acceptodds
Under review as a conference paper at ICLR 2027

Dispersion Is Not Doubt: Rethinking LLM Confidence Estimation When Multiple Answers Are Valid

Abstract

Confidence estimators often treat concentrated support as evidence of correctness, relying—explicitly or implicitly—on the assumption that concentrated support implies a correct answer. When several answers are valid, however, this assumption fails: support can spread across correct alternatives without indicating error. We call this mismatch the concentration–correctness gap, and propose two layered conjectures. (a) The inversion conjecture: as the number of valid answers grows, confidence falls even as the model answers better. (b) The dispersion-is-not-doubt conjecture: the decline stems from support being re-divided among valid answers, not from greater uncertainty about correctness. To test them, we introduce MACE (Multi-Answer Confidence Estimation), open-ended questions with controlled answer cardinalities () and exhaustive, expert-audited reference sets. Across 9 models and 15 training-free estimators, both conjectures are confirmed: a multi-answer confidence inversion is widespread; all 15 estimators lose AUROC under pooled cardinalities; and valid mass stays stable on correctly answered questions while support is unevenly re-divided—entropy decomposition, elicited confidence, and cross-model analysis () jointly attribute the decline to valid-answer diversity. These findings motivate CARE (CArdinality-Rescaled Evidence), which rescales sampled answer frequency by question-level cardinality to approximate the total support on valid answers. Using verified counts, CARE raises AUROC from to on the broadest mixture, recovering approximately of the loss and demonstrating the value of estimating the number of valid answers for more reliable confidence.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, confidence estimation, uncertainty quantification, answer cardinality, selective prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sWPXOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies uncertainty estimation for LLMs when a question has multiple valid answers. The authors introduce MACE, a benchmark with 12,000 questions from six domains. Each question has 1, 2, 4, or 6 valid answers.

The authors find that when a question has more valid answers, LLMs often give lower confidence scores, even though their answer accuracy increases. They call this problem the concentration–correctness gap. They also show that many uncertainty estimation methods perform worse on questions with multiple valid answers. To address this problem, they propose CARE, a simple method that adjusts confidence scores based on the number of valid answers. Experiments show that CARE improves AUROC compared with the original consistency-based method.

Critical strengths
  • The motivation is clear
  • The authors introduce a new benchmark with different numbers of valid answers, which is useful for studying uncertainty estimation
  • The paper evaluates many existing methods and provides findings about them across different answer cardinalities
Critical weaknesses
  • The writing part should be improved. There are many compound sentences connected by punctuation, created terms, and some paragraphs directly go to very technical details without introducing the background or motivation
  • The definition and novelty of the concentration–correctness gap need further clarification. The gap is not unique to multi-answer questions. A model can be highly concentrated on an incorrect answer even when only one valid answer exists. This is supported by Table 1&2, but it is not clear how the proposed concentration–correctness gap is formally defined and how it is associated with the ranking capability of an uncertainty method.
  • The definitions of query-level and answer-level uncertainty are unclear and potentially misleading. The paper defines query-level uncertainty mainly through the model's output distribution. However, query-level uncertainty can also be estimated before answer generation [1]. The authors should clarify this difference.
  • The proposed CARE method lacks theoretical justification and has limited practical value. The method appears to be heuristic, without a clear explanation of why this form should estimate answer correctness. CARE assumes that the true number of valid answers (K) is known. In real-world applications, this information is usually unavailable.
  • It is not clear why report the cumulative answer-cardinality
  • P(True) is very stable and robust against larger K, which undermines the contribution of this work
  • The benchmark only keeps popular facts, which is not aligned with real-world questions.

[1] Query-level uncertainty in large language models. ICLR 2026

Questions for authors

See weaknesses

7 comments
AuthorsOpenReview ↗

Response to Reviewer sWPX

Response to Weakness 1: Writing and Presentation

Thank you for your suggestions on presentation. We will make the exposition clearer, shorten compound sentences, and streamline terminology.

Response to Weakness 2: Novelty and Definition of the Concentration–Correctness Gap

Thank you for raising this distinction. We would like to clarify that our finding concerns reduced concentration on a correct response because other valid answers share the support. In your single-answer example, probability concentrates on a wrong answer, leading to high confidence in an error. In our case, total valid support remains stable, but its division among correct alternatives lowers confidence in the returned correct answer.

