acceptodds
Under review as a conference paper at ICLR 2027

What Should Black-Box LLM Confidence Predict? Estimating the Correctness of a Fixed Answer from Multiple Signals

Abstract

Black-box confidence methods score an answer by asking the model, resampling it, perturbing the input, or consulting other models, but need not score the same object. We freeze each question–answer pair before measurement and ask what survives new models and questions. Signals derived from the target itself, self-report and resampling, mostly measure commitment, including commitment to stable errors. Other models supply external corroboration. Under simultaneous model and question holdout, a small head that fuses self-report with peer evidence beats retrained text encoders and a released correctness model, and a frozen text encoder given the peer measurements nearly recovers the fusion. Asking the same peers to judge the frozen answer reads more of that evidence than counting their agreement on five of six datasets. A monotone environment link transfers the ranking while the probability scale stays local; peer value is governed by the panel's separation between right and wrong target answers; and the correctness of a unanimous consensus leaves the unlabeled agreement law unchanged. Separation predicts that stale agreement ages when an upgrade erodes it. Three frontier models showed it, the second time in a preregistered replication: the same previous-generation peers weakened whether counted or asked to judge, refreshed peers restored both readouts, above self-report at a moderate gap and indistinguishable from it at a wide one, and pooled stale peers diluted it. Freeze the answer, read capable peers as judges, refresh them when the target moves, audit the scale locally, and never equate unanimity with truth.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Confidence Estimation, Calibration, Selective Prediction, Model Agreement

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_NMF1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the confidence estimation problem for LLMs in the black-box settings. It seems to proposing training an output head on a held-out set to improve confidence estimation.

Critical strengths

None

Critical weaknesses
  1. The paper is very poorly written and it is very difficult to follow. Numerous important terms such as "survive", "commitment", "corroboration", "monotone environment link" were introduced without clear definitions.
  2. In addition to the poor writing, the three contributions (ln54-76) are either method descriptions or empirical observations, with no emphasis on the actual value related to the research around confidence estimation.
  3. Methods such as the fixed-answer framework were poorly introduced with little to none explanation or justifications to the design choices.
Questions for authors

None

3 comments
AuthorsOpenReview ↗

Reply to Reviewer NMF1

Thank you. Your points on undefined terms, the value of the contributions and the fixed-answer design guided the revision. If you can point us to the claim or experiment behind your soundness concerns, we will answer it during the discussion.

W1. Terms were introduced without definitions. We replaced two terms with plain words and now define the other two at first use.

  • "Survive" is gone. The text now says a signal keeps working on new models and new questions.
  • "Commitment" is how firmly the model being scored holds its answer, right or wrong. We call that model the target. Its stated confidence (self-report) and its resampled answers mostly measure commitment.
  • "Corroboration" is how often other models agree with the target's answer. We call those models peers.
  • The "monotone environment link" is now the plain condition that higher scores still mean a higher chance of being correct (contribution 4).

In the revision, the definitions are in the abstract, Sec. 1 and the contributions, marked in blue.

W2. The contributions do not state their value for confidence estimation. Each revised contribution now states its value. The main value is that a new model's answers can be ranked before any of them is labeled. This matters because deployed models are replaced every few months.

Contribution (pp. 2-3)Value for confidence estimationEvidence
1. What carries overmost of the gain over the model's stated confidence (self-report) comes from other models, with no labels from the new modelAUROC .938 / .933 / .859 on GSM8K / TriviaQA / MMLU-Pro, against .804 / .774 / .756 for self-report alone (Table 1)
2. When it stops, and why agreement hides itan exact identity for how much peer agreement can tell. A peer does not need to be accurate, but it must agree with right answers more often than with wrong ones (separation). The agreement rate does not show when that stops, so the rule replaces old peers at every upgradeagreement rate .517 to .510 while separation fell .278 to .044 (App. N), and Prop. 2 for unanimous answers
3. The prediction, tested at a model upgradethe prediction tells which upgrades break peer evidence and how to repair it, by replacing the peers and keeping self-reportobserved on GPQA-Diamond, then predicted on MMLU-Pro, where five of five predictions held (Table 3)
4. What a repair coststhe labels and calls each use needs are known in advanceno labels to rank, about 20 labels to recalibrate the probabilities and about 100 for a threshold with a guaranteed error rate, with six calls per answer (Sec. 6)

Compared on the same answers, our output head ranks first, or tied for first, against eight published methods, listed with their venues in point 1 of the general response. Point 3 of the general response sets out what is new.

The four revised contributions are on pp. 2-3, marked in blue.

W3. The fixed-answer framework and its design choices lacked justification. The fixed-answer framework has every method score the same answer, the one the user sees. So all of them estimate the quantity a deployment needs, which is whether that answer is correct (Eq. 1).

Design choiceProblem it preventsSupport
Fix the answer firsta score could rate an answer other than the one used, so a better score could come from a better answerbelow (App. B)
Read the target's own signals separately from the peers' signalspooling them would hide that the target's own signals stay confident on errors it repeatsresample agreement is .861 when all five models give the same wrong answer, against .712 on other errors (Sec. 6)
Hold out the model and the questions togethera random split lets the same model or question appear in training and test, which overstates transfera deployment meets models and questions that are both new (Secs. 1, 2.3)
Keep the head minimalgains could come from the head's capacity rather than the signalsa text encoder given the same measurements nearly matches the head (App. G)

Without fixing the answer, a resampling score can rate an answer the user never sees. In our runs the most frequent resampled answer is not the answer the user sees for 16% of GSM8K and 6% of MMLU-Pro answers. On those answers, a score that reads the most frequent resample reports a confidence of about .53 on GSM8K and .65 on MMLU-Pro. The answer the user sees is right only 12% and 28% of the time (App. B).

