acceptodds
Under review as a conference paper at ICLR 2027

PROVER: A Claim-Centric Framework for Auditing Reasoning Traces in Cellular Perturbation Analysis

Abstract

In the context of AI for Virtual Cells (AIVC), cellular perturbation models increasingly generate natural-language analyses alongside predictions of cellular responses. While these analyses promise biologically grounded interpretability, their validity is rarely examined beyond final prediction accuracy. We introduce (erturbation easoning utput ifier), a claim-centric hierarchical framework for auditing reasoning traces in cellular perturbation analysis. PROVER reconstructs each trace as a structured argument graph centered on Claims and evaluates it along two dimensions: and . Cogency assesses whether premises are supportable, evidence is relevant to target Claims, and inferential connections provide sufficient support for their conclusions, while Completeness measures coverage of task-relevant analytical perspectives. We apply PROVER to a 600-case benchmark spanning eight perturbation-analysis systems, examine three LLM audit backbones and holistic-judge baselines, and complement these analyses with expert feedback. The audits flag unsupported assumptions, context-misaligned evidence, and insufficiently supported inferences that answer-level metrics do not assess. Our results show that evaluating biological reasoning requires assessing not only prediction outcomes but also the evidence and reasoning used to justify them. We release the framework, auditing protocol, benchmark metadata, and audit artifacts.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Reasoning, Interpretability, AI for virtual cells, Cellular Perturbation

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_Y8muOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces PROVER, a framework for auditing natural-language reasoning in cellular perturbation analysis. It reconstructs model outputs as claim-centered argument graphs containing premises, assumptions, and inferential links, then evaluates premise acceptability, evidence relevance, inferential sufficiency, and task completeness.

Critical strengths
  1. The paper addresses a relevant limitation of prediction-based evaluation: a correct answer can still be accompanied by an unsupported or inconsistent explanation.

  2. The claim-centered representation makes individual premises, assumptions, and inferential steps inspectable. Preserving source spans and distinguishing missing evidence from contradiction are useful design choices.

Critical weaknesses
  1. The scientific motivation needs to be sharper. The introduction explains why plausible explanations can be misleading, but does not sufficiently establish how auditing them would improve perturbation research, experimental decisions, or hypothesis selection.

  2. The framework is largely heuristic. Argument reconstruction, implicit-assumption extraction, task checklists, and score aggregation involve substantial design choices whose reliability and sensitivity are insufficiently validated. Averaging inferential scores can also obscure a critical unsupported step.

  3. The necessity of the proposed dimensions is unclear. Acceptability, Relevance, and Sufficiency are conceptually reasonable, but their distinct empirical contributions are not demonstrated. Comparisons against simpler rubrics and ablations of individual dimensions would help justify the design.

  4. Independent validation is weak. All 44 complete expert reviews come from one expert, with only 33 exact-text matches for paired analysis. The primary evaluator shows negative correlations with expert ratings for relevance and inference quality, leaving the accuracy of its judgments unresolved.

  5. The benchmark provides limited evidence of generalization. Although it contains 600 traces, most systems contribute only 60 cases, and the tasks differ substantially. The study does not establish reliable auditing across broader cellular contexts, unseen task settings, or independently annotated reasoning failures.

  6. The results depend strongly on the audit configuration. Different evaluators produce substantially different scores, while prompts, reconstructed graphs, and reference exposure also vary. Without controlled comparisons, it is difficult to identify the source of disagreement or determine which judgments should be trusted.

Questions for authors

Could you clarify which underlying models and generation settings produced the 600 traces? How well does PROVER generalize to traces generated by unseen models with different reasoning styles and structures?

Reviewer_rtYSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces PROVER for auditing model-generated natural-language analyses of cellular perturbations. It reconstructs each analysis as an argument graph containing premises, warrants, and claims, and evaluates premise acceptability, evidence relevance, inferential sufficiency, and task completeness. Local judgments are aggregated through deterministic rules. The authors apply three LLM-based audit configurations to 600 analyses from eight systems, compare them with holistic judges, and provide preliminary expert feedback and case studies.

Critical strengths

1.I find the problem well motivated. Errors in biological analyses may occur in evidence use or inference without appearing in the final answer. Representing premises, warrants, and claims separately makes these concerns easier to locate. The cases involving unsuccessful retrieval being used to infer no effect, and disagreement between a trace and its submitted answer, illustrate the practical value of inspecting the reasoning itself.

