Minimal Epistemic Closed-Loop Agents for Scientific Discovery: Constraints, Information Gain, and Discriminative Testing in a Toy Genomics Environment

Track: Track 1: Technical Foundations for a Post-AGI World

Keywords: epistemic planning, closed-loop scientific agents, Bayesian experimental design, expected information gain, falsification, active learning, toy genomics, uncertainty reduction

TL;DR: We study a minimal closed-loop scientific agent that enforces feasibility, selects experiments by information gain, and uses discriminative tests to refute wrong hypotheses faster, demonstrated in a toy genomics simulator.

Abstract: Closed-loop “AI scientist” systems can generate hypotheses, call tools, and iterate, but many implicitly optimize for plausibility or predicted success rather than explicitly reducing uncertainty about an underlying mechanism. We study a minimal design for epistemic closed-loop agents that (i) enforce feasibility via hard constraints during proposal, (ii) select experiments by Expected Information Gain (EIG), and (iii) accelerate refutation by generating discriminative “Achilles” tests that maximize disagreement among hypotheses. We evaluate only algorithmic behavior in a small, reproducible toy genomics simulator with discrete latent hypotheses and noisy outcomes (not biological validation). Across random seeds, EIG-based selection reduces posterior entropy faster than success-seeking baselines, and Achilles-style testing reduces experiments-to-refute incorrect leading hypotheses under an operational refutation criterion. Finally, we use a minimal taxonomy for negative outcomes (infeasible, inaccessible, null, execution-failure) to avoid conflating distinct failure semantics during belief updates.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Manish_Sai_Kota1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 37