acceptodds
Under review as a conference paper at ICLR 2027

Reasoning-Aware Multicalibrated Reliability Predictors for Language Models

Abstract

As people increasingly rely on large language models (LLMs) for information, model providers need to communicate *meaningful* estimates of whether individual responses are *correct* or, more generally, *reliable*. In the context of hallucinations, existing work uses *multicalibration* to make such reliability predictions meaningful across application-relevant groups, such as prompt topics. Topic groups alone, however, can conceal systematic errors: within the same topic and confidence level, responses with different chain-of-thought (CoT) properties may exhibit opposing errors that cancel on average. In this work, we introduce reasoning-aware multicalibration, a framework that supplements application-relevant topic groups with groups derived from observable properties of CoT traces. We highlight that, theoretically, groupings based on CoT are *"internally important"*, i.e., groupings for whom the multicalibration guarantee is not societally crucial as it is for high-risk topics, but whose inclusion in the multicalibration could easily and significantly improve predictive accuracy of the final reliability predictor. Empirically, we apply our framework to hallucination prediction and unsafe-response detection across a variety of benchmarks, and find that adding CoT groups to topic-based calibration reduces Brier score in both tasks. Interestingly, we find that multicalibrating with respect to CoT groups alone (without any topic information) already improves calibration across topics, suggesting that the predictor's errors depend on how the model reasoned, rather than what it was asked about in the prompt.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multicalibration, uncertainty quantification, hallucination, chain of thought, AI safety, calibration

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

This paper proposes reasoning-aware multicalibration, which incorporates chain-of-thought (CoT) properties into existing topic-based calibration for LLM reliability prediction. The authors provide a theoretical characterization of useful additional calibration groups and evaluate the framework on hallucination prediction and safety detection. Experiments show consistent Brier score improvements.

Critical strengths
  1. Well-motivated framework. The paper identifies systematic calibration errors overlooked by topic-only multicalibration and proposes a practical approach using observable CoT properties.
  2. Solid empirical evaluation. The experiments demonstrate consistent improvements across four QA benchmarks, with additional analyses of different annotators, calibration algorithms, cross-benchmark transfer, and safety prediction.
Critical weaknesses
  1. Limited methodological novelty. The framework primarily incorporates additional CoT-derived groups into an existing multicalibration algorithm. The theoretical result provides a necessary condition for improvement but does not establish when CoT groups will guarantee better performance.
  2. Insufficient comparison with alternative CoT-aware predictors. The paper does not directly compare against strong supervised calibration or reliability predictors that incorporate CoT features. Therefore, it remains unclear whether multicalibration provides a meaningful advantage over simpler approaches using the same information.
Questions for authors
  1. How does the proposed method compare with a supervised reliability predictor trained on the same initial scores, topic information, and CoT diagnostic features?
  2. Can the authors clarify the practical significance of theorem beyond characterizing when additional groups may be useful?
  3. How well does the framework generalize to reasoning traces generated by other LLM families, rather than only Qwen3-8B?
7 comments

We thank the reviewer for their notes of our well-motivated problem and strong empirical results. In this and the following comments, we respond to the weaknesses and questions listed in the review. For each weakness and question, we include a "too long; didn't read" (TLDR) response as well as an extended response. All responses are entirely human written. For parts of W1 we refer the reviewer to the comment we made for all reviewers. We are more than happy to engage further during this period and will try our best to respond to additional comments/questions.

AuthorsOpenReview ↗

Response to Weakness 1

W1: Limited methodological novelty and guaranteed-good method for group selection.

We kindly ask that the reviewer reads our official comment to all reviewers for our response about limited methodological novelty. As for the comment that "The theoretical result provides a necessary condition for improvement but does not establish when CoT groups will guarantee better performance," we add the following:

TLDR

A set of conditions that is guaranteed to select groups that decrease error is a hard open theoretical problem better suited for a theory paper. Instead, we are able to show a weaker statement: CoT contains information that would likely improve the multicalibration process. This alone is an interesting and novel result.

