acceptodds
Under review as a conference paper at ICLR 2027

Sharp Recovery Thresholds for Pairwise Self-Rewarded Softmax Updates

Abstract

An evaluator can rank the optimal action above every alternative at every interior policy, yet a policy trained from sampled pairwise comparisons can still converge to a suboptimal action. We study this phenomenon for a finite softmax policy that samples two actions independently and updates its logits using their real-valued score difference. The evaluator may depend on the current policy, whereas the true rewards remain fixed. We control evaluator error by a uniform bound on the policy-weighted standard deviation of bias and an -Lipschitz bound on pairwise bias contrasts. We assume that the step sizes have a divergent sum and a finite sum of squares, and that the comparison noise has bounded conditional variance. Under these conditions, we derive the sharp worst-case recovery threshold , where is the minimum true reward gap. Below this threshold, the sampled update converges almost surely to the optimal action from any full-support initialization. At the boundary with , recovery still holds for two or three actions, but the sampled update can fail with four or more actions even when the evaluator ranks the optimal action first at every interior policy. On such failure trajectories, the optimal action is compared infinitely often, yet the cumulative conditional drift toward it remains finite. A full-information update still recovers at this boundary for every finite action set, which shows that the sampled update can fail even when the evaluator retains enough information to identify the optimum. Finally, balanced queries with an importance-corrected update recover the optimal action for without a sensitivity bound, whereas evaluator feedback alone cannot guarantee identification when .

Primary area:
learning theory
Keywords:
self-rewarding, policy-dependent evaluation, pairwise comparisons, preference learning, softmax policy optimization, stochastic approximation, convergence theory, performative learning

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_zG9eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies a tabular softmax model (the tabular size is KK) learning from pair data with a biased reward annotator. They established a so-called "recovery threshold" result, showing that if the bias scale is somehow lower than the threshold, then convergence is guarenteed; and if larger than the threshold, they can construct an evaluator wrongly prefer a sub-optimal action to the optimal action, not converging to the optimal action in some settings; and if equal to the threshold, it still converges if K=2,3K=2,3, and they constructed counterexamples for K≥4K\ge 4. They also conducted some small synthetic experiments with tiny KK.

Critical strengths
  • I think this paper's theoretical results are insteresting to me, such as the convergence analysis technique and the concept of recovery threshold. The proof technique using martingale might be of independent interest. And the studied setting is clean and related to some topics in preference learning and machine learning.
Critical weaknesses
  • The paper-writing is terrible. I find it very hard to follow the paper. There are "too many theories" and the organization is not arranged very well. Sections 3–4 explain why the transition occurs at four actions, however, this should not be viewed as a major contribution which can take so many pages. The transition at four actions is only about the case when ϵ\epsilon exactly equals to the threshold, which is a corner case, and the results can be viewed as: only K=2,3K=2,3 cases are proved where K=2K=2 is already a trivial case, and K≥4K\ge 4 cannot be proved in this way (which is unnecessary to show in such a detailed way because we already have counterexamples in Theorem 2). The idea is not very new, as there are some analogous analysis in [1][2]. I feel that the paper develops an excessive amount of technical machinery without establishing sufficiently meaningful results.
  • The results are too specific and limited. One may expect to see the application of this idea in preference learning. The current version is more like solving a mathematic problem in which the general audience of ICLR might not be very interested. For example, one can try to relate the method with some standard preference learning practice, such as [3], which studies a similar pairwise RL training objective, and [4], which studies the tabular softmax model in DPO. It's also unclear whether the proposed Balanced-query recovery is a major contribution as it is only covered in theorem 5 without further detailed discussions, such as how to deploy it in practice.

[1] Escaping the Gravitational Pull of Softmax. NeurIPS 2020.

[2] Delightful Gradients Accelerate Corner Escape. arXiv.

[3] REBEL: Reinforcement Learning via Regressing Relative Rewards. NeurIPS 2024.

