acceptodds
Withdrawn from ICLR 2027

From Tools to Instruments: Tracking Measurement Uncertainty in Autonomous Science Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI co-scientists increasingly do work that researchers have always done themselves: they test hypotheses on their own, from planning an analysis to interpreting its results. Yet inaccurate measurements from their tools can turn a correct analysis into a wrong conclusion. Such failures are hard to identify when reference answers are unavailable. We propose a framework that treats tools as measurement instruments and carries measurement assessments through autonomous hypothesis testing. Instrument disagreement supplies an assumed error scale, and a measurement counterfactual estimates the perturbation needed to change a statistical decision. Four checks combine measurement sensitivity and instrument disagreement with sampling uncertainty, verdict–evidence contradictions, and disagreement across repeats. Their flags define a risk level computed without reference answers. We implement the framework in VERITAS-M, a multi-agent clinical co-scientist, and evaluate it on hypotheses over cardiac and brain MRI, retinal images, and clinical data. VERITAS-M improves on its predecessor; using the same primary model with an additional critic, it also outperforms single-model agents in verdict accuracy. With each cohort's primary instrument, keeping conclusions flagged by at most one check lowers observed error about tenfold; the rule also lowers error for three systems it was not tuned on. Because inaccurate measurements can yield wrong conclusions that stay consistent across repeated runs, autonomous science needs what experimental science has long practiced: treating tools as instruments whose errors are measured and carried through to the conclusion.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI Co-Scientist, Multi-Agent Systems, Scientific Discovery, Medical Image Analysis, Measurement Uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_4tAROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper seems to propose an improvement over a prior LLM-agent that has tool use capabilities. The tools in question are VLMs or image classification/segmentation/detection models, working on medical imaging datasets. The proposed LLM-multiagent uses these tools to come to verdicts which are then evaluated against ground truth values to compute verdict accuracy metrics.

The paper is extremely hard to read and most likely it was written by an LLM.

The novelty or claimed contributions are unclear. There seems to be an extension to the number of tasks from the prior work, VERITAS. There also seems to be some kind of variance measurement (wrongly) framed as uncertainty.

Methods are extremely unclear due to, probably, the LLM prose of the paper. The proposed system seems to consist of agents capable of image processing tasks (i.e. the tools mentioned above), and some LLM-agents that conduct statistical testing (probably by using statistics libraries) and some user-facing LLM outputting human-readable strings as verdicts.

The reported empirical improvements lack context. The tasks are not clearly defined. There seems to be improvement in metric values but judging the benefits are difficult.

Critical strengths

I like to think the authors extended a prior multi-agent platform for hypothesis testing in medical imaging tasks with uncertainty-aware methods. This might have been a contribution if it was true but it is extremely difficult for me to say whether it is true or not.

Critical weaknesses

The paper is quite impossible to read for a human.

The contributions are unclear. Novelty is unclear. Methods are unclear. Datasets, tasks and experimental setup are unclear unless one goes and reads the prior VERITAS paper.

Questions for authors

None.

Reviewer_GD4pOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed using instruments instead of tool use in an ai co-scientist system. The authors design four gates and define a risk level as the feedback signal, integrating into the actual system flow. The integrated system gets higher performance and fewer wrong labels in the system.

Critical strengths
  • proposes a way to transform tool use as instruments and introduce uncertainty into ai co-scientist system.
  • measures and decomposes uncertainty into measurable gates.
Critical weaknesses
  • the proposed framework only works for limited special systems and the experiment are conducted on the limited scinario and benchmarks, making the argument for its generalizabiltiy less convincing.
  • the risk score design is heuristic, and lacks sufficient explanation or an ablation study
  • the system builds on an existing multi-agent framework, and I am not fully convinced about the novelty of the proposed approach beyond integrating existing techniques.
Questions for authors
  • What is the cost of introducing extra agents repeated executions, and how does the system perform under the same computational budget? If adding extra agents is token consuming, could you discuss the cost -performance tradeoff?
  • Is this design transferable to other systems outside the medical domain?
  • Why is the threshold value sets as r < 2? Although the paper reports results at different risk levels, I want to know the motivation for choosing this value.
  • Based on the ablation studies of the four checks, could you discuss the specific roles they play, how they affect performance, and why all four are necessary?
Reviewer_SAdMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VERITAS-M, a framework for assessing the reliability of autonomous scientific agents when their conclusions depend on measurements produced by external tools. Building on the previous VERITAS multi-agent system, the framework treats these tools as measurement instruments and aims to identify unreliable scientific conclusions without requiring ground-truth measurements at runtime.