Extended Response

The reviewer highlights a very good open problem: Given a collection of potential additional groups, which ones, when included in multicalibration, are guaranteed to improve error? This is a restatement of our Definition 3.1. We do not answer this in our work as it is a very non-trivial theoretical problem whose answer is better suited for a primarily theoretical paper. We hope to present this problem to the multicalibration theory community by being the first work to formalize the group augmentation setting in Sec 3. To ensure it is clear that this problem is interesting future work, in the camera-ready, we will add it to a future work subsection in Sec 6: Discussion and Limitations. As for the case of guaranteed-good CoT groups specifically, note that here we actually do not even have pre-made groups, we must also feature engineer them. For this reason, we expect that the (weaker) conclusion that CoT very likely contains information to improve error is all that is reasonably possible. To see this, consider the complexity and variation of CoT reasoning, which is natural language data. The number of mappings of this natural language data to binary groups is unreasonably large. Thus, even if we had an answer to the general group selection question, in the case of CoT, the number of groups we must select from is extremely large. Assuming the complexity of the optimal group selection depends on the number of groups, it is likely that it would not be efficient to apply to CoT without first feature-engineering a smaller number of CoT-based groups to select from.

AuthorsOpenReview ↗

Response to Weakness 2

W2: Insufficient comparison with alternative CoT-aware predictors.

TLDR

The premise of work is that we must have an LLM reliability predictor that is guaranteed to be meaningful/trustworthy on all topic groups (while also getting low error). Multicalibration algorithms are (by definition) algorithms that achieve the guarantee. We do not want to acheive the lowest error possible if it is with a generic supervised algorithm that does not have the theoretical guarantee. Thus, comparison to generic algorithms is out of scope of the setting.

Extended Response

We divide this response into a three aspects of the weakness: (A) Comparing to other generic supervised learning algorithms, (B) Comparing to other multicalibration algorithms, ( C) Comparing to other simpler approaches.

A: Why don't we test other generic supervised learning algorithms? Multicalibration with respect to some collection of groups is a theoretical guarantee that a predictor be meaningful on each of those groups. More specifically, of all the times a predictor gives a probability of 99% for group A, 99% of those times, the binary outcome actually happens (Section 2, lines 140-153). The foundational premise of our work (lines 044-047) and also Detommaso et al, 2024 (see their Section 1) is that it is necessary this guarantee holds for topic groups when building a predictor of LLM reliablity. To see why this guarantee is good, consider, for example, the high-stakes topic of chemistry questions where recommendations from an LLM could lead to harmful experimentation by a user. It is important that in responses on this topic specifically, if the predictor says the output has 99% chance of safety, that this 99% a meaningful statistic (i.e., that genuinely for 99% of such responses, the output is safe). Having perfect accuracy (i.e. the Bayes optimal predictor) would obviously achieve this, but a generic algorithm that maximizes for accuracy or minimizes error is not guaranteed to achieve this (see lines 169-181), which is generally not possible. Thus, we focus on multicalibration algorithms rather than generic supervised learning algorithms. Of course, having lower error is still nice, so while retaining the meaningfulness guarantee on topic, we consider whether CoT (which has been shown to have signal for reliability, see Sec 1.1) can be usefully incorporated into the multicalibration process. Other supervised training methods definitely could achieve lower error (and even might achieve multicalibration, as we highlight in Q1!), but they do not explicitly optimize for it and thus are not guaranteed to achieve multicalibration in the way multicalibration algorithms are.

B: Why don't we test other multicalibration algorithms? We note that our framework is purposefully general and it can work with any multicalibration algorithm (see "the algorithm can be any out-of-the-box batch multicalibration algorithm" line 347). In the main body we conduct our case study using only IGHB, but in the appendix we also run the IGLB multicalibration algorithm (see E.1:Calibration algorithm and Table 7). With IGLB, we see similar trends in multicalibration w.r.t CoT and topic groups yielding lower error than topic-only.