Sec. 1 now gives the reason for each choice, and App. B reports the resample mismatch.

Your summary reads the method as training an output head on a held-out set. We train a small output head on other models and questions, not on the held-out set. No label from the evaluated model enters training. It is tested on the held-out models and questions.

The revision states this split in contribution 1 on p. 2 and in Sec. 2.3.

Thank you for the response. My main concern on the presentation quality still remains and I will keep the original score.

AuthorsOpenReview ↗

Reply to Reviewer NMF1 (follow-up)

Thank you for reading our response and for telling us where you stand. The revised PDF, posted before your comment, already includes a change for each of your comments about presentation, and it marks the main-text changes in blue.

  • Terms. "Survive" and the "monotone environment link" are replaced by plain words, and commitment and corroboration are defined at first use (pp. 1-3).
  • Value of the contributions. Each of the four contributions now states its value for confidence estimation (pp. 2-3).
  • Design choices. Sec. 1 now gives the reason for each choice, including fixing the answer before scoring it.
  • Plain writing. The abstract, Sec. 1, the Figure 1 caption and the conclusion are rewritten in plain terms (pp. 1-3, 10), and the appendix is reworded for readability.
  • Headings and tables. The headings of Secs. 3-6 state their results. Sec. 6 has a table of the labels and calls each use needs, and Sec. 7 opens with a table that compares each result with the closest earlier work.

For the final version we will also add a worked example that follows one question from the model's answer to its final score. We cannot tell which parts still fall short for you, so one example of a passage that is hard to follow would let us fix it directly.

For reference, here are the paper's two main results. Both can be checked from their numbers, without relying on the paper's wording.

1. A simple score carries over to new models without their labels. Our score, a small logistic regression we call the head, is trained only on other models and questions. Every method is then tested on the same answers of held-out models and questions, and none uses labels from the model being scored. The head ranks first against eight published methods on GSM8K, TriviaQA and MMLU-Pro, tied with cross-model TU on TriviaQA (AUROC, as in point 1 of the general response).

Method (venue)GSM8KTriviaQAMMLU-Pro
Our head, A+C.938.933.859
Cross-model total uncertainty, TU (Hamidieh et al., ICLR 2026), as published.522.933.500
Cross-model TU, our improved version.861.933.770
Perturbation composite (Pedapati et al., TMLR 2025), five of six families.901.848.750
Correctness model, 8B (Xiao et al., ICML 2026), as released.808.863.714
APRICOT (Ulmer et al., ACL 2024), better of two variants.630.511.555
Calib-n (Xia et al., ACL 2025), our extension to unseen models.610.671.618
Fine-tuned BERT (Devlin et al., NAACL 2019).606.620.669
Stated confidence (approach of Tian et al., EMNLP 2023, and Xiong et al., ICLR 2024).804.774.756
Semantic entropy (approach of Kuhn et al., ICLR 2023, and Farquhar et al., Nature 2024).805.726.650

The perturbation composite is scored only where every perturbation is valid, and A+C keeps its lead on those answers. Where a baseline could be made stronger, we used the stronger version, as the two cross-model TU rows show.

2. An exact rule for when other models help. Suppose another model, a peer, agrees with the right answers of the model being scored, the target, at rate π\pi and with its wrong answers at rate cc. The information one peer adds is the Jeffreys divergence between these two rates, and it equals exactly (π−c)L(\pi-c)L, where L=log⁡π(1−c)c(1−π)L=\log\frac{\pi(1-c)}{c(1-\pi)} is how far one agreement moves the log-odds (Eq. 2). So a peer does not need to be accurate, but it must agree with right answers more often than with wrong ones. We call the gap π−c\pi-c separation.

A much stronger target can close this gap, since old peers cannot agree with right answers they themselves get wrong. The agreement rate mixes separation with the target's accuracy ρ\rho, as r=c+ρ(π−c)r=c+\rho(\pi-c), so it can stay flat while separation falls. The paper therefore predicts that old peers stop helping at a model upgrade while their agreement rate barely changes. After first seeing this on GPQA-Diamond, we registered five predictions before collecting the MMLU-Pro data, and all five held (Table 3). Earlier work found that the value of other models depends on their capability (Hamidieh et al., ICLR 2026). Another study found that a judge's true-positive minus false-positive gap governs aggregate evaluation with imperfect judges (Feng et al., ICLR 2026). Neither setting asks how much one peer's agreement says about a single answer of a new model, which the identity above answers exactly (point 3 of the general response).

Our question. Beyond the three points of your review, is there a claim or an experiment you consider incorrect, or a result you consider already known? If so, a pointer or a reference would let us answer it during the discussion. We would be grateful to learn where, in your view, these results still fall short.

