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

🕒 Published (v1): 2026-06-25 08:01 UTC · Source: Arxiv · link

Why this paper was selected

Pengfei Liu; Socratic closed-loop agents autonomously discover scientific laws in physical systems

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AHOIS is a five-agent framework that embeds Socratic interrogation into closed-loop physical experimentation to achieve epistemic autonomy in scientific discovery. A dedicated physics-critic agent (Duzhi) challenges candidate hypotheses through structured causal questioning, counterexample generation, and falsification-criteria formulation before any experiment is executed. On a real multimode-fibre optical platform, AHOIS autonomously discovered a random-interference encoding scheme, adaptive sparse-scanning strategies, and distinct failure modes without any pre-specified objectives.

Problem¶

Existing AI-for-science systems remain procedurally autonomous: objectives, action spaces, and success criteria are fixed by human designers. They can operate instruments and optimize parameters but lack epistemic autonomy—the capacity to construct, challenge, and revise physical explanations from ambiguous experimental evidence. Generic self-reflection evaluates linguistic plausibility rather than the causal and evidential structure of scientific claims, leaving implicit assumptions and brittle reasoning unchecked.

Method¶

AHOIS separates five functional roles across dedicated agents: - Boxue (strategy/hypothesis): decomposes a high-level goal into candidate hypotheses and machine-readable experimental plans. - Gewu (hardware abstraction): translates plans into executable routines for DMD pattern generation, laser operation, and camera triggering. - Mingde (system integrity): audits the live data stream, checks exposure, validates dark frames, and quantifies speckle repeatability. - Qiushi (modelling/analysis): implements reconstruction models, hyperparameter searches, and quantitative analyses. - Duzhi (Socratic physics critic): cannot generate plans; constrained to interrogate Boxue's reasoning by exposing hidden assumptions, unsupported causal links, missing physical constraints, plausible-but-wrong interpretations, and requiring explicit falsification criteria.

The epistemic loop is the defining feature: Boxue's candidate hypothesis is iterated with Duzhi until it is structured as a scientific state comprising mechanism, physical assumptions, competing interpretations, predicted outcomes, discriminating experiments, uncertainty estimates, and rejection conditions. Only then does Gewu convert it into hardware commands. A human supervisor gates unsafe or hardware-incompatible actions but does not substitute for the agent's planning loop. The testbed is a real multimode-fibre optical platform (laser → DMD → fibre → sample → detectors) combining coherent and fluorescence modalities.

Key Contributions¶

Socratic midwifery as a computational principle: operationalizing Socratic interrogation—clarification, physical-constraint checking, causal probing, counterexample generation, falsification-criteria formulation—as a first-class component of an agentic scientific loop.
AHOIS architecture: five-agent modular design with a shared scientific state that decouples intellectual, physical, and evaluative roles while keeping them causally coupled.
Autonomous hypothesis discovery: from an open-ended prompt with no encoding scheme specified, AHOIS proposed and experimentally validated a backscattered random-interference encoding hypothesis for multimode-fibre imaging.
Task-adaptive sparse measurement: autonomous discovery of a coarse-to-fine ROI-based scanning policy that achieves video-rate imaging (up to 135 fps) versus 0.47 fps for full scanning.
Failure-mode diagnosis: autonomous identification of three distinct failure modes—encoding instability, fluorescence contamination, detector noise—and corresponding computational interventions.
Protocol transfer: translation of a published reconstruction protocol to a non-original hardware configuration without human re-specification.
Controlled ablations: quantitative isolation of Socratic interrogation's contribution to physical consistency, hypothesis completeness, uncertainty calibration, and experimental-plan validity.

Results¶

Backscattered speckle encoder: \(16 \times 16\) measurements with effective rank 56.9 (SVD-assessed), indicating high-dimensional feature occupancy.
Classification accuracy on MNIST: 76.97%; on Fashion-MNIST: 83.17% using a compact classifier trained directly on \(16 \times 16\) encoder matrices (no reconstructed image intermediate).
Adaptive sparse scanning: background at \(1/12\) density + ROI at \(1/8\) density achieved 48.50 fps with successful ROI and content recognition; full scanning yielded 0.47 fps—a \(>100\times\) speed gap.
Highest sparse frame rate: 135.12 fps (both background and ROI at \(1/16\)), though insufficient for recognition.
Ablation baselines: removing Duzhi (Socratic critic) degraded physical consistency, hypothesis completeness, uncertainty calibration, and plan validity (quantitative scores reported in ablation section; specific numbers in truncated portion of paper).

Limitations¶

Evaluated on a single physical domain (multimode-fibre optics); generalizability to other high-dimensional physical systems (e.g., materials synthesis, particle physics) is asserted but not demonstrated.
Human supervision remains in the loop at predefined decision points for safety gating; the system is not fully hands-off.
Classification accuracies (76.97% / 83.17%) are moderate—well below state-of-the-art on MNIST/Fashion-MNIST—reflecting the compressed, indirect measurement scheme rather than any classifier limitation.
The Socratic interrogation loop's effectiveness depends on Duzhi's physical knowledge being adequate to the domain; failure modes in poorly-represented physics subfields are not characterized.
Ablation quantification appears in portions of the paper not fully provided; the exact magnitude of Socratic contribution to each metric is not fully visible in the provided text.

Relevance to Agentic AI / LLM Agents¶

AHOIS is a concrete instantiation of multi-agent LLM systems operating in closed-loop with real physical instruments, going beyond simulation or text-only benchmarks. The Socratic critic architecture directly addresses a known failure mode of LLM agents—confident but physically inconsistent reasoning—by institutionalizing adversarial self-interrogation as a structural component rather than a prompt-level afterthought. The shared scientific state object (hypothesis + assumptions + predicted outcomes + rejection conditions) is a transferable pattern for any agentic system that must maintain revisable world models under uncertainty. For researchers tracking agentic AI, this work demonstrates how role specialization and structured epistemic challenge loops can convert a generative LLM from a plausible-text producer into a system that generates falsifiable, testable scientific claims grounded in causal physical models.