C: The argument for the simplicity of multicalibration. As described in W2:B, our framework works for any generic multicalibration algorithm, but as the reviewer asks about "simpler approaches using the same information", we want to highlight that the patching-based multicalibration algorithms (IGHB and IGLB) we focus on are arguably the simpler than even many others that maximize accuracy and do not guarantee multicalibration. In particular, consider Algorithm 1. As written, it may appear complicated, but its only function is to iteratively modify the predictor's initial value for it's most uncalibrated sections such that the value over those spaces of X\mathcal{X} is more calibrated. What this means is that if, for example, the initial function is already nearly multi-calibrated, the multicalibrated version will be nearly the same, but with the values for certain xx modified. As such, patching-based multicalibration algorithms carry a certain explainability and simplicity that much harder to achieve from a generic algorithm that takes in the initial function, topics, and CoT groupings and maximizes for accuracy. Without the structure of Algorithm 1, we cannot so simply describe the transformation of f0→fTf_0 \rightarrow f_T and if f0f_0 already carries some initial meaning, understanding how it has been changed is important.

Reviewer_NT9BOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a reasoning-aware multicalibration framework for predicting the reliability of LLM responses. The work is motivated by the observation that topic-based multicalibration can overlook systematic prediction errors associated with different chain-of-thought (CoT) characteristics. To address this limitation, the authors extend topic-based multicalibration by incorporating additional groups derived from observable CoT characteristics, such as uncertainty, self-correction, and consideration of alternative answers. These groups are termed internally important because they can reveal residual calibration errors and potentially improve predictive accuracy.

Critical strengths
  1. The research problem is well-motivated. Figure 3 demonstrates that topic-level multicalibration can conceal substantial differences in response correctness associated with different reasoning characteristics.
  2. The experimental results demonstrate that the proposed joint multicalibration framework consistently outperforms topic-only multicalibration in terms of Brier score across four QA benchmarks.
Critical weaknesses
  1. While the methodology is well-motivated by empirical evidence, the proposed methodology somewhat fall short in novelty and technical depth. The key idea of the proposed method is to construct CoT-derived groups and incorporate them into an established multicalibration algorithm (Detommaso et al, ICML 2024). The main contribution of this paper is the selection of additional groups based on observable reasoning characteristics.
  2. In addition, the theoretical formulation is also relatively disconnected to (or insufficient to support) the claimed contribution. Proposition 3.1 establishes that internal importance is a necessary condition for additional groups to improve the Brier score under greedy multicalibration. However, this result applies to arbitrary additional groups and is not specific to CoT-derived groups. It does not establish guidance about how to construct groups that satisfy the conditions. I think it is critical for the authors to connect the theoretical analysis to the empirical evidence that they have.
  3. Despite the encouraging empirical evidence, there lacks a principled way to select those groups. The current group selection relies on (1) semantic embedding clustering; and (2) 10 manually designed diagnostic questions to construct CoT groups. The paper should present a more systematic investigation about why these 10 manual diagnostics are necessary and sufficient conditions to improve calibration.
Questions for authors
  1. Can you please analyze the impact of the group selection strategy, e.g., removing, adding, or combining groups.
7 comments

We thank the reviewer for their notes of our well-motivated problem and strong empirical results. In this and the following comments, we respond to the weaknesses and questions listed in the review. For each weakness and question, we include a "too long; didn't read" (TLDR) response as well as an extended response. All responses are entirely human written. For W1 we refer the reviewer to the comment we made for all reviewers. We are more than happy to engage further during this period and will try our best to respond to additional comments/questions.

AuthorsOpenReview ↗

Response to Weakness 1

W1: Adding groups to existing multicalibration algorithms lacks novelty.

Thank you for noting this concern, in very short, we have chosen such a simple method very intentionally as it represents an understudied theory setting. We kindly ask that the reviewer reads our official comment to all reviewers. Additionally, we want to point out a minor discrepancy to the reviewer: they are correct to point out that we use an "established multicalibration algorithm" in the form of IGHB, but we note that IGHB is a very generic patching-based multicalibration approach. It was not a novel algorithm as used in Detommaso et al. 2024 because their IGHB is equivalent to Roth 2022 Alg 15, which is itself a minor modification of the patching algorithm in Hebert-Johnson 2018. We reference IGHB as being from Roth 2022 in Sec 3.

AuthorsOpenReview ↗

Response to Weakness 2

W2: Theoretical formulation connections

TLDR

1. Our theory applies to arbitrary groups purposefully because the question of useful group augmentation for multicalibration in the general setting is an open theory problem and our results contribute (see also response to W1)

2. We make the general theory/CoT bridge in Section 3.2.1. We elaborate on this subsection below and are happy to revise anywhere our intentions were not successfully fullfilled.