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

The paper finds that a model's own confidence and agreement from other models provide complementary information. Combining them can predict answer correctness more accurately than several existing methods, including on unseen models and questions. Asking peers to judge a fixed answer is often better than counting agreement. Peer-based confidence estimates become unreliable when the peers are substantially weaker than the model being evaluated.

Critical strengths
  1. Great problem to work on; verification is obviously very challenging and important to figure out.
  2. Nice robust experimental setup, with many benchmarks and different existing methods. There’s practical guidance with the result that checking a model’s answer against other models only helps if those models are about as capable as the one being checked (shown on GPQA and on MMLU pro).
Critical weaknesses
  1. The paper was hard to read, unfortunately. For example, the abstract uses the type of terms that only are understandable later in the paper. It’s hard to parse these: “a monotone environment link transfers the ranking while the probability scale stays local”, “the correctness of a unanimous consensus leaves the unlabeled agreement law unchanged”.

  2. The method relies on having peer models that are relatively close in capability to the target model. Although the authors acknowledge this limitation, I wonder how useful the approach would be for frontier models, where equally capable peers may not always be available.

  3. A simpler method in the paper beats the method the paper recommends. The main method (A+C) combines the model’s own confidence with how many other models give the same answer.

Questions for authors
  1. A single Haiku 4.5 judge beats A+C on every dataset where it was tested. When should someone use A+C instead?
  2. The frontier test covers one generation change and three targets. Does the stale-peer result hold for other model upgrades?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer tFYc

Thank you for the questions on the Haiku 4.5 judge, frontier models and other upgrades, and for flagging the hard-to-read abstract. You call the peer-capability result practical guidance, and Reviewer MZgZ names the same result as a strength. The frontier test bears it out. Peers that fall far behind the model they check stop helping, and current peers restore the gain without new labels.

Q1 and W3. A single Haiku 4.5 judge beats A+C. When should someone use A+C? You are right that Haiku 4.5 alone beats A+C on all four datasets where it was run. A strong judge is a good signal, and it is even better on top of A+C. Scores here are AUROC, the chance that a right answer scores above a wrong one. Our rule is the following.

SituationScore to useWhy
Calls for a verdict are affordableA+C plus a judge's verdict (a strong model or the four peers)higher than the judge alone in 15 of the 16 judge-dataset cells of Table 11, and equal in the other
No calls beyond A+C's ownA+Csix calls per answer, under a cent (App. F)
You want a judge alonea judge alone, only after a check against labels from the new targetat the frontier upgrade Haiku 4.5 alone fell to .617 on the new models' GPQA answers, below their own stated confidence (self-report) at .784. A+C keeps self-report and replaces old peers, while a judge alone has nothing to fall back on (App. O)

The judge does not have to be an extra model. A+C's own four peers can judge each answer, at four extra calls. As judges on GPQA they reach .845, about the same as Haiku 4.5 alone at .841 (Table 11).

Sec. 6 now states this rule.

W2. Is the approach useful for frontier models without equally capable peers? Yes, with self-report as the base. The target's own stated confidence kept working at every capability gap we tested, and at the GPQA upgrade its AUROC was .784 (Table 3). Current peers can add to it. On MMLU-Pro they raised the pooled AUROC from .691 for self-report alone to .778, mainly for the target whose self-report was weakest (App. O).

Peers do not need equal capability, as point 2 of the general response shows. So a peer does not need to be accurate, but it must agree with right answers more often than with wrong ones. This gap is separation. At the GPQA upgrade, separation fell from .278 to .044 while the agreement rate moved only from .517 to .510 (App. N). The old peers' agreement then ranked near chance, and current peers restored separation. Weaker peers can still help, as Xue et al. (2026) and Hamidieh et al. (2026) saw on TriviaQA. Separation says exactly when they stop. Our rule replaces the peers with current models whenever the target changes and always keeps self-report.

The rule to replace the peers is new in Sec. 4.3.

Q2. Does the stale-peer result hold for other model upgrades? Yes, in a capability upgrade that we simulated on the paper's stored runs (App. U). In the theory, old peers fail because of the capability gap and not because of the release date. So the simulation splits the models by accuracy. The four least accurate models supply the training data for the score and serve as the old peers. The four most accurate are the targets, and each target's current peers are the other three. We fixed this split before computing any AUROC, and the score is trained with no label from the stronger models.

Change in AUROC on the same answers (Table 24, four targets pooled)GSM8KMMLU-ProTriviaQA (smallest accuracy gap)
Replace old peers with current peers (agreement score alone)+.189+.170+.016
Use A+C with old peers instead of self-report alone-.013-.045+.159
Add old peers to current peers (agreement score alone)-.031-.056+.011

On GSM8K and MMLU-Pro, replacing the peers helps, old peers hurt A+C, and mixing old peers in dilutes the current ones. All three effects appear for each of the four targets, and their pooled 95% intervals exclude zero. The gain from replacing the peers grows with the accuracy gap between old and new models, with a Spearman correlation of .83. On TriviaQA, where the gap is smallest, old peers still help, as the theory predicts.

The new App. U gives the design and the per-target results.

W1. The paper was hard to read. We agree. The two phrases you quote, on the "monotone environment link" and on "the correctness of a unanimous consensus", are now two plain sentences of the abstract. The first reads "Ranking therefore needs no labels from the new model as long as higher scores still mean a higher chance of being correct." The second reads "A small labeled sample still checks and calibrates the ranking and shows how often unanimous answers are right, which agreement alone cannot."

