acceptodds
Under review as a conference paper at ICLR 2027

VEA-Aware Uncertainty for LLM Chain of Actions

Abstract

For the prediction of a binary outcome, Variability–Epistemic–Aleatoric (VEA) uncertainty quantification separates the error of a single model-reported probability into three components: the variability (V) of the prediction across repeated runs of the model, the epistemic gap (E) between the average prediction and the true conditional probability of the outcome, and the aleatoric uncertainty (A) of the outcome itself, which no predictor can impact. This decomposition presents an important distinction that calls for different strategies to improve model performance: averaging over repeated runs of the model of interest reduces variability, and better modeling choices and training procedures reduce the epistemic gap while reporting the intrinsic aleatoric error. We propose an estimation procedure of VEA components in the setting of predictions through a chain of actions, namely a single structured generation mechanism that emits an ordered sequence of intermediate judgments before its final action. First, when the true conditional probability is unobserved, we estimate the epistemic gap (E) against a disclosed proxy model, and we provide an aggregate diagnostic that checks the adequacy of that proxy using the observed outcomes. Second, we further decompose the variability (V) into an exact sum of individual action-specific contributions so that action-level instability can be identified and targeted. We apply our VEA framework to evaluate the replicability of published scientific findings, an important but resource-intensive task in practice, using three core benchmark replication datasets (RPP, CB and SSRP). Across runs per study, the VEA decomposition shows that the overall prediction error is dominated by the epistemic component (E) and is concentrated in a small number of specific actions along the chain. This implies that increasing the reasoning effort of the LLM under evaluation improves prediction accuracy, whereas lowering the sampling temperature does not, because the already small variability (V) has no room for further shrinkage.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
uncertainty quantification, large language models, epistemic uncertainty, aleatoric uncertainty, variance decomposition, chain of actions, replicability prediction, reproducibility

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_K4fxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper propose a decomposition of the squared error of LLM's action (conditioned on sequential actions) with three parts, variability (for different repeated predictions), epistemic (for error between the average prediction and the true posterior), and aleatoric (the entropy of posterior itself). The authors give a good and detailed discussion on estimation of these components. They also demonstrate this framework on three real-world datasets and give interesting interpretations on suggested interventions based on the decomposition results.

Critical strengths
  • The paper is technically sound, offering definitions and proof for every statement. I didn't find any error in it. It is also fun to read and easy to follow.
  • The idea of decomposition is useful for identifying the reason of prediction error. The demonstrations show interesting use case and practical takeaway.
Critical weaknesses
  • I didn't understand the section 4.1 completely. It's a little hard to see the connection between this section and the big picture of the paper. Therefore, I'm not sure how the proposed decomposition is related to a sequence of predictions, compared to just a single prediction.
  • A miscalibrated proxy could make the framework useless. In figure 2c and d, we wouldn't know whether it is because the average prediction by LLM is bad or the label is uncertain by itself. It would be great if the authors can include some results between the calibration of the proxy and the quality of the decomposition in the main texts.
Questions for authors
  • What is the usage of the section 4.1 and its connect to VEA decomposition? And how does it make the framework related to sequence of predictions?
  • How would the authors interpret the results in 5.3 given that the proxy is not calibrated (or truthful)?
2 comments

We thank the reviewer for the positive assessment of the paper and for two insightful and constructive questions, which have helped us improve the paper greatly.

Weakness 1 and question 1

Section 4.1 and its connection to VEA. The VEA decomposition is developed for the total squared error, but it does not tell practitioners when, and how often, VV and EE lead to a very large error in a single run. We are interested in the agreement between one probability prediction pip_i and the true replication probability πi\pi_i. We show (page 4, lines 184-185) that the sum of VV and EE, the key ingredients of the VEA decomposition, equals E[(pi−πi)2]\mathbb E[(p_i-\pi_i)^2]. This connection lets us study the probabilistic behavior of the agreement. It motivated the two bounds (Freedman and empirical Bernstein) on the deviation ∣pi−πi∣|p_i-\pi_i|, which describe the chance that the deviation is unduly large and identify the factors that control that chance. This knowledge is useful for designing suitable interventions. We have revised the opening of Section 4.1 (page 6, lines 309-316) to state this connection.