2.The authors are relatively transparent about the framework’s limitations, including the distinction between unknown and false, differences in evaluator inputs, and the preliminary nature of expert feedback. Retaining source spans, local judgments, and repair records also makes the outputs inspectable.

Critical weaknesses

After reading the paper, I am still unsure how much confidence to place in PROVER’s diagnoses. The framework is clearly laid out, but the experimental evidence for the reliability of its judgments remains limited. The following points are my main concerns.

1.Passing structural checks on all 600 samples establishes compliance with format and dependency constraints, but not correct extraction of claims, preservation of qualifiers, or reconstruction of hidden warrants. Errors here could cause the evaluator to judge problems introduced by reconstruction. I would like to see a manual examination of a subset of graphs, including whether reconstruction errors change the resulting diagnoses.

2.I am also uncertain about sensitivity to graph decomposition. Equations (3) and (4) average Link scores within Paths and then average across Paths, so alternative decompositions of the same argument could change the summary by introducing different numbers of steps or overlapping paths. The authors clarify that these are not proof probabilities, but they are still used to characterize argument quality. Could a small set of meaning-preserving alternative decompositions establish whether the scores and diagnoses remain stable?

3.My main concern is that the reliability of PROVER’s judgments has not yet been established. The 44 complete reviews in Appendix J come from one expert, only 33 match the benchmark text exactly, and Table 14 does not show consistent expert alignment. For Qwen, the correlations for relevance and inference quality are −0.38 and −0.20. I understand that this small pilot does not support strong conclusions, but it also provides limited evidence of evaluator validity. Could the authors directly assess diagnostic correctness on a small subset independently annotated and adjudicated by multiple experts?

4.The evaluation is difficult to interpret because evaluator prompts and reference inputs differ, the holistic baseline uses fewer calls, and the GSE rubric extends beyond the original task. A controlled comparison with comparable budgets and task-matched criteria would help establish how much the argument graph itself contributes.

Questions for authors

See above weaknesses.

Reviewer_D9icOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PROVER audits the presented reasoning in cellular perturbation analyses. It reconstructs traces as claim-centered argument graphs with premises, warrants, inferential Links, and Paths, then evaluates Acceptability, Relevance, Sufficiency, and task-specific Completeness. The paper applies three LLM pipelines to 600 traces from eight systems, compares the resulting scores with holistic judges and answer correctness, and presents qualitative cases, an AROMA sensitivity study, and preliminary expert feedback.

Critical strengths
  1. A perturbation prediction can have the correct label despite an unsupported explanation. PROVER explicitly targets the presented argument rather than claiming to recover complete biological truth.

  2. PROVER is an inspectable representation with useful safeguards. Premises, warrants, and inferential dependencies provide concrete locations for criticism. The paper also reports evaluator sensitivity and important limitations.

Critical weaknesses
  1. The benchmark is very limited compare to the width of the intended findings, and its evidential role is unclear. The 600 cases span 8 systems, but most systems contribute only 60 traces. DE and DIR tasks already cover 442 cases (74%) so the remaining tasks are small, distinct extensions. More importantly, these are mostly traces scored by PROVER, rather than independently annotated examples of reasoning quality.

  2. The independent evidence does not validate the auditor's central biological judgments. Appendix J reports 44 completed reviews from external expert, with 33 exact text matches for comparison. Table 14 gives primary-Qwen correlations of −0.38 for Relevance and −0.20 for Sufficiency versus expert inference quality. These limited comparisons do not prove that PROVER is wrong, but they do not establish reliable diagnosis.

Questions for authors
  1. What is the 600 case corpus intended to show: applicability of the framework, accuracy of the auditor, or general findings about perturbation reasoning? Given its concentration on DE/DIR tasks and limited cases for each system, which conclusions does the current coverage support?

  2. How should the mixed or negative expert alignment in Table 14 and the high automated scores for SynthPert-0 affect our confidence in PROVER's biological judgments? What existing evidence distinguishes correct detections from plausible but mistaken audit explanations?

  3. Which recurring failure patterns or biological-reasoning insights are shown across the corpus, rather than illustrated by selected cases? Given the substantial pipeline sensitivity, which conclusions remain supported across task-specific analyses and evaluator choices?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.