AHOIS Socratic agents for autonomous discovery: 76.97% MNIST

AHOIS, a multi-agent AI scientist submitted to arXiv on 25 Jun 2026 by Xianrui Zeng and colleagues, embeds Socratic midwifery into closed-loop experimentation and autonomously generated, challenged and revised physical explanations. The system ran on a real multimode-fibre optical platform and proposed a random-interference encoding that it validated through iterative experiments.

How does AHOIS work?

AHOIS pairs a multi-agent control loop with Socratic interrogation: a physics-critic agent interrogates hypotheses through causal questioning, constraint checking, counterexample generation and falsification-criteria formulation, and the system executes closed-loop experiments on hardware. The paper frames AHOIS as an "AI scientist" that constructs hypotheses, stresses them with targeted queries, converts critiques into experimental plans and uses measured outcomes to accept, revise or reject explanations.

The architecture centers on an explicit critic role. That physics-critic agent performs four classes of tests the authors name: causal questioning, constraint checking, counterexample generation and falsification-criteria formulation. The loop runs without prior encoding schemes, classifiers or speckle models; instead AHOIS synthesizes hypotheses and experimental plans then translates them into executable workflows on the optical hardware.

What did AHOIS discover and measure?

AHOIS autonomously proposed and validated a random-interference encoding hypothesis on a high-dimensional multimode-fibre system, discovered task-adaptive sparse-measurement strategies, diagnosed distinct failure modes and translated a published imaging protocol into an executable workflow on a non-original configuration. The discovered encoding produced 16x16 measurements with effective rank 56.9 and yielded classification accuracies of 76.97% on MNIST and 83.17% on Fashion-MNIST.

The system also identified three distinct failure modes during its experiments: encoding instability, fluorescence contamination and detector noise. Ablation experiments reported in the paper show that Socratic interrogation improves multiple experimental attributes, specifically physical consistency, hypothesis completeness, uncertainty calibration and experimental-plan validity. The result set spans both algorithmic discoveries (sparse-measurement strategies) and practical engineering outcomes (porting a protocol to different hardware).

Why it matters

AHOIS moves beyond procedural automation by adding epistemic autonomy: the capacity to construct, challenge and revise physical explanations in response to evidence. That shift matters because it narrows the gap between human-driven hypothesis cycles and an AI that can both propose mechanistic accounts and generate the targeted tests needed to falsify them. For high-dimensional, indirectly observed systems such as multimode fibres, the paper demonstrates a concrete path from workflow automation toward evidence-grounded, self-correcting autonomous discovery.

The combination of hypothesis-level critique and closed-loop experiment execution also lowers the reliance on hand-designed encodings or pre-existing models, enabling the system to find strategies such as task-adaptive sparse measurements that a human operator might miss or take longer to identify.

What to watch

Confirmatory reproduction on other hardware platforms and scaling the approach to new high-dimensional physical systems would be a direct next milestone; success would show the method generalizes beyond the multimode-fibre testbed. The paper’s ablation results offer a measurable signal: improvements in physical consistency, hypothesis completeness and uncertainty calibration when Socratic interrogation is present. Tracking those metrics in independent replications will indicate whether the route from workflow automation to epistemic autonomy holds up.

Authors and provenance: the paper is titled "Socratic agents for autonomous scientific discovery in high-dimensional physical systems," submitted 25 Jun 2026, by Xianrui Zeng, Pengfei Liu, Yirui Zang, Yang Shen, Fei Yu, Chunlei Yu, Minghao Liu and Yang Du. The PDF, HTML and TeX source are available from the arXiv entry.