Sequential versus a single zero-shot prediction. The actions are naturally sequential, in a divide-and-conquer spirit, which gives more insight into the profiles of variability, uncertainty and error in a complex decision process. In a designed chain of actions, we can identify the "trouble-makers" that contribute unduly large errors or uncertainties, so that interventions can be applied at the right place and at the right time. A chain may cost more tokens and computation than a single zero-shot prediction, and more actions may introduce more errors. Nevertheless, in important scientific studies, knowing the detailed sources of uncertainty is of great interest for reproducibility and replicability. The opening of Section 4 now says this (page 6, lines 281-286).

Weakness 2

A toolbox to check the calibration of the proxy. We agree on the importance of a well-calibrated proxy for an appropriate evaluation of uncertainty. At a minimal level of assurance, we propose the Tier-3 diagnostic as a toolbox to check the calibration of the proxy. The observed mean squared error minus the estimated VV and EE terms, n−1∑iR−1∑r(pi(r)−Yi)2−n−1∑i(V^i+E~i)n^{-1}\sum_i R^{-1}\sum_r(p_i^{(r)}-Y_i)^2-n^{-1}\sum_i(\hat V_i+\tilde E_i), estimates the mean aleatoric term n−1∑iπi(1−πi)n^{-1}\sum_i\pi_i(1-\pi_i), and it agrees with the plug-in π^(1−π^)\hat\pi(1-\hat\pi) only when the proxy is calibrated in aggregate. The rationale of the check rests on a sufficient condition: a proxy calibrated in aggregate passes it, so a failure flags the proxy. Figure 4 in Appendix E shows the check for each dataset, Appendix E adds a discussion of the results, and Section 3.2 now refers to both (page 5, lines 246-247). RPP passes; CB and SSRP do not. Section 5.4 now also identifies the results that do not depend on the proxy at all (pages 9-10, lines 485-489).

Question 2

We reiterate that the Tier-3 diagnostic checks whether the proxy (the logistic prediction model) is adequately calibrated. When the check fails, as on CB, we do not suggest stopping the analysis.

Some results in Section 5.3 for CB and SSRP deserve special attention: the small VV values, the dominant action-specific EE values, and the closeness of the scoring action to its proxy. Because the proxy fails the calibration check on both datasets, results that involve the proxy at the scoring action should be interpreted with caution. There are two exceptions, where the miscalibration of the proxy does not affect the estimates.

  1. The variability estimates, because they do not depend on the proxy.

  2. The epistemic components of the extraction and statistical-review actions. These actions are upstream of, and separate from, the scoring action, and their references are the dataset's recorded values (N,p,r)(N,p,r) and the statistical power computed from the original sample size and effect size, rather than the proxy, so their epistemic components can be evaluated appropriately.

Section 5.4 now states this (pages 9-10, lines 485-489). For the scoring action on CB, Section 3.2 already notes that a large epistemic component there measures the disagreement between two imperfect predictors, and that the LLM predicts the CB outcomes better than the proxy does (page 5, lines 250-253).

We are grateful for your encouraging assessment that the paper is technically sound and easy to follow. Your two questions led us to explain how Section 4.1 connects to the VEA decomposition and to make the calibration check of the proxy visible in the main text. We would be happy to answer any further questions.

Thank you again for your constructive feedback. We have included additional analyses and clarifications in our rebuttal in response to your comments. We know you have a busy schedule and would appreciate any further feedback if you have a chance to review our response. We would be happy to address any remaining concerns.

Reviewer_9xeyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

High-level goal: Split uncertainty of LLMs into 3 different sources: V: variability of outputs over different answer generations of the same LLM (due to random seed rr used for the next-token generation); E: epistemic uncertainty of the average result A: irreducible aleatoric uncertainty.