3. We do not give general theoretical guidance on how to extract good groups that are guaranteed to decrease error, but we do use Prop 3.1 to explain why CoT specifically contains the information necessary to do so (Sec 3.2.1). Unfortunately, a method that is guaranteed to extract groups that decrease error is a hard open theoretical problem. We will specify this in Sec 6.

Extended Response

To our understanding, this weakness encapsulates three primary critiques: (A) the theory is for generalized groupings and not specifically CoT, (B) the theory lacks connection with the rest of the work, ( C ) we do not provide a method by which the practioner can find good internally important groups. We address each of them separately and are happy to address any additional concerns the reviewer feels we missed.

A: Theory is for arbitrary groups. The reviewer accurately points out that that our theoretical results (Prop 3.1) and Section 3 setting more generally are for multicalibration with arbitrary additional groups (not just topic and CoT) and that is correct, but we have purposefully presented the generalized theoretical setting. Basically, it is a feature, not a bug. As described in our response to W1, the more general theoretical set-up is a novel problem for the multicalibration theory community who have not previously studied group augmentation while at the same time, Section 3 broadly highlights that group addition (in any case of multicalibration) is not an obviously always good thing to do to a reader unfamiliar with the theoretical aspects of patching-based multicalibration.

B: Lack of Theory/Framework connection. As for the theory's direct connection to the LLM reliability predictor framework, the main idea is that our theory sets up the reasoning for why additional groups might be useful and then the remainder of the paper highlights that CoT-based groups satisfy this. Our intention is that this bridge is made in Section 3.2.1: Are CoT-based Groups Internally Important?. We apologize this bridge seems to not have been made successfully and would be keen to hear how we may improve it. As is, Section 3.2.1 says the following:

  1. In lines 279-282, we highlight why Prop 3.1 is important in determining whether groups engineered from CoT are likely to be useful in the multicalibration process. Specifically, we say that CoT is already known to provide signal, thus satisfying the first aspect of internal importance (which is necessary by Prop 3.1), so as long as this signal is unique to that which is from topic, we will have both aspects.
  2. In lines 282-283, we explain that we've created our Section 4 framework of adding CoT groups to the multicalibration process to harness this likely utility of CoT.
  3. On line 283-284, we restate that our theory indicates that "we reasonably suspect that CoT is useful", but of course this is not a guarantee, so in Sec 5 then even confirm that our theoretical conditions hold and that CoT-based groups are indeed useful to decrease error

We are happy to improve anything in this section (or in others) the reviewer feels hurt the intended connection between theory/framework and are very open to further discussion.

C: No guidance on how to construct groups that satisfy the conditions. It is true that we do not give guidance on how to find additional groups that satisfy Prop 3.1. However, this is because we set up Sec 3 to explain why we theorize CoT will already very likely yield for good additional groups (see discussion of the theory/framework connection). The reviewer highlights a very good open problem: Given a collection of potential additional groups in the general setting, what is a method to select which ones are worthwhile to include? It is a very non-trivial theoretical problem and is outside the scope of this work where we want to primarily study CoT as additional groups, but we hope to open this problem to the multicalibration theory community open by being the first work to formalize the group augmentation setting in Sec 3. To ensure it is clear that this is for future work, in the camera-ready, we will add it to a future work subsection in Sec 6: Discussion and Limitations.

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

This paper studies how to build more calibrated binary estimators of a LLM response on some given property (e.g. hallucination). It extends an existing approach called multicalibration, which calibrates the estimator so that the error bound holds within each of a collection of (possibly overlapping) groups, such as prompt topics. It proposes to add an additional group on top of the primary grouping, prompt topics, to also consider the characteristics of the CoT chains. Because the fact that the groups can be overlapping, and that the algorithm uses an iterative greedy approach, adding more groups is not always leading to lower calibration error. This paper hence provides a theoretical argument that CoT groups are plausible candidates for improving accuracy: They show empirically that CoT groups carry information about the estimator’s errors that topic groups don't, which according to their theoretical analysis leads to lower Brier score.