For a returned correct answer, the gap can be formally defined as the difference between total probability mass on all valid answers and the mass on the returned answer, measured over the semantic answer distribution. It captures support allocated to other correct alternatives. Table 2 illustrates this: among correctly answered questions at each answer count, total valid support remains around 90–92%, while support for the returned answer falls from 90.2% to 70.0%.

This gap affects correctness ranking because many estimators derive confidence from concentration through answer likelihood or sampling agreement. Valid-answer dispersion can therefore lower a correct response’s score enough to rank it below an incorrect response. Ranking metrics measure this loss of alignment between confidence and correctness, complementing the distributional evidence for the gap.

We agree that this definition and connection were not clear enough and will clarify them in the revised manuscript.

Response to Weakness 3: Query-Level and Answer-Level Uncertainty

Thank you for pointing out this distinction and reference [1]. Our study follows the established post-generation evaluation setting. Within this setting, our taxonomy distinguishes query-level uncertainty over the sampled output distribution, represented by semantic entropy [2], from answer-level confidence in a particular response, represented by P(True) [3]. We acknowledge that “query-level” may be ambiguous given its use for pre-generation answerability in [1]. We will clarify this terminology and discuss this complementary direction in Related Work.

[1] Query-Level Uncertainty in Large Language Models. ICLR 2026.

[2] Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation[J]. arXiv preprint arXiv:2302.09664, 2023.

[3] Language models (mostly) know what they know[J]. arXiv preprint arXiv:2207.05221, 2022.

Response to Weakness 4: Motivation, Design, and Practical Value of CARE

Thank you for raising this point. We would like to clarify the motivation behind this transformation. As Table 2 shows, when multiple answers are valid, support for a correct answer can decrease because other valid answers share the model’s support. CARE uses answer count to compensate for this potential dilution without identifying which generated answers are correct.

Linear rescaling is intuitive, but valid support is not equally distributed. CARE instead uses a saturating transformation: for a fixed count, it behaves like count multiplication near zero, then smoothly saturates as frequency increases. With verified counts, it preserves single-answer scores and rankings within each fixed count.

Experiments with verified counts isolate the value of cardinality information. Since counts are usually unavailable in practice, we also tried predicting them from model-output features using a regressor (Appendix D.12). All results below are for LLaMA-3.1-70B-Instruct (Tables 3 and 34). Recovery measures the fraction of uncorrected Consis’s observed AUROC loss regained in the full mixture:

MethodSingle-answer AUROCFull-mixture AUROCFull-mixture gain over ConsisLoss recovered
Consis80.174.6——
Linear (verified counts)80.176.4+1.9≈34%
CARE (verified counts)80.178.9+4.3≈77%
Linear (estimated counts)73.573.6−1.0≈−18%
CARE (estimated counts)78.975.6+1.1≈20%

CARE outperforms linear rescaling in both settings: it recovers more of the loss with verified counts and remains beneficial with estimated counts, where linear rescaling degrades AUROC. Estimated-count gains remain limited, but the verified-count results demonstrate the value of cardinality information. We identify reliable count estimation directly from the question as future work in Section 6.

AuthorsOpenReview ↗

Response to Reviewer sWPX

Response to Weakness 5: Why Evaluate Cumulative Mixtures

Thank you for raising this question. AUROC measures how well confidence ranks correct responses above incorrect ones. In deployment, questions with different numbers of valid answers arrive together, and a system may apply one confidence threshold to decide which responses to accept or reject. Scores must therefore be comparable across answer counts. For example, a correct response to a multi-answer question may receive a lower score than an incorrect response to a single-answer question because its support is spread across valid alternatives. Separate-cardinality AUROC never tests this comparison: it only compares correct and incorrect responses within the same answer count. An estimator can therefore rank well within each count while ranking poorly when the groups are mixed.

Table 3 progressively includes multi-answer questions to assess this comparability. Consis falls from 80.1 AUROC on single-answer questions to 74.6 on the full mixture. When comparing correct responses to questions with six valid answers against incorrect responses to questions with one valid answer, AUROC is only 59.5 (Table 20), showing why cross-count comparisons matter.