Setting: The setting is not explained very well; I understood it in the following way: They study 3 datasets, each consisting of multiple scientific studies. The task is to predict the probability for each scientific study if it will reproduce. These 3 datasets have binary labels telling us if these studies actually reproduced. They ask a LLM to predict these probabilities in multiple steps. They call each of these steps an “action”. The sequence of steps is then called a chain of actions (CoA). Each of these steps generates some kind of score. After these steps, a final score pi,s(r)p_{i,s}^{(r)} (or pi,M(r)p_{i,M}^{(r)}?) is generated for study ii, seed rr after all ss (or MM?) steps? The goal is to decompose the uncertainty associated with pi,s(r)p_{i,s}^{(r)} into V, E and A. The 2nd goal is to estimate how much each of the ss (or MM?) steps contributes to the V, E and A. The 3rd goal is to derive some theoretical bounds. The 4th goal is to discuss that these decompositions (in particular the attribution to specific steps) might inspire some useful interventions. However, the empirical evidence provided for the usefulness of these interventions is quite weak in my opinion. I still find the idea interesting.

Method: Estimating V is trivial via prompting the LMM with multiple different random seeds rr. Estimating E would usually be challenging, but if you have a proxy for A, it becomes trivial too. Estimating A would usually be challenging, but if you have a proxy for A, it becomes trivial too, by definition. They propose fitting a logistic regression on their labeled dataset to obtain a proxy for A. For 2 out of these 3 datasets, this proxy seems to be usable based on some medium-convincing arguments. For one of the 3 datasets, this proxy seems to fail.

Critical strengths

In principle, the core idea seems simple and interesting. If the paper were presented better and the empirical evaluation were improved, it could be quite interesting.

Critical weaknesses

W1: Weak empirical evidence for interventions: Guiding interventions would be very interesting. However, the empirical evidence on this is very weak, and there is at least one flaw in the experiments.

W1a: VEA decomposes the uncertainty of one individual pi(r)p_{i}^{(r)}. However, you evaluate pˉi\bar{p}_{i} averaged over the R=100 random seeds. The V component of pˉi\bar{p}_{i} should be R=10\sqrt{R}=10 times smaller than the one of pi(r)p_{i}^{(r)}. Do you agree? Your experiments show that V is even very low for pi(r)p_{i}^{(r)}. This would suggest that changing from pi(r)p_{i}^{(r)} to pˉi\bar{p}_{i} would result in very little improvement. If you’ve stored your intermediate results, you can run these experiments basically for free. I think this would be a much more natural experiment for evaluating VV-interventions rather than varying the temperature.

W1b: In line 437, you say “A reversed ordering is observed in SSRP, possibly due to its small sample size.”. Does your method estimate lower epistemic uncertainty for SSRP? For RPP it estimates high epistemic uncertainty, and increasing the effort actually helps in terms of AUC (how about Accuracy, Brier, and NLL?). For CP the proxy is not reliable, so all the CP results should be taken with a grain of salt. If the results point in the wrong direction for SSRP, the overall empirical evidence for guiding the interventions would be extremely weak.

W1c: There are no empirical experiments on the step/action-specific interventions?

The last sentence of the paper is: “The framework is descriptive: it does not improve the forecaster; it makes the forecaster auditable and indicates where improvement effort should be directed.” Do you have any strong empirical evidence that your method “indicates where improvement effort should be directed”?

W2: The presentation is quite weak. I think the main ideas could actually be explained in quite simple words to make this paper a very quick and effortless read that could reach a broader audience. However, you present it in an unnecessarily complicated way. Directly from the start, you place a lot of technical terms that are not really essential for the main storyline, which makes it harder to read (e.g., Freedman tail bound, empirical-Bernstein population guarantee, Doob martingale, Neyman-orthogonal score, nuisance functions). Also, when starting to read the abstract in the introduction, it was often not clear to me what you meant by “repeated runs of the model” (training runs or inference runs) and “action” in the abstract. Undefined notation is a similar problem: In Figure 1: when reading Figure 1, none of the variables D,m,v^i,U,Nipival,ri,...D, m, \hat{v}_i, U, N_i p_i^{val}, r_i,... were defined at this point. Lines 104-108 are hard to understand. Which “event” are you referring to in “it is also the event the bounds of Section 4 control” in line 182. In line 240, the sentence “The check is informative in practice.” could be replaced by something more to the point. Also, in lines 244-246 the sentence “An aleatoric term πˆi(1−πˆi) is reported only where the proxy is a genuine probability of a defined outcome, not for agreement, power, or plausibility scores” is confusingly written. For which indices ii is it not a probability? Or are you talking about some steps mm for which A is not reported, but why don’t you mention mm at least as an index in the equation then? What are “epistemic statuses”? What are “identifiable actions”? Line 409: “Aleatoric A is shown only for actions with a probabilistic proxy.” Lines 482-484 are hard to understand.