Experiments show that CoT grouping is more effective than topic grouping when used alone, and that adding CoT groups to topic-based multi calibration reduces Brier error by 4.5 - 15% compared to topic-only calibration.

Critical strengths
  1. The proposed approach is sound, and the findings that CoT grouping is more effective than topic grouping is interesting and useful to the community. The extension to multicalibration with additional grouping is well-justified, and backed by empirical results.
  2. The paper is relatively clearly (and human) written, despite the substantial terminologies and theoretical analyses.
Critical weaknesses
  1. The paper mostly relies on multicalibration, an existing approach, and the extension is mostly about a justification on which additional groups can be added to an existing grouping. Technically, the approach is essentially the same multicalibration with more groups.
  2. CoT grouping alone achieves most of the error reduction. In fact, adding topic-grouping on top of CoT grouping leads to marginal improvements (within error bounds) on all but one datasets (PopQA). It seems like the same effect can be achieved by simply identifying CoT as the groups instead of topics during multicalibration.
  3. The diagnostic questions are the key to group CoT reasoning chains, yet they are not sufficiently justified, tested, or compared with alternatives.
Questions for authors

How are the diagnostic questions selected? How much does the performance vary if using a different set of questions? Previous work found that CoT patterns vary greatly depending on the question (e.g. math vs. fact-seeking questions; see Section 5.2 of https://arxiv.org/pdf/2508.05618). How well does this fixed set of diagnostic questions generalize to different kinds of prompts?

5 comments

We are grateful for the reviewer's kind comments on the good justification of our approach and are especially glad to see their acknowledgement of the human effort that went into our submission. In this and following comments, we respond to the weaknesses and answer the listed question(s). For each weakness and question, we include a "too long; didn't read" (TLDR) response as well as an extended response. All responses are entirely human written. For W1 we refer the reviewer to the comment we made for all reviewers. We are more than happy to continue discussion during this period and will try our best to respond in a timely fashion to any follow-up questions/comments.

AuthorsOpenReview ↗

Response to Weakness 1

W1: our approach is simply multicalibrating on more groups.

Thank you for noting this concern, in very short, we have chosen such a simple method very intentionally as it represents an understudied theory setting. We kindly ask that the reviewer reads our official comment to all reviewers.

AuthorsOpenReview ↗

Response to Weakness 2

W2: It seems that in many cases CoT should be used as the only groupset.

TLDR

Yes, in some cases multicalibration with only CoT (and not both CoT + topic) yields very low error, but a multicalibration process that excludes topic does not satisfy the requirement of our setting, which is that LLM reliability predictors need a guarantee of meaninfulness on topic. Even if low error is achieved, multicalibration with only CoT cannot theoretically guarantee this in all settings.

Extended Response

The reviewer is correct to point out that in many cases, including both CoT and topic groups is essentially unnecessary for error reduction. However, this does not imply that it makes sense to use only CoT-based groupings because the primary utility of including groups in the multicalibration process is to achieve the prediction meaningfulness guarantee on those groups. In Sec 2, we highlight that multicalibration w.r.t. a particular group yields a function whose predictions within that group are reliable/trustworthy. This is a seperate desirable property from decreased general error (e.g., "Multicalibration is not equivalent to accuracy maximization", line 165-166). The foundation of our claim is that the reliability property on prompts/responses of different topics is something inherently desirable in the real-world (lines 052-053 and 145-149). This is why we call topic groups "externally important". To see why this might be the case, consider, for example, the high-stakes topic of chemistry questions where recommendations from an LLM could lead to harmful experimentation by a user. It is important that in responses on this topic, if the predictor says the output has 99% chance of safety, that is a meaningful statistic (i.e., that genuinely for 99% of such responses, the output is safe). Such a guarantee is similar to accuracy in various ways, for example, the Bayes optimal predictor (which achieves minimal Brier error) is also fully multicalibrated (lines 166-177). However, the definitions of the multicalibration guarantee (Def 2.1) and Brier error (Def 2.2) are not equivalent in general. For this reason, we purposefully frame our CoT groups as additional and not replacement: we do not care about achieving the meaningfulness guarantee on them because they do not represent groupings of the data that is inherently important to the downstream user, but we can get bonus accuracy from including them in the process. In Sec 5.3 we do see that multicalibration w.r.t CoT only does often recover multicalibration w.r.t topics, but this is not the same as having a theoretical guarantee that the predictor is multicalibrated w.r.t topics generally. Therefore, we feel it would be incorrect to make any claims that CoT-only is truly sufficient for the task. We are grateful the reviewer has pointed out that this is under-discusssed as we do not intend to sweep under the rug the fact that CoT-alone often does very well in error reduction. In the camera-ready, we will add a remark on this subject to ensure the reader has full clarity.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Addressing concerns about limited methodological novelty

