acceptodds
Withdrawn from ICLR 2027

Auditing Discovery Claims: A Two-Sided Criterion for Agentic Science

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When an agentic science system improves against a fallible evaluator, what establishes a new capability rather than evaluator-specific optimization? We propose a two-sided audit that separates an exact exclusion for the prior verifier from empirical gains under a shared adjudicator. RNA topology supplies the negative side: a pseudoknot-free verifier cannot certify crossing structures at any search budget. The positive side is model-relative, assessed by three structure predictors with one held outside every optimization loop. The audit returns six findings rather than a pass/fail bit. A hand-built operator solves 43 of 60 crossing targets under the predictor it optimizes, but only one under all three. On the same 43 targets, an unoptimized predictor confirms two of its designs versus 26 for an external solver. Auditing our revised operator finds a narrow improvement but no measured advantage over matched-cap undirected search. The instrument nevertheless discriminates: two failure-free frozen LLM-written operators attain held-out carry-over of 0.293 versus 0.095 on 951 paired units, using 4.6–10 times fewer oracle calls. Seven interventional tests do not explain the difference. These results support externally adjudicated, compute-accounted capability comparisons, not physical validation, a policy-level grammar transition, or discovery of a transferable design principle.

Primary area:
datasets and benchmarks
Keywords:
agentic science, capability evaluation, verifier dependence, reward hacking, RNA inverse design, formal languages, held-out validation, reproducibility

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_st9hOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

When an agentic scientist improves a score, it's unclear if they did so meaningfully or gamed the evaluator. The authors claim to provide an audit to disentangle these cases, and I am guessing that they do this in the domain of RNA topology. The authors produce a list of six results about their audit, but do not provide any takeaways and claim that they cannot conclusively say that there was a "pass" or "fail" verdict in the audit.

Critical strengths
  1. None, this paper should be desk rejected for abusing AI systems in the research process
Critical weaknesses
  1. This paper is impossible to understand, so I cannot even begin to judge it. I think I was lost immediately after the first paragraph of the introduction. The authors use terms like "RNA inverse design", "Nested secondary structures", "context-free descriptions", and "seudoknot-free folding oracle" without defining them, citing papers that discuss them, or connecting their relevance to the research question. This paper needs a substantial rewrite. I would recommend the authors follow a more standard flow of: Provide context on the problem -> discuss related work -> introduce the contribution or hypotheses -> discuss experimental setup -> provide results, discussion, and takeaway. Even Figure 1 is very clearly AI-generated. There are two explanations for this: 1) either I am extremely unfamiliar to this field; or 2) the paper uses extensive AI writing which makes it impossible to understand

  2. The experimental results seem extremely weak. There are only 60 examples used for evaluation. I am not an expert on the field of RNA topology, but it is not discussed whether this is a small-scale experiment or what we would expect from this domain

Questions for authors
  • It seems like there is another paper by the same authors under review at ICLR 2027 which is cited four times:

“solved” is a choice of verifier, not a fact: Deterministic cost–robustness frontiers of verifier choice in RNA and SAT

I would recommend the ACs to look into this paper as well, but from the title it seems like the domain and experiment setup could be very similar. Can the authors explain the contributions of this paper beyond that other ICLR paper?

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

This paper studies how to evaluate claims that a scientific agent has gained a new capability. It separates changes in what a verifier can recognize from performance gains under the same evaluation tools. The authors propose a two-sided audit and apply it to RNA pseudoknot inverse design. One side proves that a pseudoknot-free verifier cannot confirm a target with crossing pairs. The other side compares design methods using pKiss, ProbKnot, and ShapeKnots, with ShapeKnots kept outside the optimization process. A hand-written method succeeds on 43 of 60 targets under pKiss, but only one design is confirmed by all three predictors. The authors also find no clear drop in the final confirmation rate when objective guidance is removed from their revised method. Two selected LLM-written programs achieve higher confirmation rates under the unused predictor in a matched comparison.

Critical strengths
  1. The authors distinguish what a verifier can output, whether a system performs better under the same evaluation rules, and whether we understand the reason for that improvement. These distinctions help prevent a tool upgrade or a higher score from being treated as proof that an agent has gained a new ability.
  2. The experiments include useful controls. The authors compare methods on the same targets and test their own method after removing key parts. In particular, removing objective guidance does not produce a clear drop in the final confirmation rate.
Critical weaknesses
  1. The methodological contribution seems limited. The framework combines checks on the old verifier’s output limits, fixing candidates before an unused predictor checks them, shared evaluation rules, resource accounting, and ablations. But it lacks a substantial new method or insight beyond a carefully designed evaluation workflow.
  2. The exact negative result may not apply when another task lacks a provable restriction on the old verifier. Once this special condition is removed, the added value over ordinary held-out evaluation and controlled comparisons remains unclear.
  3. The drop from 43/60 to 1/60 shows dependence on the evaluation tools, but does not by itself prove that the method exploited a predictor. Figure 2 changes the success rule from acceptance by pKiss to acceptance by all three predictors. The second rule is stricter by construction. The drop shows poor agreement across predictors, but it does not by itself tell us whether the designs are wrong, whether the search exploited pKiss, or whether the other predictors failed to recognize valid designs. In Section 4.1 the additional DesiRNA comparison still changes the search algorithm, energy model, and other method choices, and the relative performance changes across predictors.
Questions for authors
  1. I suggest providing further evidence for the stronger claim of exploiting a predictor. Given the low recovery rates on native RNA and the changes in method ranking across predictors, what evidence separates exploitation of a predictor from differences in the predictors’ models or their ability to recognize valid designs?
  2. Can the current experiments show what the framework reveals that would be difficult to learn from standard cross-predictor checks and controlled comparisons?
Reviewer_UpTxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-sided audit framework for evaluating scientific discovery claims made by AI agents. The framework distinguishes limitations of the original verifier from actual capability improvements measured using external evaluators. The authors instantiate this framework in RNA inverse design, using multiple folding predictors to examine evaluator overfitting. The experiments show that a method achieving 43/60 successes under its optimized verifier retains only 1/60 under three predictors. The paper also evaluates LLM-generated design operators and finds improved held-out agreement with fewer oracle calls.

Critical strengths

A meaningful evaluation problem. I think distinguishing genuine capability improvements from evaluator-specific optimization is an important problem for agentic science. The proposed framework provides a useful way to audit such claims, especially in domains where multiple verifiers are available.

Convincing empirical analysis. The 43/60 to 1/60 result clearly illustrates the risk of relying on a single verifier. I also appreciate the controlled comparisons, ablation studies, and compute analysis. The authors are careful about distinguishing observed improvements from explanations of those improvements.

Critical weaknesses

Limited generalizability. The framework works naturally for RNA inverse design because multiple folding predictors are available. It is less clear how the proposed audit could be applied to other scientific discovery tasks, especially those without multiple reasonably independent verifiers.

Reliability of external verification. The three folding predictors are not fully independent and have relatively low recall on native RNA structures. Therefore, cross-verifier agreement does not necessarily imply scientific correctness. This limits how strongly the positive results can be interpreted.

Questions for authors

How would the proposed framework generalize to scientific discovery domains where only one verifier is available, or where independent verification is expensive?

Given the low recall and correlated predictions of the RNA verifiers, how sensitive are the main conclusions to the choice of held-out verifier? Is there additional evidence supporting the robustness of the reported improvements beyond this particular predictor panel?

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

Related papers

Discussion (0)

Sign in to comment.