(W3: A better proxy of A would make the paper more compelling. In line 251-252, you write: “At deployment the frozen proxy applies to a new study directly”. If you think that the proxy generalizes well to deployment, then this proxy would be a better estimate than the LLM, so why bother asking the LLM in the first place? For which actions do you have a probabilistic proxy?)

Questions for authors

I think a lot of changes would be necessary to push this paper above the bar: better presentation and much more and stronger evidence. The core idea sounds promising.

8 comments

W1

We are grateful to the reviewer for such a careful reading of our manuscript, especially Comment W1c, which points out a gap (termed a "flaw" in the experiments) between the claim and the evidence included in the main text. In this revision we have fully addressed this gap in a rewritten Section 5.4.

W1a

We agree that the variability of a 100-run average is Vi/100V_i/100, and that comparing the AUC of two such averages cannot fully reveal whether temperature changes the VV component. However, our decomposition targets the single-run predicted probability pip_i, and ViV_i characterizes the run-to-run variability of that single prediction. The predicted probabilities from the 100 zero-shot runs are used to estimate ViV_i; the average pˉi\bar p_i is not our target, although it is needed to estimate ViV_i.

Given this design, varying the temperature remains a valid way to evaluate a VV-intervention, provided that its effect is read from ViV_i, the variability of pip_i rather than of pˉi\bar p_i. We measured ViV_i at both temperatures; the results are in Table 10 of the appendix (Table 8 in the initial submission). To avoid the confusion caused by reporting only the AUC of the averages in the main text, Section 5.4 now points readers to Table 10 for the full evidence (page 9, lines 470-474).

To show this directly, the table below reports the single-run standard deviation of pip_i at two temperatures (0.20.2 and 0.70.7) and two reasoning efforts (low and high).

high reasoning effortlow reasoning effort
Datasettemp. 0.70.7temp. 0.20.2temp. 0.70.7temp. 0.20.2
RPP0.0530.0520.0560.057
CB0.0660.0650.0570.056
SSRP0.0560.0550.0520.053

The table shows that lowering the temperature changes the single-run variability only slightly. Moreover, as shown in Section 5.3 (page 8, line 420), ViV_i is a small share (a median of 5%5\%) of the reducible error Vi+EiV_i+E_i, so this small change has little effect on the total uncertainty. This supports directing improvement effort to the epistemic component.

These points are now made throughout the paper. Section 5.1 (page 7, lines 373-375) clarifies that (i) ViV_i always refers to a single run and (ii) the R=100R=100 runs are used to estimate it. Section 5.4 (page 9, lines 470-474) discusses Table 10. The Abstract now states that these insights are supported by designed interventions that vary the sampling temperature and the reasoning effort (page 1, lines 32-34), and the Introduction reports the run-to-run variability directly, in addition to the AUC comparison (page 3, lines 133-135).

Following your suggestion to use the stored results, we took the average of the 100 runs as a new predicted probability, which is a VV-intervention. Averaging does not change the epistemic and aleatoric components, because the expected value of the 100-run average equals pˉi\bar p_i. It does, however, lower the expected squared error, as the reviewer notes. The table below compares the Brier score of a single run (Table 8 in the appendix) with that of the 100-run average (Table 2 in the main text).