In the revision, the rest of the abstract, Sec. 1, the contributions and the conclusion use the same plain terms, defined at first use (pp. 1-3, 10).

Thanks for the detailed response and for running the additional experiments! I'll keep my original score of 3.

AuthorsOpenReview ↗

Thank you to Reviewer tFYc

Thank you for reading our response and for your positive assessment. We are glad the additional experiments were useful, and your questions made the paper stronger. The revised PDF already includes the analyses we added for them, the simulated upgrade (App. U) and the per-target upgrade results (App. O). The final version will keep these results and the clarifications from this fruitful discussion.

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

This paper studies how to predict whether an LLM's answer is correct using only black-box confidence.

The authors study three types of signals: self-reported confidence, sampling consistency, and agreement with other LLMs.

They find that self-report and sampling consistency mainly reflect the model's commitment to an answer, while peer agreement provides additional evidence about correctness.

The paper proposes a simple method that combines self-report and peer evidence to improve correctness prediction across different models and datasets.

It also shows that peer agreement becomes less reliable when the target model is much stronger than its peers, and that updating the peer models can improve performance.

Critical strengths
  • The paper evaluates different confidence signals across multiple LLMs and datasets. These experiments provide useful evidence about the generalization of confidence methods
  • The paper shows that peer agreement becomes less reliable when the target model is stronger than its peers. This helps us understand the limitations of peer-based confidence estimation.
Critical weaknesses
  • The proposed method mainly combines existing confidence signals using a simple regression model, which lacks novelty
  • Several key findings, such as the fact that self-consistency does not guarantee correctness and that peer agreement depends on peer model capabilities, are already known.
  • This work explains why peer agreement works or fails, but the method requires correctness labels from the target model, which limits the practical value
  • The peer-based confidence requires multiple runs, which introduces additional inference cost
Questions for authors
  • In Section 4.3, the proposed assessment requires correctness labels from the target model. How can we determine whether peer agreement is reliable for a new target model without knowing the correctness of its answers?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer MZgZ (1/2)

Thank you for naming two strengths, and for your questions on target labels, novelty and inference cost. You write that the experiments are "useful evidence about the generalization of confidence methods". This evidence comes from running together two tests that earlier work ran separately. It shows that on new models and new questions, most of the gain over self-report comes from the agreement of other models, which we call peers. Your second strength is the paper's central prediction, that peer agreement "becomes less reliable when the target model is stronger than its peers". Reviewer tFYc names the same result as practical guidance. To our knowledge, no earlier work has tested this at an actual model upgrade with predictions registered before the data were collected. The revision now states this prediction on page 1.

W1. Is the method a simple regression over existing signals, and does it lack novelty? The regression is simple on purpose, and the signals are known. What is new is knowing which signals to combine, and why. The paper shows that a model's own signals mostly measure how firmly it holds its answer, right or wrong. Information about correctness that these signals miss must come from elsewhere, and peer agreement supplies it. A logistic regression on self-report and peer agreement then beats much larger trained models, including fine-tuned text predictors up to an 8B model, on every dataset in the general response table. Its AUROC is .938 / .933 / .859 on GSM8K, TriviaQA and MMLU-Pro, against at most .808 / .863 / .714 for any fine-tuned predictor. A text encoder given the same measurements nearly matches it (Table 10). So the gain comes from choosing the right measurements (point 3 of the general response).

As for the regression itself, we respectfully see its simplicity as the method's main strength. It needs no GPU training, each gain can be traced to a signal, and it uses only answers and stated confidence, so it works with any black-box API, including the many that return no token probabilities or hidden states. Beyond the regression, Prop. 1 says when the ranking carries over without labels, and Prop. 2 says which questions the few labels must come from.

In the revision, Sec. 1 explains why the head is minimal.

W2. Are these findings already known? Yes, and the paper builds on both. It explains the first, since a model's own signals measure how firmly it holds its answer (W1). It turns the second into an exact identity and a rule, shows why a failure stays hidden, and tests the rule at a real upgrade.

  • An exact identity. A peer agrees with the target's right answers at rate π\pi and with its wrong answers at rate cc. In the classical agreement model (Dawid and Skene, 1979), each agreeing peer moves the log-odds that the answer is right by the same amount, so the count of agreeing peers carries all of their evidence. The information one peer adds is the Jeffreys divergence between these two rates, and it equals exactly (π−c)L(\pi-c)L, where L=log⁡π(1−c)c(1−π)L=\log\frac{\pi(1-c)}{c(1-\pi)} is how far one agreement moves the log-odds (Eq. 2). So a peer does not need to be accurate, but it must agree with right answers more often than with wrong ones. The gap π−c\pi-c, which we call separation, enters twice, so a peer's information shrinks with roughly the square of the gap. Across the thirty source environments, larger separation goes with better ranking from agreement, and the frontier upgrade moves along the same curve (Figure 3).
  • How this differs from Feng et al. (ICLR 2026). Feng et al. use the same gap, a judge's true-positive minus false-positive rate, to decide when testing a model's overall accuracy with an imperfect judge beats testing it on labels. Our question is one their setting does not ask, namely how much each peer's agreement says about one answer of a new model with no labels from it. For that question the gap sets the information exactly, and it can vanish while the agreement rate stays flat.
  • Why the failure is hidden. The agreement rate can stay flat while separation falls (Q1 in our reply 2/2).
  • A test at a real upgrade. The failure appeared first on GPQA-Diamond. We then registered five predictions for MMLU-Pro before collecting the data, and all five held. Current peers restored the signal without new training labels (Table 3).