We therefore prioritize pooled AUROC, with separate-cardinality results in Appendix Table 18 as complementary diagnostics. We will present this rationale more clearly in Section 5.

Response to Weakness 6: Stability and Discrimination of P(True)

Thank you for raising this point. We would like to clarify that PTrue-Consis-Cand’s relative stability is itself one of our findings (Section 5.3). Our mechanism analysis suggests that separate verification may reduce reliance on generation dominance, making it an important direction for further research. Stability matters, but so does absolute discrimination. Table 3 reports the following AUROC results for LLaMA-3.1-70B-Instruct:

EstimatorSingle-answer AUROCFull-mixture AUROC
PTrue-Consis68.463.3
PTrue-Prob72.566.2
PTrue-Prob-Cand75.170.6
PTrue-Consis-Cand62.562.4
Consis80.174.6
CARE (verified counts)80.178.9

The nearly stable variant achieves 62.4 pooled AUROC versus 74.6 for uncorrected Consis, while the other three P(True) variants decline. CARE addresses the instability of a stronger estimator: with verified counts, it raises Consis’s pooled AUROC to 78.9 and reduces its decline from single-answer performance. Improving separate verification and correcting strong estimators such as Consis are complementary directions toward confidence that is both stable and discriminative.

Response to Weakness 7: Purpose of Popularity Filtering

Thank you for raising this point. We would like to clarify that MACE provides controlled, closed-book confidence evaluation across answer counts. Studying valid-answer dispersion requires some model support for valid alternatives. Popularity filtering (Section 3.2; Appendix B.3) reduces the risk of unfamiliar entities dominating the evaluation, which would mainly test knowledge gaps or the need for retrieval. Limiting the candidate pool also controls expert verification costs.

Excluding obscure entities does not mean that the remaining questions are easy. Even after filtering, LLaMA-3.1-70B-Instruct achieves only 48.0% single-answer accuracy (Table 2), showing that the benchmark remains challenging. It reflects a practical use case: factual questions answered from the model’s existing knowledge without external retrieval. This design supports controlled diagnosis of confidence reliability in a relevant application setting.

Dear Reviewer sWPX,

Thank you again for your valuable comments on our paper. As the rebuttal period draws to a close, we would appreciate your feedback on whether our responses and additional experiments have adequately addressed your concerns.

If any questions remain, we would be happy to provide further clarification. If our responses have resolved your concerns, we would be grateful if you would reconsider your assessment and score in light of the additional evidence.

Thank you for your time and consideration. We look forward to hearing from you.

Best regards, The Authors

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

The paper shows that, as the number of correct answers grow, the confidence assigned to each correct answer becomes lower. This is done by using their newly introduced benchmark MACE. The paper then shows a confidence score adjustment that empirically helps with the AUROC score but does not provide any intuition on why the score is good or how it is derived.

Critical strengths

The implication of the pheneomenon seems believable: If there are K correct answers, then the probability of LLM answer for each correct tends to be lowerred. People normally use this as the score. The inference normally have a cutoff for that if score is lower than a certain threshold, it. will not show (to minimize hallucinated answer). Thus, if K is big, even a correct answer can be flagged as being incorrect. Therefore, we have to redefine score to be higher when K is big.

Critical weaknesses

The paper suggests that the score will change from c to 1-(1-c)^K without saying why. However, this is equivalent to the probability of K-independently drawn answers contain each correct answer at least once.

Still, there is no basis on why this will be a good score theoretically. I also feel that the improvement in the AUROC score seems not so big. However, my background is not much on empirical work, so I may misjudge the improvement.

Questions for authors

Why do we choose the new score to be in that way? Is there a way that we can model the distribution of answer so that such new scoring can be rationalized for increasing AUROC?

4 comments
AuthorsOpenReview ↗

Response to Reviewer fZCE

Response to Weakness 1 and Question 1: Why We Use This Score Transformation

Thank you for this observation. We would like to clarify that the independent-draw interpretation concerns seeing one particular answer at least once, not every correct answer. The connection to CARE is the function’s behavior: adding more independent opportunities increases this probability, but with diminishing gains. CARE borrows this saturating behavior, using answer count to control the correction strength while keeping the actual sampling budget fixed.

