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

The evolution of scientific automation has transitioned from simple data collection to complex reasoning systems. Early efforts focused on optimizing parameters via Bayesian methods or reinforcement learning, but these lacked deep physical understanding. Recent advances incorporate large language models and multi-agent architectures, enabling hypothesis generation, literature synthesis, and experimental control. Platforms like EvoMaster and AutoMOOSE have demonstrated partial autonomy, yet they rely heavily on predefined models or tasks, limiting adaptability. High-dimensional systems such as multimode fibers and scattering media pose unique challenges due to their complex wave transformations, environmental sensitivity, and indirect measurements. These systems require autonomous reasoning about physical encoding, noise sources, and measurement strategies, which current methods cannot fully address. Therefore, developing a framework that combines physical modeling, hypothesis testing, and adaptive experimentation remains a critical goal for autonomous scientific discovery.

The main challenge is enabling AI systems to autonomously generate and validate physical hypotheses in complex, high-dimensional environments without relying on pre-encoded models. Existing approaches often depend on fixed workflows, limiting flexibility and interpretability. In multimode fiber systems, the difficulty lies in distinguishing meaningful physical signals from artifacts caused by environmental drift, noise, or system instability. Achieving epistemic autonomy requires the system to actively interrogate its assumptions, identify failure modes, and adapt measurement strategies dynamically. This entails integrating causal reasoning, uncertainty calibration, and hypothesis falsification into a cohesive loop, which is currently lacking in most AI-driven experimental platforms.

This work introduces several innovations: 1) A multi-agent architecture that separates hypothesis generation, physical validation, and system monitoring, facilitating modularity and interpretability; 2) The incorporation of Socratic questioning as a core mechanism for hypothesis refinement, exposing hidden assumptions and causal links; 3) A novel, encoder-agnostic physical measurement strategy based on structured inquiry, enabling autonomous discovery of information encoding methods; 4) An adaptive sparse measurement protocol that reallocates sampling efforts based on real-time evidence, significantly improving efficiency. These innovations collectively enable a system that not only automates experimental procedures but also actively constructs, challenges, and revises its physical understanding in complex environments.

• �� Hypothesis generator analyzes optical platform, proposing candidate explanations for observed speckle patterns.
• �� Physics critic employs causal inference, singular value decomposition, and modal analysis to verify the physical richness and consistency of proposed hypotheses.
• �� Socratic interrogator challenges hypotheses through clarification, constraint checking, counterexample generation, and causal probing, exposing assumptions and missing links.
• �� Measurement strategies are iteratively refined: initial coarse scans identify regions of interest, followed by adaptive, task-specific sparse sampling.
• �� Hardware agents execute optical modulation, data acquisition, and system monitoring, providing real-time feedback.
• �� The reasoning loop integrates measurement outcomes, updates hypotheses, and guides subsequent experiments, ensuring continuous model refinement.

Experiments involved autonomous hypothesis generation on a multimode fiber platform, testing the encoding hypothesis via structured backscattering measurements. Data included classification tasks on MNIST and Fashion-MNIST, with metrics such as accuracy, effective rank, and information entropy. Baselines included fixed scanning and predefined encoding schemes. Ablation studies assessed the impact of Socratic questioning on physical consistency and experiment validity. Multiple scenarios tested adaptive sparse sampling, environmental robustness, and failure mode diagnosis, with hyperparameters tuned for optimal performance. The system operated under real-time constraints, demonstrating high-speed classification and task adaptation across different imaging modalities.

The system successfully discovered a physical encoding mechanism based on backscattered speckle, achieving a 16×16 measurement matrix with an effective rank of 56.9, enabling classification accuracies of 76.97% (MNIST) and 83.17% (Fashion-MNIST). It diagnosed multiple failure modes, improving physical model fidelity. Adaptive sparse sampling increased efficiency, reaching real-time classification at over 70 fps, with significant reductions in measurement effort compared to full scans. Ablation experiments confirmed that Socratic questioning enhanced hypothesis robustness, physical consistency, and experimental validity. These results demonstrate the system’s ability to autonomously generate, validate, and optimize physical hypotheses in complex environments.

This framework can be applied to high-dimensional optical imaging, quantum sensing, biological microscopy, and materials science, where autonomous understanding of complex physical interactions is crucial. It reduces reliance on human expertise, accelerates discovery, and improves robustness against environmental variability. The adaptive measurement strategies enable real-time applications such as medical diagnostics, industrial inspection, and environmental monitoring. Long-term, this approach could lead to fully autonomous scientific laboratories capable of exploring uncharted physical phenomena with minimal human intervention.

The current system depends on specific physical models of optical platforms, limiting immediate generalization. Its robustness under high noise or environmental shifts needs further validation. Computational complexity during multi-agent reasoning and iterative questioning remains high, constraining real-time scalability. Future work should focus on improving generalization, reducing computational costs, and extending to other physical systems such as quantum or biological environments.