Secs. 1 and 7 now state these additions.

AuthorsOpenReview ↗

Reply to Reviewer MZgZ (2/2)

W3. Does the method need correctness labels from the target model? No. Ranking needs no labels from the new model, and the head comes out first among the methods we compared both without labels and with a few.

  • No labels from the new model. The head (A+C) ranks the new model's answers first on all three datasets in the general response table, ahead of four retrained text encoders and a released 8B correctness model. Its AUROC is .938 / .933 / .859, tied with cross-model TU on TriviaQA.
  • A few labels. When every method gets the same 20, 80 or 120 labels, the head has the lowest log loss in all 15 model-dataset pairs (table below). About 20 labels recalibrate its probabilities, and about 100 set a threshold with a guaranteed error rate (Sec. 6).

It is also cheap and easy to run on black-box LLMs. It uses only answers and stated confidence, makes six calls per answer for under a cent, and needs no GPU training (W4 below).

Labels from the new model, the same for every methodHead (A+C)Best retrained encoderCorrectness model
20 labels, median log loss over 15 model-dataset pairs (lower is better, App. K).154.559.546
80 labels.151.544.510
120 labels.149.534.505

In the revision, Sec. 6 has a new table of the labels and calls each use case needs.

W4. How much inference cost do the extra calls add? The head uses six model calls per answer, which cost 0.4 to 0.9 cents for the five targets on the 512-question datasets. Several calls are normal for black-box confidence. Resampling methods such as semantic entropy make eight extra calls to the target, more than our six, and still rank worse on all three datasets (.805 / .726 / .650, Table 1). Our six calls are one self-report prompt, four peer answers and one call that groups the answers by meaning. Most of the cost is the target's own self-report. This version ranks as well as the full head, which uses three self-report prompts. Its AUROC on GSM8K, TriviaQA and MMLU-Pro is .938 / .933 / .857, against .938 / .933 / .859 for the full head. Adding the peers' verdict, which ranks even better, takes four more calls. The text encoders make no calls when they score, but they need GPU fine-tuning on thousands of labels from the deployed model.

For example, take a workflow in which an agent handles cases and an expert reviews the uncertain ones. An expert who spends 30 minutes on a case at 60 dollars an hour costs 30 dollars per case, more than 3,000 times the cost of the added calls. Even if each case had 100 times our context length, the calls would cost 90 cents per case. The added calls pay for themselves once the score lets experts skip one case in thirty. On TriviaQA, with 120 labeled answers, a threshold certified to keep errors under 5% among the answers it accepts lets experts skip about half of the answers (App. Q). For a queue of 10,000 such cases, experts review about 4,700 cases for 140,000 dollars instead of all 10,000 for 300,000 dollars. The calls cost 90 dollars at our context length, or 9,000 dollars at 100 times that. Labeling the 120 check cases costs 3,600 dollars. So with the score, the queue costs about half as much as reviewing every case.

App. F lists the calls and their cost, and the new table in Sec. 6 lists the calls each use case needs.

Q1. Without labels, how can one tell whether peer agreement is still reliable for a new target? The method does not need this check, and our theory explains why the agreement rate alone could not provide it. Our rule replaces the peers with current models whenever the target changes and always keeps self-report. Self-report kept working at every capability gap we tested. With no label from the new models, this rule worked at the frontier upgrade. On MMLU-Pro, A+C with current peers reached an AUROC of .778, against .691 for self-report alone and .683 with the old peers (Table 3).

The agreement rate alone could not do this job. A peer agrees with the target's right answers at rate π\pi and with its wrong answers at rate cc, and separation is the gap π−c\pi-c. The agreement rate mixes separation with the target's accuracy ρ\rho, as r=c+ρ(π−c)r=c+\rho(\pi-c), so it can stay flat while separation falls. At the GPQA upgrade, separation fell from .278 to .044 while the agreement rate moved only from .517 to .510 (App. N). As the theory predicts, two tests that use no labels, one registered in advance, did not detect the stale peers.

A few labels, where available, add two things. A small sample confirms that higher scores still mean a higher chance of being correct, and it shows how often unanimous answers are right. The second needs labels whenever only the models' answers are seen and they can share misconceptions, because the agreement counts then look the same whether a shared answer is right or wrong (Prop. 2).

Sec. 4.3 now states these rules for a new target.

AuthorsOpenReview ↗

Follow-up to Reviewer MZgZ

Thank you again for your review. You named two strengths, the evidence on how confidence methods generalize and the finding that peer agreement becomes less reliable when the target is stronger than its peers. Reviewer tFYc also called the second practical guidance. For each of your concerns, here is what we did and what it showed, starting with labels.