Using answer count is motivated by the support dilution in Table 2: support for a correct answer can decrease because other valid answers share the model’s support. CARE uses the count to compensate for this potential dilution without identifying which generated answers are correct.

Linear rescaling is a natural alternative, but valid support is unevenly distributed. Multiplying by answer count can therefore apply a large correction even when one answer dominates. CARE instead behaves like count multiplication near zero and smoothly saturates as frequency increases, for any fixed count.

To compare these corrections, Tables 3 and 34 report the following results for LLaMA-3.1-70B-Instruct using verified counts. Recovery is measured relative to uncorrected Consis’s single-answer-to-full-mixture AUROC loss:

MethodSingle-answer AUROCFull-mixture AUROCFull-mixture gain over ConsisLoss recovered
Consis80.174.6——
Linear (verified counts)80.176.4+1.9≈34%
CARE (verified counts)80.178.9+4.3≈77%

CARE achieves more than twice the gain of linear rescaling, supporting the saturating correction’s effectiveness. We will present this motivation and comparison more clearly in Section 6.

Response to Weakness 2: Recovery of the Observed Performance Loss

Thank you for raising this point. Our objective is to mitigate an estimator’s loss of discrimination when answer cardinalities are mixed, rather than to maximize absolute AUROC. We seek to keep the same estimator comparably reliable in single-answer and multi-answer settings. The +4.3 AUROC gain should therefore be assessed against the observed loss: it recovers approximately 77% of Consis’s degradation. Thus, a seemingly modest absolute gain represents substantial mitigation of the failure mode we study. We will make this objective and interpretation clearer in Section 6.

I still feel that the motivation behind the choice of new score seems lacking.

AuthorsOpenReview ↗

Response to Reviewer fZCE

Thank you for following up—we are glad to clarify further. Below we explain (1) how 1−(1−c)K1 - (1 - c)^K arises from a simplex-based stick-breaking model of the answer distribution, (2) why this mathematical structure improves pooled AUROC, and (3) offline validation against effective valid mass.

1. Probabilistic Motivation and Simplex-Based Stick-Breaking Model

Target quantity for cross-KK ranking. Let A∼p(⋅∣x)A \sim p(\cdot \mid x) denote a semantic answer drawn from the model's output distribution, and let V(x)V(x) be the set of KK mutually exclusive valid answers. What reflects whether the model knows a valid answer is the total valid probability mass m(x)=P(A∈V(x))∈[0,1]m(x) = P(A \in V(x)) \in [0, 1]. When K=1K = 1, a valid returned answer a^\hat{a} holds all valid mass, so sample agreement c=p^(a^∣x)c = \hat{p}(\hat{a} \mid x) directly estimates m(x)m(x). When K>1K > 1, however, m(x)m(x) remains high (≈0.91\approx 0.91 on correct responses, Table 2) but is partitioned across KK valid answers on the probability simplex (whose probabilities sum to 11), so cc observes only one slice—which can cause a correct multi-answer response to receive a lower score than a confident single-answer error. Without knowing which other sampled answers are valid at test time, we therefore seek to estimate total valid mass m(x)m(x) from (c,K)(c, K) alone.

Simplex-based stick-breaking model. Once the returned answer claims probability mass M(1)=cM(1) = c, only 1−c1 - c of residual probability budget remains on the simplex for all other answers combined. Naive linear scaling (KcK c) assumes every valid alternative takes an unconditional mass cc, which violates the simplex bound m(x)≤1m(x) \le 1 whenever c>1/Kc > 1/K and ignores empirical answer skewness (Table 2). Instead, we use a simple symmetric stick-breaking model on the simplex in which each additional valid answer captures the same conditional fraction cc of the remaining probability mass. Letting M(j)M(j) denote the cumulative mass across the first jj valid answers gives the recursion

M(1)=c,M(j)=M(j−1)+c(1−M(j−1)).M(1) = c, \qquad M(j) = M(j - 1) + c\big(1 - M(j - 1)\big).