[4] The Crucial Role of Samplers in Online Direct Preference Optimization. ICLR 2025.

Questions for authors

Please see my comments in "Weakness".

3 comments
AuthorsOpenReview ↗

Presentation and Theoretical Significance

Thank you for the detailed comments. We address your concerns below.

“I find it very hard to follow the paper. There are "too many theories" and the organization is not arranged very well.”

The original presentation introduced several technical ideas before explaining how they fit together. We have revised the exposition to make the main argument clearer. The Introduction now explains bias dispersion and pairwise bias contrasts when they first appear, and Section 2.2 motivates the recovery threshold through endpoint cancellation. We also revisited the original four-action example in Section 3 to explain the role of the auxiliary actions. Their comparisons continue to favor the designated suboptimal action, while the auxiliaries themselves become rare. Through the conserved-product structure, the optimal action's mass also decreases, and its shrinking score margin can make the total corrective drift finite despite infinitely many comparisons.

Section 4 has also been reorganized around the proof itself. We first use the evaluator assumptions to bound the optimal action's aggregate score advantage. After removing the convergent martingale term, we introduce reciprocal coordinates whose conserved product controls how rare the optimal action can become. These bounds lead to the integrability test, and we now explain where its exponent K−2K-2 comes from. We also moved the bias-geometry figure and full sample-path rate theorem to Appendices A and E, so that the main text can focus on the central argument.

“Sections 3–4 explain why the transition occurs at four actions, however, this should not be viewed as a major contribution which can take so many pages.”

The original presentation focused on the exact threshold, which may have made the distinction seem specific to that case. We therefore investigated whether it persists strictly below the threshold. The new Theorem 2 shows that the same distinction governs near-critical recovery budgets.

Let κ=Δ−ΨL−1(ε)>0\kappa=\Delta-\Psi_L^{-1}(\varepsilon)>0. A cumulative step-size budget of O(κ−(K−1))O(\kappa^{-(K-1)}) suffices for high-probability first hitting. For every fixed K≥4K\ge4, we construct a family of correctly ranked, strictly subcritical evaluators, one for each κ\kappa, with a matching Ω(κ−(K−1))\Omega(\kappa^{-(K-1)}) lower bound from an open set of initializations independent of κ\kappa. For K≤3K\le3 and fixed L>0L>0, the required budget remains uniformly bounded up to the threshold (Section 2.3; Appendix D). The difference comes from the integral ∫0∞(1+τ)−(K−2),dτ\int_0^\infty(1+\tau)^{-(K-2)}\\,d\tau, which diverges for K≤3K\le3 but converges for K≥4K\ge4. The same argument therefore explains both recovery at the boundary and the slowdown approaching it.

“The idea is not very new, as there are some analogous analysis in [1][2].”

There is a connection to earlier work on escaping suboptimal softmax corners, and we have made it more explicit in the Introduction and Appendix G.7. Escaping the Gravitational Pull of Softmax and Delightful Gradients Accelerate Corner Escape study slow escape under fixed rewards or modified gradients. Our setting has an additional difficulty: the evaluator's score margin can itself shrink as the policy changes. As a result, the total corrective drift can be finite even when the optimum remains correctly ranked and is compared infinitely often. The policy can then converge to a wrong vertex rather than merely recover slowly.

AuthorsOpenReview ↗

Connections to Preference Learning

“The results are too specific and limited. One may expect to see the application of this idea in preference learning.”

The original analysis focused on real-valued contrasts and a particular evaluator family. To see whether the failure mechanism extends beyond this setting, we added Proposition 4 (Section 3.1), which gives a local failure criterion that also applies to Bradley–Terry binary feedback. For K≥4K\ge4, the optimal action can beat every alternative with probability above 1/21/2 at every interior policy, yet the sampled update can converge to a suboptimal vertex with high probability from a suitable open set of initializations. The actor remains a finite tabular softmax policy.