  • W3 and Q1, the need for labels from the target. Our score, which we call the head, is a small logistic regression on the target's stated confidence and on peer agreement. Without any label from the new model, it already ranks first against eight published methods on GSM8K, TriviaQA and MMLU-Pro, tied on TriviaQA (point 1 of the general response). A few labels then make it better. About 20 labels recalibrate the probabilities and also check the ranking. With the same 20, 80 or 120 labels for every method, the head still has the lowest log loss in all 15 model-dataset pairs (App. K, added for this point). For a new target, the method does not need to test whether the peers are still reliable. It replaces old peers with current models at every upgrade and always keeps the target's stated confidence. At the frontier upgrade, this raised the head's AUROC on MMLU-Pro from .683 with the old peers to .778, with no new labels (Table 3). The revision now states in the abstract and Sec. 6 that ranking needs no labels, and Sec. 4.3 gives the rule for a new target.
  • W1, a simple regression without novelty. The regression is simple on purpose, and we see this as its main strength. A logistic regression with at most eleven parameters beats every fine-tuned predictor in the comparison, up to an 8B correctness model, on every dataset. Its AUROC on GSM8K, TriviaQA and MMLU-Pro is .938 / .933 / .859, against at most .808 / .863 / .714 for the fine-tuned predictors (point 1 of the general response). What makes this possible is knowing which signals to combine. A model's own signals mostly measure how firmly it holds its answer, so information about correctness has to come from other models. A text encoder given the same measurements nearly matches the regression, so the gain comes from the measurements and not from the size of the predictor (Table 10). Sec. 1 now explains this choice.
  • W2, findings already known. We agree that both findings are known. W1 above gives the reason for the first. For the second, the paper adds its cause and a test of it at a real upgrade. How much one peer's agreement tells us depends only on how often it agrees with right answers and with wrong ones (Eq. 2). It tells nothing when the two rates are equal, whatever the peer's own accuracy. We call the gap between the two rates separation. At the GPQA upgrade, separation fell from .278 to .044 while the agreement rate moved only from .517 to .510, so the agreement rate did not show the fall (App. N). At the frontier upgrade, five predictions registered before collecting the MMLU-Pro data all held (Table 3). Secs. 1 and 7 now state these additions, and Sec. 7 opens with a table that compares each result with the closest earlier work.
  • W4, inference cost. Without access to a model's internals, any score that must work on a new model has a cost somewhere, in extra calls, in a large model of its own, or in labels and training. The head has only a small, fixed call cost and needs no training or labels to rank. It makes six calls per answer, which cost 0.4 to 0.9 cents (App. F), and a new table in Sec. 6 lists the calls each use needs. The other methods have costs too. Those that read the target's behavior or other models' answers make extra calls. For example, semantic entropy makes eight extra calls to the target and still ranks lower (Table 1). Those that read only the text pay in model size or training. The released correctness model runs an 8B model on every answer, and the original recipes of the fine-tuned text encoders need GPU fine-tuning on thousands of labels from the deployed model. None of them ranks above the head on new models.

We know this is a busy time, and we appreciate your attention. As the discussion closes soon, we would be glad to hear which point, if any, is still open, so that we can answer it in time. If these answers address your concerns, we would be grateful if you would take them into account in your final rating.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response (1/3)

We thank all the reviewers. Reviewer tFYc found the problem important, the setup "robust" with "many benchmarks and different existing methods", and the peer-capability result "practical guidance". MZgZ found the experiments "useful evidence about the generalization of confidence methods".

All three reviews rated the presentation low. We agree. Our original presentation also understated two properties of our results. Below, we call the model being scored the target, its stated confidence its self-report, the other models its peers, and our score the head (A+C).

First, we compare the three kinds of black-box signals on the same answer and on models and questions that are both new. The three are self-report, the target's resampled answers and the peers' agreement. Earlier studies did one or the other (Bouchard & Chauhan, TMLR 2025, and Xiao et al., ICML 2026). Doing both shows which signal still works on a new model, and why. The target's own two signals mostly measure how firmly it holds its answer. So they stay confident when it repeats a mistake, and the information about correctness has to come from the peers. The head, a logistic regression on self-report and peer agreement, ranks a new model's answers without any of its labels, and it beats much larger trained models, including an 8B correctness model. Its AUROC is .938 / .933 / .859 on GSM8K, TriviaQA and MMLU-Pro, against at most .808 / .863 / .714 for any fine-tuned predictor (point 1; MZgZ, NMF1).

Second, earlier work found that the value of other models depends on their capability (Hamidieh et al., ICLR 2026). We turn this into an exact identity and a tested rule. Treat each peer as a noisy judge. A peer agrees with the target's right answers at rate π\pi and with its wrong answers at rate cc. The information one peer adds is the Jeffreys divergence between these two rates, and it equals exactly (π−c)L(\pi-c)L, where L=log⁡π(1−c)c(1−π)L=\log\frac{\pi(1-c)}{c(1-\pi)} is how far one agreement moves the log-odds (Eq. 2). So a peer does not need to be accurate, but it must agree with right answers more often than with wrong ones. We call the gap π−c\pi-c separation. A much stronger target can close this gap while the agreement rate stays the same, so that rate gives no warning.

Our rule replaces the peers with current models whenever the target changes and always keeps self-report. We tested it at a frontier model upgrade. There, the old peers and every score built on old models, including a strong judge, ranked the new target's answers worse than its own self-report. Current peers restored the gain without new labels. On MMLU-Pro, all five predictions we registered in advance held (point 2; MZgZ, tFYc). MZgZ named this result as a strength, writing that peer agreement "becomes less reliable when the target model is stronger than its peers".