DatasetSingle run100-run average qˉ\bar qDifference (%)
RPP0.2320.2290.003 (1.3%)
CB0.2230.2190.004 (1.8%)
SSRP0.2240.2200.004 (1.8%)

These results confirm the reviewer's insight: the 100-run average improves the prediction very little. The decomposition explains why: VV is small.

W1b

We respond in two parts.

  • Does the decomposition estimate a lower epistemic component for SSRP? Yes. At the scoring action, the estimated epistemic component is E~=0.03\tilde E=0.03 on SSRP, against 0.050.05 on RPP and 0.080.08 on CB. The aleatoric component takes the largest share on SSRP: only about 13%13\% of its squared error is reducible, against about 30%30\% on CB (per-action values in Appendix E).

Since the LLM vote rate on SSRP (AUC 0.8270.827) is already very close to the proxy (AUC 0.8460.846), the decomposition leaves little room for an EE-intervention on SSRP. This point has been added to the revised manuscript (Section 5.4, page 9, lines 479-483). For clarity, we have also deleted the sentence "A reversed ordering is observed in SSRP, possibly due to its small sample size" (line 437 of the initially submitted version).

  • Other metrics. In addition to AUC, Table 8 in the appendix now reports accuracy, Brier score and NLL at both reasoning-effort levels. Unlike SSRP, where an EE-intervention has little effect, on both RPP and CB higher reasoning effort consistently improves AUC and accuracy and reduces the Brier score and NLL, at both temperatures.

In response to the remark that "all CB results should be taken with a grain of salt", we added a paragraph to Section 5.4 (pages 9-10, lines 485-489) on which results depend on the proxy, and a sentence on the effort comparison to Appendix E (page 21, lines 1113-1115). The Tier-3 failure affects the proxy-dependent quantities at the scoring action, namely its epistemic and aleatoric terms. It affects neither the comparison of reasoning efforts nor the extraction action, because both are evaluated against the observed CB data rather than against the proxy.

W1c

We appreciate this valuable suggestion. We now include an experiment with an action-specific intervention on RPP, using gpt-5.4-mini, the model of the main experiment, at temperature 0.70.7 and high reasoning effort (Section 5.4, page 10, lines 490-501; Appendix E). The intervention introduces spurious statistics into the statistical-review action (action 3). In principle, asking the LLM to follow a spurious rule for assessing statistical power should increase the epistemic component of this action, and comparing the VEA decompositions with and without the intervention lets us evaluate the design of the chain for a specific task. The epistemic component of this action doubles, from 0.150.15 to 0.300.30, whereas the epistemic components of the other actions remain nearly unchanged. Moreover, as Figure 5 shows, the intervened chain attributes less of the run-to-run variability to the statistical-review action and more to the researcher-degrees-of-freedom action (action 4), and its residual term also grows. The AUC, averaged over the 100100 runs, drops from 0.6110.611 to 0.5810.581 (Table 9 in Appendix E). This experiment shows that the decomposition captures the effect of an intervention at a specific action. This action-level sensitivity makes the action-level uncertainty profile a useful guide for designing and evaluating action-specific interventions in practice.

W2

To address this constructive comment, we have made the following changes, highlighted in blue in the revised paper.

  • Abstract and Introduction: (i) run and action are now defined (page 1, lines 13-14 and 45-47); (ii) "repeated runs" are described as inference-time re-executions of the same prompt, not retraining.

  • Figure 1 caption: all symbols used in the panels are now defined, with references to the sections that introduce them (page 2, lines 72-75).

  • Contribution (C2) has been rewritten in plain language (pages 2-3, lines 103-112). The terms Doob martingale, Freedman, empirical Bernstein and Neyman orthogonality no longer appear in the Abstract or in (C2); they are introduced in Section 4 and Appendix C.

  • The indices ss and MM asked about in the reviewer's summary are now stated in Section 4 (page 6, lines 274-278). M=7M=7 is the number of actions in the chain, and s=6s=6 is the position of the scoring action, at which the LLM reports the replication probability pi=pi,sp_i=p_{i,s}, the quantity the decomposition analyzes. The verdict, action 7, follows the score and makes a binary decision (replicable or not) rather than reporting a probability. The final predicted probability is therefore pi,s(r)p_{i,s}^{(r)}, not pi,M(r)p_{i,M}^{(r)}, and the attribution of the variability runs over actions m=1,…,sm=1,\dots,s.