Across all reviews, we have seen a consistent concern of limited methodological novelty in our work. In particular, the concern is that the process of simply adding groups to multicalibration is seemingly an overly simplistic method. To engage all reviewers equally, we would like to leave our response to this concern as a general comment. We present the response with a "too long; didn't read" version and an extended version. We are happy to discuss this concern further with all reviewers.

TDLR

Adding groups to a multicalibration process is actually a theoretically rich problem. Adding groupings can make the error of your resulting multicalibrated predictor worse, which is very strange and unintuitive! This issue is well-known, but good realistic group augmentation has never been properly studied and formalized. To this end, our multicalibration approach is just multicalibration with more groups, but that is both very intentional and a novel problem for the multicalibration theory community. Theory work in this setting would be useful to practioners wanting to take advantage of additional binary group data such that the important multicalibration guarantees are still retained while overall error is lowered.

Extended Discussion

Our LLM reliability predictor method is indeed simple: add CoT-based groups to multicalibration w.r.t topic. There is nuance in the fact that engineering CoT-based groups is itself non-trivial and we intend that our suggestion of diagnostic questions is useful in this respect, but the fundamental concept is not novel from an algorithmic perspective. We hope this is made sufficiently clear in our description of (1) the Reasoning-aware Multicalibration contribution of lines 091-099 because we purposefully avoid language that would indicate when using "a multicalibration algorithm [that] fits score corrections across both group families" it is a novel algorithm and (2) the multicalibration step of the framework in Sec 4.3 line 347 where we explicitly state "the algorithm can be any out-of-the-box batch multicalibration algorithm". We apologize if any phrasings imply a novel multicalibration algorithm as this is definitely not our intention and are happy to modify any writing the reviewers consider unreasonable.

The reason we focus on such a seemingly non-novel method is because, surprisingly, the field of multicalibration theory actually does not have (to the best of our knowledge) any established understanding of how to tackle this problem. In lines 177-180 and other places throughout Section 3, we highlight that even though intuition may suggest that providing more information about the space (via additional groups) cannot hurt, in general, adding groups to the multicalibration process actually does not guarantee a function that is even weakly better than one without the added groupings. In fact, adding groups can make your resulting multicalibrated predictor worse in terms of accuracy (see example C.1, which is referenced on line 277). This is very strange! This simple fact of multicalibration theory is well-known, but realistic group augmentation (i.e., the setting we present in Section 3 where a practioner must decide which additional groups to include s.t. he creates a multicalibrated predictor with decreased error) has never been properly studied and formalized (see sec 1.1 for more details on existing work) even though it is a very natural setting if a practioner has some additional data on top of that which must obtain the multicalibration guarantee. To this end, our multicalibration approach is, methodologically, just multicalibration with more groups, but that is very intentional. We have written the work with the dual intent of (1) presenting this interesting theoretical setting (and some results in the form of Prop 3.1) and (2) showing that CoT in the case of LLM reliability prediction in particular contains information that could likely improve multicalibration w.r.t. topic.

We are grateful to the reviewers for bringing this concern to light as we would like this dual purpose to be clear to all readers and explains a bit better why the actual method may be simple. To improve this, in the camera-ready, we will add a future works subsection to Section 6 that explicitly lists interesting open questions in both directions and highlight at the beginning of Section 3 that this is a novel theoretical setting, making clearer that selecting groups that are guaranteed to satisfy Defintion 3.1 is important to a practioner yet unknown.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.