The revision puts these two results first and marks changes in blue. The answers to the main questions are these.

  • Does it need labels from the new model? (MZgZ, NMF1) No, ranking needs none. About 20 labels recalibrate the probabilities, and about 100 set a threshold with a guaranteed error rate (point 1, Sec. 6).
  • What is new? (MZgZ, NMF1) The signals and the regression are known. What is new is the exact identity for how much one peer can tell, the reason a failure stays hidden, the test of the rule at a real upgrade, and a proof of what agreement cannot tell about unanimous answers (point 3).
  • Which score should one run? (tFYc) A+C is the cheap default, at six calls per answer for under a cent. When more calls are affordable, adding the peers' verdict ranks even better. A judge used alone needs a labeled check first (point 4).
  • What about frontier models and other upgrades? (tFYc) Self-report kept working at every capability gap we tested, current peers add to it, and a second, simulated upgrade shows the same pattern (reply to tFYc).
  • Was the paper hard to read? (all three) Yes. The abstract, Sec. 1, the contributions and the conclusion are rewritten in plain terms, each defined at first use.
AuthorsOpenReview ↗

General response (2/3)

1. Does it need labels from the new model, and how does it compare with other methods? (MZgZ, NMF1) It needs no labels to rank, and its AUROC is the highest on all three datasets below, tied with cross-model TU on TriviaQA. Every method, ours included, scores the same answer to each question, the one the user sees. None uses labels from the model being scored.

Method (venue)What it readsTraining and labelsGSM8KTriviaQAMMLU-Pro
Our head, A+Cself-report (A) and four peers' agreement (C)logistic regression, at most eleven parameters, trained on other models and questions.938.933.859
Cross-model total uncertainty, TU (Hamidieh et al., ICLR 2026), as published (App. R)the target's resamples and other models' answers, by text similaritynone.522.933.500
Cross-model TU, our improved versionthe same, with answers grouped by meaningnone.861.933.770
Perturbation composite (Pedapati et al., TMLR 2025), five of six familiesthe target's answers to perturbed questionslogistic regression, trained on other models and questions.901.848.750
Correctness model, 8B (Xiao et al., ICML 2026), as releasedquestion and answer text8B model fine-tuned on eight models' past correctness.808.863.714
APRICOT (Ulmer et al., ACL 2024), better of two variantsquestion and answer textDeBERTa fine-tuned on other models and questions.630.511.555
Calib-n (Xia et al., ACL 2025), our extension to unseen models (App. F)the question with several models' answerstext encoder fine-tuned on other models and questions.610.671.618
Fine-tuned BERT (Devlin et al., NAACL 2019)question and answer textBERT fine-tuned on other models and questions.606.620.669
Stated confidence (approach of Tian et al., EMNLP 2023, and Xiong et al., ICLR 2024)the target's stated confidence, three promptsnone.804.774.756
Semantic entropy (approach of Kuhn et al., ICLR 2023, and Farquhar et al., Nature 2024)eight resamples of the targetnone.805.726.650

The perturbation composite is scored only where every perturbation is valid, and A+C keeps its lead on those answers. Judges are compared separately (Table 11), and point 4 says when to use one.

App. V now holds this table.

2. When do peers stop helping, and how is it fixed? (MZgZ, tFYc) Peers stop helping when the target is much stronger than they are. The paper's theory predicts this, the frontier upgrade confirmed it, and replacing the peers with current models restored the gain. On MMLU-Pro, A+C reached an AUROC of .778 with current peers, against .683 with the old ones (Table 3).

The reason is separation, the gap π−c\pi-c between how often a peer agrees with the target's right answers (π\pi) and with its wrong answers (cc). Part 1 gives the identity that makes it decisive. In the classical agreement model (Dawid and Skene, 1979), each agreeing peer moves the log-odds that the answer is right by the same amount, so the count of agreeing peers carries all of their evidence. Intuitively, a much stronger target answers correctly many questions its old peers get wrong. On those questions the old peers cannot agree with its right answer, so π\pi falls and the gap closes.

The agreement rate mixes separation with the target's accuracy ρ\rho, as r=c+ρ(π−c)r=c+\rho(\pi-c), so it can stay flat while separation falls. At the GPQA upgrade, separation fell from .278 to .044 while the agreement rate moved only from .517 to .510 (App. N).

The headings of Secs. 4 and 5 now state this, and App. U adds a second test.

AuthorsOpenReview ↗

General response (3/3)

3. What was known, and what is new? (MZgZ, NMF1) The signals and the regression are known. The table shows what the paper adds on each question.