  • Lines 104-108 belong to Contribution (C2), rewritten as above. Line 182: the sentence now names the event (page 4, lines 188-190). The deviation event {∣pi−πi∣>τ}\{|p_i-\pi_i|>\tau\}, in which one run misses the true probability by more than a tolerance τ\tau, is the event whose probability the bounds of Section 4.1 control.

  • Line 240: "The check is informative in practice" is replaced by "The check separates a proxy that can serve as a reference for the LLM from one that cannot" (page 5, lines 249-250).

  • Lines 244-246: the sentence on the aleatoric term has been rewritten and now refers to the actions mm explicitly (page 5, lines 256-260). Each action mm has its own reference π^i,m\hat\pi_{i,m} (Table 3), and its components Vi,mV_{i,m}, Ei,mE_{i,m} and Ai,mA_{i,m} are computed against that reference. For the extraction and verdict actions the reference is a fixed value or the observed outcome, so Ai,m=0A_{i,m}=0. The Tier-3 check of the proxy applies only to the scoring action, whose output is the predicted replication probability; the other actions are intermediate actions.

  • "Epistemic statuses" is replaced by "The three components differ in what data they require and in what can be claimed for them" (page 5, lines 262-263). "Identifiable actions" is replaced by "specific actions" (page 8, lines 422-423; Figure 2 caption, page 9, line 462).

  • Line 409 (Figure 2 caption): the sentence now reads "Each action's AA is computed against its own reference; it is zero for extraction and verdict, whose references are a fixed value and the observed outcome" (page 9, lines 463-464).

  • Lines 482-484 (the fourth limitation): the term "nuisances" is removed, and the limitation now states plainly that confidence intervals are given only for each action's average contribution and that the per-study breakdowns are descriptive (page 10, lines 535-537). Appendix C explains why (page 18, lines 952-953): the regularization bias of the fitted regressions averages out across studies, but not within a single study.

W3

In our paper, the proxy is a supervised logistic model trained on the observed replication labels, and such labels are rarely available in practice. We use it to measure EE on labeled benchmarks, not to forecast new studies. In contrast, the LLM is a zero-shot procedure that applies to any paper. On CB, the LLM also predicts replicability better than the logistic proxy does (AUC 0.700.70 versus 0.580.58).

On the question of probabilistic proxies: the probabilistic proxy in that sentence is the logistic-regression proxy of the scoring action, whose output is the predicted replication probability pip_i, the quantity of interest. The Tier-3 check applies to this proxy only; the other actions are intermediate actions, each with its own reference (Table 3). We clarified this in Section 3.2 (page 5, lines 256-260).

Questions for authors

Your positive review is really appreciated. We have carefully addressed all the points raised in your review and incorporated the corresponding changes throughout the revised manuscript. We look forward to any further insightful and constructive comments you may have.

Reviewer_x7irOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework to decompose LLM prediction error into three parts: run-to-run variability (V), epistemic gap (E), and aleatoric uncertainty (A), specifically for models outputting a structured chain of actions. Since true probabilities are unknown, they estimate the epistemic gap using a proxy model and introduce an aggregate diagnostic to check if that proxy is actually reliable. They also use a Doob martingale setup to attribute run-to-run variance to specific steps in the chain, along with Freedman and empirical-Bernstein tail bounds. They test this on predicting scientific study replicability across three datasets (RPP, CB, SSRP) and show that epistemic error dominates, meaning increasing reasoning effort helps much more than lowering temperature.

Critical strengths

The decomposition of uncertainty along the chain of actions is very well thought out. Using a Doob martingale with Neyman-orthogonal cross-fitting to attribute run-to-run variance to specific intermediate steps (and tying it to practical interventions like reasoning effort vs. temperature) makes the UQ framework genuinely actionable rather than just descriptive. Also, the Tier-3 aggregate diagnostic is a clever and practical way to validate whether the external proxy is actually good enough to judge the LLM against, which pays off nicely when analyzing the CB dataset failure.

