vast text corpora for tasks including generation and analysis, though biased training data can affect reliability [11]. The preprint appears on arXiv, an open-access repository that hosts e-prints across physics, computer science, and related fields [9]. As of late 2024, the platform was receiving about 24,000 submissions per month [9]. The work was posted under the Computer Science > Artificial Intelligence category [1].