We also tested the mechanism with a less structured evaluator, using PRM800K mathematical reasoning steps and human ratings. The text-kernel evaluator is refitted under policy-dependent weights as the actor changes. We compare the sampled update (U), full-information update (FI), and balanced-query update (BQ). After 100,000 scalar evaluator calls, the mean optimal-action masses are:

ProblemsUFIBQ
All 37 eligible official-test problems0.0280.9190.657
16 initially correctly ranked problems0.0631.0000.688

The 16-problem group was fixed before the actor runs. Across all 128 U trajectories in this group, the evaluator ranks the optimum first at every visited policy. Yet on 15 of the 16 problems, none of eight U runs reaches p≥0.9p\ge0.9. The median score margin falls from 0.0132 to 0.00060, while the probability of comparing the optimal action falls from 0.0198 to 0.00113. Since the evaluator is fitted using each task's offline human ratings, this is a controlled study on real reasoning text rather than live LLM retraining (Section 6; Appendix H.10).

“For example, one can try to relate the method with some standard preference learning practice, such as [3], which studies a similar pairwise RL training objective, and [4], which studies the tabular softmax model in DPO.”

Section 5.2 now makes this connection precise at the update level. For a sampled pair (i,j)(i,j) with real-valued contrast CtC_t, the relative-reward regression loss can be written in tabular softmax logits as

Lt(θ)=[(θi−θj)−(θt,i−θt,j)h−Ct]2.L_t(\theta)=\left[\frac{(\theta_i-\theta_j)-(\theta_{t,i}-\theta_{t,j})}{h}-C_t\right]^2.

At θt\theta_t, a gradient step of size ηth/4\eta_t h/4 gives exactly θt+1=θt+(ηt/2)Ct(ei−ej)\theta_{t+1}=\theta_t+(\eta_t/2)C_t(e_i-e_j), our update (U). If the conditional mean contrasts equal the evaluator score differences, exact minimization of the expected loss over a full-support pair distribution gives the noiseless full-information multiplicative update with step hh. A DPO step with its reference reset to the current policy gives the binary-feedback form of (U). These identities concern the specified updates, not the full REBEL algorithm or standard fixed-reference DPO. Reference [4] studies fixed-reference DPO with fixed Bradley–Terry rewards and a KL-regularized target, whereas our target is the optimal vertex under policy-dependent evaluation (Appendix G.7).

“It's also unclear whether the proposed Balanced-query recovery is a major contribution as it is only covered in theorem 5 without further detailed discussions, such as how to deploy it in practice.”

The purpose of balanced queries is to distinguish failure of the sampled update from loss of information in the evaluator. We have expanded Section 5.3 (now Theorem 7) to explain this distinction and the access that BQ requires. Instead of evaluating pair (i,j)(i,j) at the actor's current policy, BQ uses μij=(ei+ej)/2\mu_{ij}=(e_i+e_j)/2. The dispersion bound then gives ∣cij−(ri−rj)∣≤2ε|c_{ij}-(r_i-r_j)|\le2\varepsilon, so the optimal action retains a positive pairwise margin whenever ε<Δ/2\varepsilon<\Delta/2, without a sensitivity assumption. Theorem 7 proves recovery up to this evaluator-only identification limit.

BQ requires access to the evaluator at counterfactual policies, which a fixed judge API does not automatically provide. In the PRM800K experiment, we implement this access by reweighting and refitting the evaluator. We also compared the updates under matched evaluator-call budgets. On quadratic compression, BQ reaches p≥0.9p\ge0.9 in 59/64 runs versus 28/64 for U after 200,000 calls, although its global πmin⁡\pi_{\min} factor makes it slower in other settings (Section 6; Appendix H.8).

We would be glad to clarify any remaining points.