Critical weaknesses

The evaluation relies on a single model family (gpt-5.4-mini) and a single rigidly structured 7-step prompt format. Because of this, it is a bit unclear how well these empirical takeaways and action-level attributions generalize to other model architectures or less constrained agentic workflows. Additionally, as the authors candidly show in their contamination probe, pretraining leakage on older public benchmarks like CB is pretty high (nearly 30% five-hit recall), which somewhat clouds the zero-shot performance claims on that dataset.

Questions for authors

In Appendix C, 55 out of 158 CB claims had to be dropped from the cross-fitted stage attribution due to missing extraction values. Does dropping a third of the dataset introduce selection bias into the variance attribution profile for CB? It would be nice to see if this exact decomposition holds up when evaluating an open-weights model where pretraining cutoffs or data can be more strictly controlled.

2 comments

We thank the reviewer for the positive assessment and for the insightful question on selection bias, which is thoroughly addressed later in this response.

Question 1

Selection bias from dropping 55 of 158 CB claims. The 55 claims on the CB data were dropped because the extraction-action failed to score, mainly because the original papers do not report a jointly usable sample size, pp-value and effect size. Such missing information does not seem to be intentionally removed or masked by the authors to avoid the checking of research replicability. This assumption can be justified, at least partially, by the fact that a study may report a claim for an experiment with the complete information and another for a different experiment with partial information. So, we would safely assume that the data attrition is very likely to follow the missing completely at random mechanism at the claim-level. Of course, it would be ideal to directly verify this assumption when suitable data were available to do so.

In this paper, we invoked a sensitivity analysis to partially show that data attrition had little effect on our results, partially reflecting the above assumption. This verification includes all actions prior to the final scoring-action. Our results of this sensitivity analysis on the CB data is summarized in the following table (available in Appendix C, Table 6).

Action-1Stat reviewRes. DoFTheoryResidual
Sample (claims)Inclusionnn (claims)GG (papers)Total VV (×10−3\times 10^{-3})θ2\theta_2θ3\theta_3θ4\theta_4θres\theta_{\mathrm{res}}
Claims analyzedYes103204.6552.0%2.8%1.0%44.2%
Claims analyzedNo103204.6551.9%2.8%1.0%44.2%
Claims excludedNo55145.0154.1%1.4%1.5%43.0%
All claimsNo158234.7753.1%2.2%1.2%43.4%

As shown in the Table it is evident that excluding 55 claims in the CB analysis had virtually no difference for the total run-to-run variance of the final score (0.005010.00501 against 0.004650.00465), despite a very minor increase in the unit of 10−310^{-3}. Moreover, the compositional profiles of the action-specific attributions are comparable across different samples of claims used in the analysis. This indicates that the data attrition did not exhibit a noticeable impact in the CB analysis, which reflected the missingness mechanism discussed above.

Weakness, first part, and question 2

To address this limitation of the single-model based analysis, we expanded and added an open-weights model, Qwen3.8-Flash, and repeated the experiments and analysis on dataset RPP, which serves as a better test bed than the other two datasets because of its popularity and higher data quality. Computational cost and the time constraint for the rebuttal also limited the new analysis to one dataset. The table below compares the two LLM models head to head (mean over the 100100 runs; also see Table 15 in Appendix E).

ModelAUCAccuracyBrier
gpt-5.4-mini0.6110.6590.232
Qwen3.8-Flash0.6520.6780.221

The predictive performance of the two models is broadly comparable. In addition, Figure 6 in the appendix shows that the two models have consistent patterns in the per-action variability VV and epistemic component EE of the VEA decomposition and in the tail bounds; Figure 5 in the appendix shows that the two models respond to the spurious-statistics intervention in a highly similar way.