Under this model, the KK valid answers receive geometric shares c,  c(1−c),  …,  c(1−c)K−1c,\; c(1 - c),\; \dots,\; c(1 - c)^{K-1} and invalid answers receive the residual (1−c)K(1 - c)^K, forming a proper simplex distribution that sums to 11 for all c∈[0,1]c \in [0, 1] and K≥1K \ge 1. Solving the recursion above gives CARE:

C(c,K)=M(K)=c+c(1−c)+⋯+c(1−c)K−1=1−(1−c)K.C(c, K) = M(K) = c + c(1 - c) + \dots + c(1 - c)^{K-1} = 1 - (1 - c)^K.

2. Mathematical meaning for pooled AUROC.

  1. Exact K=1K = 1 anchoring and within-KK invariance: C(c,1)=cC(c, 1) = c, and for any fixed KK, C(c,K)C(c, K) is strictly increasing on c∈[0,1)c \in [0, 1), so within-cardinality rankings and AUROCs are strictly preserved.
  2. Adaptive cross-KK calibration via remaining-budget damping: Expanding 1−(1−c)K=Kc−(K2)c2+O(c3)1 - (1 - c)^K = K c - \binom{K}{2} c^2 + O(c^3) shows two natural regimes: when cc is small (strong multi-answer dilution), 1−c≈11 - c \approx 1 and C(c,K)≈KcC(c, K) \approx K c recovers dispersed valid support; when cc is large (one dominant answer), the shrinking residual budget 1−c1 - c tapers the increment via −(K2)c2-\binom{K}{2} c^2, limiting the amplification of moderately concentrated multi-answer responses (unlike unbounded KcK c) while avoiding hard-clipping ties at 1.01.0 for all c≥1/Kc \ge 1/K (unlike min⁡(Kc,1)\min(K c, 1)).

3. Offline Validation Against Effective Valid Mass

Because C(c,K)C(c, K) is motivated as an estimator of valid probability mass from (c,K)(c, K) alone, we directly evaluate its alignment with the reference-audited effective valid mass q(x)q(x) (Table 2) on LLaMA-3.1-70B-Instruct (n=11,388n = 11{,}388):

Score transformationSpearman ρ\rho with effective valid mass q(x)q(x) [95% bootstrap CI]MAE vs. q(x)q(x)Pooled AUROC (<=6)
Raw concentration cc0.614 [0.602, 0.627]0.30874.6
Linear KcK c0.576 [0.563, 0.588]0.88176.4
Clipped linear min⁡(Kc,1)\min(K c, 1)0.565 [0.552, 0.578]0.25674.5
CARE 1−(1−c)K1 - (1 - c)^K0.661 [0.650, 0.672]0.24178.9

CARE achieves the highest rank correlation with effective valid mass q(x)q(x) (paired bootstrap Δρ=+0.047  [+0.039,+0.055]\Delta\rho = +0.047\;[+0.039, +0.055] over raw cc, and ≥+0.086\ge +0.086 over linear variants), the lowest MAE, and the highest pooled AUROC. Hard clipping erases the gain of linear scaling because every multi-answer response with c≥1/Kc \ge 1/K collapses onto 1.01.0, creating extensive ties, whereas CARE saturates smoothly and raises pooled AUROC from 74.6 to 78.9 (versus 76.4 for KcK c).

Thank you again for your insightful comments. We will incorporate this simplex stick-breaking model and offline validation into Section 6 of the revised manuscript, and we would be happy to clarify further if any questions remain.

AuthorsOpenReview ↗

Response to Reviewer fZCE

Dear Reviewer fZCE,

We thank you again for taking the time to review our response, and for stating your points so concretely. Reviewing is demanding, and our additional analyses have added to your verification burden—we are sorry for that and sincerely appreciate your patience. With the discussion period now drawing to a close, we remain available throughout it and are happy to answer any further questions or provide any clarification. We hope our additions address your concerns, and we would be grateful to hear any further comments before the discussion ends.

Reviewer_9vksOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how multiple correct answer affect existing confidence estimation methods. The paper first introduces MACE, a benchmark dataset with varying number of valid answers. Also, it introduces CARE, a cardinality-aware confidence estimator.

