Socratic agents for autonomous scientific discovery in high-dimensional physical systems

Traditional autonomous scientific discovery systems are largely procedural, executing fixed workflows defined by human designers. This paper asks whether an artificial intelligence can achieve epistemic autonomy—the ability to construct, challenge, and revise physical explanations based on experimental evidence—by embedding Socratic reasoning into a closed-loop multi-agent architecture.

The authors introduce AHOIS, a multi-agent framework comprising five specialized agents: Boxue (planning and hypothesis generation), Gewu (hardware execution), Mingde (system integrity monitoring), Qiushi (data analysis), and Duzhi (the Socratic physics critic). Unlike monolithic AI systems, AHOIS forces candidate hypotheses through a 'Socratic midwifery' loop. The Duzhi agent interrogates proposed explanations by demanding explicit physical assumptions, causal links, and falsification criteria. The framework was tested on a high-dimensional multimode-fiber optical platform, where it had to interpret complex wave transformations and environmental noise without prior models.

AHOIS successfully moved beyond predefined tasks by autonomously proposing that backscattered speckle patterns could function as a random-interference encoder. The system formulated testable physical criteria, validated the hypothesis through experimental measurements, and successfully classified MNIST and Fashion-MNIST datasets using the discovered encoding. Furthermore, the agent diagnosed hardware-specific failure modes, such as fluorescence contamination and detector noise, and adapted its experimental strategy accordingly. Ablation studies confirmed that the Socratic interrogation process significantly improved the physical consistency, completeness, and uncertainty calibration of the agents' scientific output.

This work provides a blueprint for transitioning from simple workflow automation to true autonomous scientific discovery. By operationalizing Socratic inquiry as a computational principle, the authors demonstrate that AI can move from merely operating instruments to actively reasoning about physical reality, identifying its own knowledge gaps, and designing experiments that discriminate between competing scientific explanations.