Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Criteria in Automated Scientific Discovery

Forum topic · 小凯 · 2026-10-19

Summary

This paper by Tianyu Wu (arXiv:2610.22445) examines whether passing tests in automated scientific discovery actually establishes the intended claim. The authors develop four executable operations for changing physical descriptions—identifying redundancies (L1), coarse-graining dynamics (L2), transferring structure (L3), and inferring missing quantities from conservation imbalances (L4)—and introduce three evaluation criteria: null acceptance rate (NAR), positive acceptance rate (PAR), and consequence agreement (CAR). A key finding is that a degenerate reduction mapping every state to zero perfectly predicts its own future, passing closure tests on 96% of 129 public chaotic systems while retaining no dynamical variation; only an independent observable-preservation test rejects it. On controlled benchmarks, combined criteria accept 24/24 genuine reductions and reject 72/72 nulls. In interpretability settings, linear CKA is insensitive to random rotations of proposed correspondences in language and ESM-2 protein models, while localized interventions separate genuine mappings from nulls; in sparse autoencoders, enrichment tests reject all 16 nulls but independent ablation supports only 45% of enriched features versus 0.5% of matched controls. The results expose two acceptance gaps, showing reliable automated discovery requires assessing both what tests accept and what consequences they establish.

Overview

Field: Machine Learning
Author: Tianyu Wu
Published: 2026-10-19
arXiv: 2610.22445

Abstract (translated)

Automated discovery systems propose scientific explanations and test which to accept. Does passing a test establish the intended claim? The paper develops four executable operations for changing physical descriptions: identifying redundancies (L1), coarse-graining dynamics (L2), transferring structure (L3), and inferring missing quantities from conservation imbalances (L4). These operations and three interpretability settings are evaluated through null acceptance (NAR), positive acceptance (PAR), and consequence agreement (CAR).

Key Findings

The degenerate zero reduction: A reduction that maps every state to zero predicts its own future perfectly. It passes closure checks on 96% of 129 public chaotic systems despite retaining no dynamical variation; an independent observable-preservation test rejects it throughout.
Controlled benchmarks: The combined criteria accept 24/24 genuine reductions and reject 72/72 nulls.
Other operations: Distinguish symmetries from redundancies, validate structure transfers, and reject spurious conservation imbalances.
Interpretability (CKA): In language models and ESM-2 protein models, linear CKA remains unchanged when a proposed correspondence is randomly rotated, whereas localized interventions distinguish genuine mappings from tested nulls.
Sparse autoencoders: Enrichment tests reject all 16 tested nulls, yet independent ablation supports only 45% of evaluable enriched features, compared with 0.5% of matched controls.

Conclusion

The results reveal two acceptance gaps: descriptive tests can accept scientifically empty candidates, and accepted candidates can lack independently verified consequences. Reliable automated discovery requires evaluating both what a test accepts and the scientific consequences it establishes.

> Code and data: available via the arXiv page.