Weakness, second part

We agree that the CB analysis may exhibit inflated zero-shot performance due to the fact that this CB data may have been possibly used already in the LLM-training, causing the so-called pretraining data leakage. We discussed this data contamination as the first limitation of the paper; see page 10, lines 529-532. Specifically, we used a recognition prompt with the same LLM to conduct the contamination probe and the accuracy analysis, and the results are reported in Appendix E (Table 14). On CB, we observed that the zero-shot LLM performance appears more accurate than the proxy even on some claims that the LLM cannot recognize and recall, so contamination cannot fully explain the model performance.

Thank you again for the positive evaluation and for recognizing the action-level attribution and the Tier-3 diagnostic as practical contributions. Your comments prompted the sensitivity analysis of the dropped CB claims, the replication with an open-weights model, and a clearer discussion of contamination. We welcome any further suggestions.

Thank you again for your constructive feedback. We have included additional analyses and clarifications in our rebuttal in response to your comments. We know you have a busy schedule and would appreciate any further feedback if you have a chance to review our response. We would be happy to address any remaining concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision (to all reviewers and the Area Chair)

We thank the three reviewers for their careful reading and constructive comments, which have improved the paper substantially. The revised PDF shows every change since the initially submitted version in blue. The main changes are summarized below.

1. Evidence that the decomposition guides interventions (Section 5.4, Appendix E)

  • Which predictor is evaluated. The decomposition targets the single-run prediction, and the predictive-performance tables (Tables 2, 8, 12 and 13) now evaluate single runs, reported as the mean and SD over the 100 runs. The run-averaged predictor appears only in Table 9, as an intervention on VV.
  • Intervention on VV. Averaging 100 runs removes almost all of VV but lowers the Brier score by only 0.003 to 0.004, and lowering the temperature leaves the single-run SD nearly unchanged. Both results agree with the decomposition: VV is a small share of the reducible error. Appendix E also notes that the single-run decomposition gives the error of any kk-run average, V/k+E+AV/k+E+A, without further runs.
  • Intervention on EE. The reasoning-effort comparison is now judged by the Brier score and NLL as well as the AUC. Raising the effort lowers the Brier score on CB and RPP. Appendix E estimates the change in EE without using the proxy and reports the per-action epistemic component at both effort levels on RPP (Table 10). The SSRP result (n=21n=21) is reported as inconclusive.
  • Action-level intervention. A new experiment introduces spurious statistics into the statistical-review action on RPP, with gpt-5.4-mini and Qwen3.8-Flash (Figure 5, Table 11). The epistemic component of that action doubles, while those of the other actions barely change, and the chain attributes less of the run-to-run variability to the altered action.

2. The proxy and the Tier-3 check (Sections 3.2 and 5.4, Appendix E)

  • The Tier-3 check now has an explicit pass/fail criterion. The proxy passes on RPP and fails on CB and SSRP, and the verdicts are the same for any threshold from 0.010 to 0.025.
  • A new paragraph in Section 5.4 states which results depend on the proxy. The variability estimates and the epistemic component of the extraction action do not. The Limitations state that the split between EE and AA is supported only on RPP.
  • Section 3.2 clarifies that the proxy is an audit tool, also when it is applied to a new study at deployment; it does not replace the LLM's prediction.

3. Robustness

  • The analysis on RPP is repeated with the open-weights model Qwen3.8-Flash, which shows the same VEA patterns (Appendix E).

  • A sensitivity analysis shows that excluding the 55 CB claims without a scoreable extraction action leaves the action attribution essentially unchanged (Appendix C, Table 6). 4. Presentation

  • Run and action are defined in the Abstract and Introduction, all symbols in Figure 1 are defined, and Contribution (C2) is rewritten in plain language.

  • Section 4.1 now explains how the tail bounds connect to the VEA decomposition.

  • Section 3.2 explains that each action has its own reference for its components, and that the Tier-3 check applies to the scoring action only.

5. Claims aligned with the evidence

The Abstract and Introduction now state that the epistemic component dominates the reducible part of the prediction error, not the total error. The phrase "clearly supported by strong empirical evidence" has been removed. The Limitations now state that the main results use one model family and three datasets and that several effects are small.

We hope these revisions address the reviewers' concerns, and we would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.