The framework introduces a structured statistical analysis backend and a measurement assessment layer. It estimates measurement error scales from disagreements between instruments and uses measurement counterfactuals to assess how much perturbation would be needed to change a statistical decision. The authors further combine four reliability checks, including instrument disagreement and measurement sensitivity, verdict–evidence contradictions, proximity to effect-size thresholds, and inconsistency across repeated runs, into a risk score for identifying potentially unreliable conclusions.

The authors evaluate the framework on 279 clinical hypotheses across cardiac MRI, brain MRI, retinal imaging, and clinical datasets, comparing it with the previous VERITAS system and several single-model baselines. Experimental results suggest that VERITAS-M improves final verdict accuracy, while the proposed risk-based selection reduces the observed error rate from 15.3% to 1.5% when retaining 61% of eligible runs. Additional experiments with less accurate measurement instruments show that measurement-related checks can identify incorrect conclusions that remain consistent across repeated runs. Overall, the paper argues that explicitly tracking measurement uncertainty and combining it with statistical and consistency checks can help identify unreliable conclusions in autonomous scientific workflows.

Critical strengths
  1. The paper provides a comprehensive experimental evaluation of the proposed framework across multiple clinical datasets, measurement instruments, and agent configurations. I particularly appreciate that the authors do not rely solely on overall accuracy improvements, but also conduct extensive ablation studies, instrument substitution experiments, controlled error injection, and evaluations across different agent systems. These experiments help examine the contributions of different components and provide a more detailed understanding of when the proposed reliability checks are effective. I think this relatively thorough experimental design is one of the main strengths of the paper.

  2. The paper provides several interesting and practically relevant empirical findings regarding the reliability of autonomous scientific agents. In particular, the experiments demonstrate that inaccurate measurements can lead to incorrect scientific conclusions that remain consistent across repeated runs, highlighting a limitation of relying solely on agent consistency. The proposed reliability checks also achieve encouraging results, reducing the observed error rate from 15.3% to 1.5% while retaining 61% of eligible runs, with similar improvements observed across additional agent systems. These findings provide useful evidence that measurement-aware reliability assessments can complement existing approaches to evaluating scientific agents.

Critical weaknesses
  1. My main concern is that I do not see substantial methodological novelty in the proposed framework. The approach mainly combines existing ideas from measurement error estimation, statistical sensitivity analysis, and consistency checking within the VERITAS workflow. The paper would benefit from a clearer explanation of which aspects of this combination constitute a methodological advance over existing approaches. The incremental value of the measurement counterfactual is also not consistently established. For example, Table 10 reports a lower label-error AURC for instrument disagreement alone than for the full instrument check under substituted instruments (0.090 vs. 0.104). The counterfactual provides benefits in other settings, but it remains unclear when and why it offers a meaningful advantage over directly comparing instrument outputs. Despite the comprehensive experiments and encouraging overall results, I remain uncertain whether the proposed method provides a sufficiently distinctive contribution.

  2. I also have concerns about the interpretation of the assumed measurement-error scales and the final risk score. Appendix H shows that scales derived from instrument disagreement can substantially underestimate expert-derived error dispersion when instrument errors are correlated. Replacing these scales with expert-derived dispersion does not improve overall error ranking, but it remains unclear whether scale underestimation leaves particular measurement-induced errors in the retained low-risk subset. Relatedly, the risk score assigns equal binary weights to four checks with different implications. A strict contradiction between the agent’s verdict and its computed evidence label necessarily indicates an error in at least one output under the paper’s joint error definition, whereas a thin-margin flag indicates potential uncertainty. Nevertheless, both contribute one point, and a run with only a strict verdict–evidence contradiction can still be retained under r<2r<2. The authors provide risk–coverage analyses and examine alternative check orderings, but the rationale for treating these signals equally remains insufficiently clear. These choices affect how the risk score should be interpreted and which conclusions are considered suitable for retention.

Questions for authors
  1. Could the authors clarify the main methodological innovation of the proposed framework and its differences from existing measurement error estimation and sensitivity analysis methods? I understand how the components are integrated into the scientific agent workflow, but the methodological advance of this integration is not clear to me. In particular, what new capability or insight is provided by the proposed measurement counterfactual?

