acceptodds
Under review as a conference paper at ICLR 2027

Epistemic Control for Open-world AI Scientists

Abstract

AI-scientist systems increasingly automate hypothesis generation, experimentation, and analysis. Yet predictive success alone does not determine whether a scientific explanation should be retained, revised, or replaced. We study epistemic control: the decisions connecting evidence acquisition to accepted explanations and scientific claims. We organise the problem around representation revision, abductive hypothesis generation, and the control of inquiry and belief revision, with six operational criteria for their coordination. We investigate selected criteria using a literature-conditioned discovery prototype and separate controlled policy studies. The prototype acquires variables from a predefined catalogue and generates expressions within a fixed executable language; the policy studies control candidates, evidence access, and revision opportunities. The results show that evidence-selection and revision policies affect inquiry efficiency and accepted conclusions, that a genuine cause can look redundant for prediction, and that recovering the right variables need not recover the generating mechanism. These findings identify limitations of predictive criteria for scientific decisions and provide controlled tests of components of epistemic control. Autonomous diagnosis and revision of an inadequate scientific representation remains the broader challenge.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI Scientist, Causal Discovery, Epistemic Control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_noM8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes six operational criteria for managing evidence acquisition and explanation revision in AI scientists. It implements selected criteria through a literature-guided variable-acquisition and symbolic-regression prototype, alongside separate controlled policy studies. The evaluation primarily uses author-designed synthetic tasks and shows that predictive accuracy alone does not guarantee mechanism recovery or appropriate claim revision.

Critical strengths
  1. The distinction between prediction, mechanism recovery, and claim revision provides a useful perspective for evaluating scientific agents.
  2. Several paired comparisons isolate specific policy effects, and the paper transparently reports limitations and a failed acceptance criterion.
Critical weaknesses
  1. The prototype acquires variables from a predefined catalogue and operates within a fixed domain-specific language, while the ordered-program condition is enabled externally. These experiments do not demonstrate that the system can autonomously recognize when its representation is inadequate and revise it accordingly, which is central to the paper's motivation.
  2. Evidence acquisition and formula discovery are evaluated in the prototype, whereas reopening and claim withdrawal are tested separately using supplied candidates and evidence. Consequently, the results do not establish that these capabilities can operate coherently together throughout an autonomous scientific discovery process.
Questions for authors
  1. What is the primary methodological contribution that distinguishes the proposed framework from a combination of task-specific prompting, conventional symbolic fitting, and explicit claim-tracking rules? How does the framework advance beyond organizing established ideas from experimental design, model criticism, and belief revision?

  2. In the auxiliary-variable study, the baseline optimizes predictive performance within the initial operating range, whereas the proposed approach prioritizes mechanistic relevance. Could the authors clarify how the comparison controls for these differences in scientific objectives?

  3. In the component-revision study, the baseline can only add components, whereas the treatment can also remove them, and two tasks require removal. How much of the observed performance difference can be attributed to the broader action space rather than to the quality of the revision policy? Could the authors provide a comparison with matched action spaces?

Reviewer_EZkVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces a framework for epistemic control in open-world scientific discovery, focusing on representation revision, hypothesis generation, evidence acquisition, and belief revision. It evaluates selected components through a literature-guided discovery prototype and controlled synthetic experiments. It shows that predictive accuracy alone may not lead to correct scientific explanations.

Critical strengths

The paper provides an interesting and well-motivated discussion of an important problem in AI-driven scientific discovery. I appreciate its emphasis on distinguishing predictive performance from scientific understanding, particularly the need for agents to critically evaluate evidence and revise previously accepted hypotheses.

Critical weaknesses
  1. The paper has insufficient connection to existing AI-scientist literature. It presents an interesting conceptual framework, but I find its connection to existing AI-scientist methods insufficient. Several prior works have already studied similar problems, including hypothesis falsification and evidence validation (POPPER, Huang et al., 2025), experiment selection and mechanistic model revision (Model Discovery Agent, Murphy, 2026), and iterative experimental design and model discovery (BoxingGym, Gandhi et al., 2025). It is unclear what the proposed epistemic-control framework contributes beyond these existing methods, especially since the paper mainly evaluates isolated decision policies rather than a unified discovery system.

  2. The experiments do not convincingly validate the proposed framework. For example, Figure 2 compares policies with different scientific objectives, making the improvement somewhat expected. Figure 4 shows that broader hypothesis search and joint variable acquisition can improve performance, but these results are not particularly surprising and do not demonstrate open-world representation revision. I would like to see stronger comparisons against existing scientific discovery methods and evaluations showing that the proposed framework offers capabilities beyond standard hypothesis testing and experimental design.

Questions for authors
  1. How does the proposed framework differ from existing AI-scientist approaches (mentioned above), which already address hypothesis falsification, experimental design, and model revision?

  2. In Figure 2, the two policies optimize different objectives, making the improvement somewhat expected. How does this comparison demonstrate the effectiveness of epistemic control rather than simply the consequences of different instructions?

  3. Could the authors compare against existing experimental-design or scientific-discovery methods, rather than primarily using a prediction-guided policy?

  4. Since the discovery prototype and epistemic-control policies are evaluated separately, how can we assess whether the proposed framework improves an actual AI-scientist system when these components are integrated? Like make it end-to-end.

Reviewer_8qqtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an epistemic-control framework for open-world discovery. It claims that predictive accuracy alone is insufficient for discovery. Representation revision, abductive hypothesis generation and epistemic control are three key abilities. The authors designs a literature-conditioned discovery prototype that select variables and genreates executable hypothesis. Experiments show that epistemic-control policies can improve efficiency and preserve structure compared with prediction-guided policies.

Critical strengths
  1. interesting motivations: This paper points out a critical limitation in current AI scientist systems. Good predictive performance does not necessarily imply good discovery.

  2. clear conceptual framework: This framework provides a perspective on representation revision, hypothesis generation, evidence acquisition and belief revision.

  3. clear experimental design: The authors conducts controlled comparisons with matched evidence and evaluation budgets to make sure it is fair comparison.

Critical weaknesses
  1. limited number of model backbone for evaluation: Most experiments are conducted using only Qwen-32B, with limited evaluation on GPT-5.6. Since the claims are general principles for AI scientists, it is unclear whether the findings generalize to larger and more capable language models. More extensive evaluations across different model families and scales are needed.

  2. Complicated story, but limited technical novelty: The paper mainly introduces a conceptual framework that combines existing ideas from hypothesis generation, experimental design, causal reasoning, and belief revision. A lot of concepts are proposed but not technically used. The actual implementation largely relies on prompting and predefined decision rules, without introducing a substantially new algorithm for epistemic control. It is unclear what the main methodological contribution is beyond reorganizing existing concepts.

  3. Toy and oversimplied evaluation: Experiments are conducted on small and synthetic tasks with predefined hypothesis and with limited complexity. These settings are actually too easy to check the difference of different frameworks. Authors should consider evaluating on more realistic and complicated scientific discovery benchmark like DiscoveryWorld and DiscoveryBench. The improvement in such datasets can provide stronger evidence to method.

Questions for authors

What is the main technical novelty compared with existing Bayesian experimental design and belief-revision frameworks?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.