Critical strengths

The paper introduces a benchmark supporting confidence estimation with multiple correct answers.

Critical weaknesses
  1. A straightforward solution when multiple correct answer is just aggregating the probability assigned to all valid answers. While the paper has included it as valid mass q(x)q(x) in Table 2, it is unclear why its not evaluated alongside CARE as a baseline.

  2. The inversion conjecture states that as k grows, "confidence will fall even as the model answers better", and uses "accuracy rises from 48.0 to 61.7" (Ln316-317) as the supporting evidence. This, however, might be because that larger K relaxes the question constraints, which might change question difficulty. The increase in accuracy might simply because more answers are considered correct, rather than the model demonstrate stronger knowledge ("model answers better").

Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9vks

Response to Weakness 1: Why Valid Mass Is Used as a Diagnostic

Thank you for raising this point. We would like to clarify the reference information required at test time:

Method or quantityAdditional reference information at test time
ConsisNone
CARE with verified countsNumber of valid answers
Valid massReference answer set identifying which answers are valid

Computing valid mass requires identifying valid semantic clusters through reference-based correctness judgments and summing their normalized probabilities (Appendix D.2). It is therefore a reference-dependent diagnostic (Table 2). Given these references, a system could directly check correctness instead of estimating confidence.

CARE needs only the number of valid answers as additional reference information, without identifying which generated answers are correct. This smaller information requirement still yields a substantial improvement in confidence ranking. The following results use verified counts on LLaMA-3.1-70B-Instruct (Tables 3 and 34); recovery is measured relative to uncorrected Consis’s single-answer-to-full-mixture AUROC loss:

MethodSingle-answer AUROCFull-mixture AUROCFull-mixture gain over ConsisLoss recovered
Consis80.174.6——
Linear (verified counts)80.176.4+1.9≈34%
CARE (verified counts)80.178.9+4.3≈77%

CARE outperforms linear rescaling and recovers approximately 77% of the observed loss without access to reference answer identities. This demonstrates the value of answer count for confidence estimation, while valid mass remains a reference-dependent diagnostic. We will make this distinction clearer in the main text.

Response to Weakness 2: Relaxed Constraints and Increasing Accuracy

Thank you for highlighting this distinction. We agree that allowing more valid answers can increase accuracy without increasing model knowledge, and will revise “the model answers better” accordingly.

However, this does not change our central finding: AUROC still declines when accuracy is held constant across answer counts. In Appendix D.5, we randomly subsample correct predictions so that each answer-count group has the same proportions of correct and incorrect responses, while leaving confidence scores unchanged. Table 21 reports the results for LLaMA-3.1-70B-Instruct, averaged over five subsamples:

EstimatorSingle-answer AUROCFull-mixture AUROC
Semantic Entropy78.975.4
Consis80.177.6
Perplexity80.175.9

Matching accuracy reduces the decline but does not eliminate it. The distributional analysis in Table 2 helps explain this remaining loss: among correctly answered questions at each count, total valid support remains around 90–92%, while support for the returned answer falls from 90.2% to 70.0%.

Together, these results support our central finding: valid-answer dispersion can lower confidence in correct responses, and correctness ranking can deteriorate even without rising accuracy. We will present this control more clearly in Section 5 and distinguish task accuracy from model knowledge.

Dear Reviewer 9vks,

Thank you again for your valuable comments on our paper. As the rebuttal period draws to a close, we would appreciate your feedback on whether our responses and additional experiments have adequately addressed your concerns.

If any questions remain, we would be happy to provide further clarification. If our responses have resolved your concerns, we would be grateful if you would reconsider your assessment and score in light of the additional evidence.

Thank you for your time and consideration. We look forward to hearing from you.

Best regards, The Authors

Thank you for the response. My first concern has been largely addressed. For the second one, I appreciate the additional experiment. However, I am not sure if matching accuracy represents question difficulty. Dataset with identical accuracy may still differ in the way they are constructed, so the observed AUROC degradation cannot be attributed exclusively to K. Given that the inversion conjecture is central to the paper, this concern requires a more careful interpretation of the results. I will therefore maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.