  2. Under what conditions does the counterfactual provide useful information beyond instrument disagreement? Table 10 shows mixed results for its incremental contribution. Could the authors report how often the counterfactual identifies incorrect evidence labels that instrument disagreement misses, and how often it additionally flags correct labels? It would be helpful to know whether these additional detections improve the final risk-based selection, including how many otherwise-retained errors are excluded because of the counterfactual.

  3. How does underestimation of the measurement-error scale affect the assignment of low risk? Appendix H.2 reports substantial differences between the assumed scales and expert-derived dispersion for some instruments, while replacing the scales does not improve overall error ranking. Could the authors examine whether particular cases of scale underestimation cause measurement-induced label errors to remain at r=0r=0 or r=1r=1? Distinguishing errors that change the evidence label from inaccurate measurements that leave the label unchanged would help clarify the practical consequences of this scale mismatch.

  4. What is the rationale for assigning equal binary weights to the four checks, given their different implications? In particular, a run with a strict verdict–evidence contradiction as its only flag would still be retained under r<2r<2. How many retained runs fall into this category, excluding cases flagged solely by the additional pp-value warnings? Could the authors explain why such contradictions are treated in the same way as signals of potential uncertainty, and how the alternative check-ordering results informed the final aggregation rule?

Reviewer_dTB2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The papers looks at the problem of autonomous agents that are trying to draw novel hypothesis or conclusions in a process where they utilize external tools to gather evidence to draw conclusions from. These measurements may be inaccurate, which raises a different failure category than agent's capability. The authors propose a framework to identify which conclusions are unreliable, and introduce a measurement counterfactual that estimates how much the measurements need to change before the statistical conclusion changes. they propose four different kind of checks in their experiments each for a different source of inconsistency in the process, and combine them to make a risk score. This is my best understanding of the proposed methodology and problem, though the exact formulations and technical details are very hard to follow and the writing and format is not well-suited for an academic publication.

Critical strengths

The problem the paper tries to address seems interesting.

Critical weaknesses

The methodology for addressing the problem is not coherent, hard to understand, and inconsistent.

I found the problem formulation and the methodology proposed very hard to follow or understand. The motivation of the paper seems interesting, which is arguing that the failures in hypothesis generation can come from two sources: a) inconsistent measurements that are collected b) wrong or inconsistent agent's interpretation of the underlying measurements. for each source of error, we need to have specific checks and verification tailored for it. The high level proposal of trying to change the measurements and find counterfactual effects of if the conclusions the agent draws form the measurements changes by the perturbations seems interesting? thought I was not able to clearly understand the proposed solution or the justification behind several of its components.

There are many terms and quantities introduced without sufficient explanation, and the presentation makes it very difficult to asses the technical or any contributions of the paper.

To be concrete: here are some examples from the paper that I found difficult to understand:

"An agent specifies an analysis ϕ and cohort D." What is the cohort D here? It is not clearly defined.

"where δ is the smallest effect size considered substantively relevant and plausibility gates are held fixed." What are the plausibility gates? Why are they held fixed? This is not explained when the quantities are introduced.

"Proposition 1 (Observable and latent error). A run is wrong, meaning L ≠ Lref or v ≠ m(Lref), if and only if L ≠ Lref or v ≠ m(L)." What is this proposition trying to establish? The equivalence seems to follow immediately from the definition of an incorrect run.

"Disagreement supplies an assumed error scale σ, conditional on error covariance and allocation between instruments." What does this mean? How is the error scale obtained from disagreement, and why is this a meaningful estimate of the actual measurement uncertainty?

"For a continuous biomarker, let Td shift readings differentially across the comparison by dσ, with T0 = T." I do not understand what it means to shift readings differentially across the comparison by dσ, or why this is an appropriate way to model measurement errors.

"This ordinal score supports selection (Geifman & El-Yaniv, 2017) without assigning error probabilities." What is the justification for this score? Why does counting the number of triggered checks provide a meaningful ranking of the reliability of scientific conclusions?

The paper needs significant improvements in its presentation and technical formulation for its contributions to be properly assessed.

Questions for authors

Please see weaknesses above.

I am also concerned about the plots in the experiments? the curves and figures look unusual to me (AI generated), and it is not clear how they were generated from the experimental observations.

I would appreciate clarification on how the plots were produced, what data points they represent, and whether any smoothing or interpolation was applied.

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

Related papers

Discussion (0)

Sign in to comment.