QuestionEarlier workThis paper
What is scored?how stable a model's answers are (Kuhn et al., 2023, and Farquhar et al., 2024), or correctness learned from in-domain labels (Ulmer et al., 2024, and Xia et al., 2025)the chance that the exact answer the user sees is correct, on models and questions never seen (Eq. 1)
Which signal carries over to a new model?several signals combined, but tuned and tested within each model and dataset (Bouchard & Chauhan, 2025), or one trained scorer tested on new models (Xiao et al., 2026)the three kinds of signals on the same answer, on new models and questions. Peer agreement carries most of the gain, and a logistic regression with no labels from the new model beats fine-tuned text encoders and an 8B correctness model (point 1)
When do other models help, and does it show?peer value depends on capability (Hamidieh et al., 2026), and weaker or fixed judges lose value on stronger models (Dorner et al., 2025, Zhou et al., 2026, and Singh et al., 2026). A judge's true-positive minus false-positive gap governs aggregate evaluation with imperfect judges (Feng et al., 2026)an exact identity for each single answer. One peer's information is (π−c)L(\pi-c)L, so only separation π−c\pi-c matters, not accuracy (Eq. 2). The agreement rate can stay flat while separation collapses. We predicted this for a model upgrade, and five of five registered predictions held (point 2)
What can agreement say about unanimous answers?models can be consistently and confidently wrong (Tan et al., 2025)when only the models' answers are seen and they can share misconceptions, no test without labels can tell how often a unanimous answer is right (Prop. 2). A few labels from those questions can

Sec. 7 now opens with a table of this comparison.

4. Which scorer should a user run? (tFYc) We recommend A+C as the default because it is cheap. Our strongest score adds the peers' verdict, which raises AUROC on all six datasets (Table 2) for four more calls per answer (App. G). So add the verdict when you can afford the calls. Adding a judge's verdict to A+C raised AUROC in 15 of the 16 judge-dataset cells (Table 11).

A judge used alone is riskier. At the frontier upgrade, Haiku 4.5 alone fell to an AUROC of .617 on GPQA, below the new models' own stated confidence at .784 (App. O). A+C keeps the model's own stated confidence and replaces old peers with current ones. A judge used alone has nothing to fall back on, so check it against a few labels from the new model first.

Sec. 6 adds this rule, and the conclusion ends with the whole deployment rule.

Analyses added for this discussion. Each answers a reviewer's question.

  • App. U, for tFYc, repeats the upgrade test with models split by accuracy. Replacing the peers helps more as the gap between old and new models' accuracy grows, with a Spearman correlation of .83.
  • App. K, for MZgZ, gives the text encoders the head's label budgets. The head has the lower log loss in all 15 model-dataset pairs at 20, 80 and 120 labels.
  • App. O, for tFYc, reports the MMLU-Pro upgrade for each target. Current peers add most for the target whose self-report is weakest.
  • App. B, for NMF1, finds that for 16% of GSM8K and 6% of MMLU-Pro answers the most frequent resample is not the answer the user sees.

Changes in the revision.

  • We rewrote the abstract, Sec. 1, the Figure 1 caption, the contributions and the conclusion around what keeps working when the model changes (pp. 1-3, 10).
  • Sec. 1 gives the reason for each design choice.
  • Section headings state their results, and terms are defined at first use.
  • We also reworded the appendix for readability. These edits are not colored.

A closing note. The reviewers' concerns were reasonable, and addressing them made the paper clearer and more useful. Many deployed LLMs are reached only through an API, which hides the model's internal signals. They are also replaced before anyone labels their answers. The paper gives a scorer that a deployment team can run as is, with API calls, a small logistic head and no GPU or fine-tuning. We kept the head small on purpose, so that each gain can be traced to a signal. It ranks a new model's answers before any of them is labeled, and its theory says when the peers must be replaced.

AuthorsOpenReview ↗

Summary of the discussion

We thank the reviewers and the Area Chair. The paper shows that a new black-box model can borrow its confidence from other models, by checking its answer against theirs, and it predicts and tests when this stops working. The reviewers found the problem important and the setup "robust" with "many benchmarks and different existing methods" (tFYc), the experiments "useful evidence about the generalization of confidence methods" (MZgZ), and the peer-capability result "practical guidance" (tFYc), a result MZgZ also named as a strength. During the discussion we added the analyses the reviewers asked for. The main results and their evidence are these.

  • A new model's answers can be ranked without its labels. Our score, a small logistic regression on the model's stated confidence and other models' agreement that we call the head (A+C), is trained only on other models and questions. It ranks first against eight published methods on all three datasets, tied on TriviaQA (general response, point 1). With the same 20, 80 or 120 labels, it also has a lower log loss than the retrained text encoders and the correctness model in all 15 model-dataset pairs (App. K, added for MZgZ).
  • The paper predicts when other models stop helping, and the prediction held. An exact identity gives how much one other model, a peer, can tell (Eq. 2). At a frontier upgrade, five predictions registered before collecting the MMLU-Pro data all held, and current peers restored the gain with no new labels (Table 3). A second, simulated upgrade shows the same pattern (App. U, added for tFYc). Sec. 7 opens with a table that compares each result with the closest earlier work.
  • The method is cheap to run. A+C makes six calls per answer for under a cent and needs no GPU training. When more calls are affordable, adding a judge's verdict raised AUROC in 15 of the 16 judge-dataset cells (Table 11, Sec. 6).
  • The presentation is revised (all three reviews). The abstract, Sec. 1, the contributions and the conclusion are rewritten in plain terms (pp. 1-3, 10). Terms are defined at first use, each contribution states its value, Sec. 1 gives the reason for each design choice, and the headings of Secs. 3-6 state their results. The final version will also add a worked example that follows one question from the model's answer to its final score.

Reviewer tFYc thanked us "for the detailed response and for running the additional experiments".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.