Thank you for your rebuttal. I appreciate the reorganization of the paper. The rate analysis is better than previous results. However, I still have concerns remained:

  • As the authors admitted, "Since the evaluator is fitted using each task's offline human ratings, this is a controlled study on real reasoning text rather than live LLM retraining", the supplemental PRM800K experiments are still toy experiments with only four actions, while in the practice of REBEL and online DPO, the number of actions can be much larger than 44. For the same reason, the theorems showing a different between K=2,3K=2,3 and K≥4K\ge 4, as a separate theoretical tabular result, are not very interesting to me.
  • This paper points a theoretically-identified issue in preference training when K≥4K\ge 4, however, it's not identified in practical preference learning training beyond the 4-action toy experiments. What's more, the authors proposed BQ as a remedy, while the practical effect of BQ is still unclear. This further limits this paper's contribution.
Reviewer_To4NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the problem of recovering the optimal action in a self-rewarding model, under constraints on the probability of selecting the optima action, as well as on the policy dependent score-margins. The authors provide a class of finite-action examples that show that, depending on some structural parameters of the problem (bias-dispersion bound, minimum true reward gap, a Lipschitz condition), different conclusions follow. Depending on the bias-dispersion bound, the policy converges to the optimal action or concentrates on sub-optimal choices. At the threshold, the behaviour is additionally governed by the total number of actions. The idea of the paper is to provide a simple explicative abstract example, with no extra-frills, to be able to understand the basic mechanisms that govern these phenomena.

Critical strengths

The author identify a sharp recovery threshold (Theorem 1). With lower bias dispersion bound, policies converge to the optimal action, while (this is more interesting) above the threshold policies may converge to sub-optimal choices.

Failure is observed in the model even if the model include a consistent ranking of the actions

Failure happens also at the recovery threshold (Theorem 2), depending on the number of actions. A thorough explanation of the behaviour of the model at the threshold is also given.

Failure is shown to depend on structural properties of the model, such as the update rule, and not on the evaluator.

Critical weaknesses

Generalization value of the example: it is unclear which of the author's findings in their model are of general value, and which are specific of their model. A valuable addition to the paper would be to find some of these conclusions in non-synthetic models. This is addressed as a limitation in the Conclusions, but I feel this is a too strong limitation.

Experiments could have provided insights on the paper's findings on more realistic models, on which the analysis carried on by the authors could be be more complex. Unfortunately, the experiments only illustrate the conclusions of the paper.

Failure conditions require special constructions, and it is unclear how typical are such constructions.

The authors could have provided a more comprehensive literature review, and could have placed the problem from a broader perspective.

Questions for authors

Do the authors have any (even preliminary) result extending their analysis to more realistic frameworks?

Do they have experiments investigating the conclusions of their paper on more realistic models?

How typical is failure at and above the threshold? Do you have results/esperiments in this direction?

1 comment
AuthorsOpenReview ↗

Generality and Empirical Evidence

We appreciate the reviewer's focus on generality and evidence beyond synthetic examples. The revision adds a local failure criterion, two new evaluator families, and an experiment using human-rated reasoning steps.

Q1. Does the analysis extend to more realistic frameworks?

Partly, at the level of comparison feedback. The new Proposition 4 keeps the finite tabular-softmax actor but no longer requires the original evaluator formula. For K≥4K\ge4, suppose that at a suboptimal vertex the optimal and suboptimal actions have tied scores, the remaining auxiliary actions have strictly lower scores, and pairwise score differences are Lipschitz. Then an open set of nearby initializations leads to wrong-vertex convergence with arbitrarily high probability. So failure needs only this local structure, not our specific score formula (Section 3.1).

The same argument applies to Bradley–Terry binary labels: if the win probability is [1+exp⁡(−a(si−sj))]−1[1+\exp(-a(s_i-s_j))]^{-1}, the signed label has mean tanh⁡(a(si−sj)/2)\tanh(a(s_i-s_j)/2). This link is zero at a score tie and is Lipschitz in the score gap, so the local argument still applies. With such labels, the optimum in the critical construction wins every comparison with probability above 1/21/2 at every interior policy, but the update can still converge to the wrong vertex. We also show one-step connections with relative-reward regression and reset-reference DPO (Section 5.2).

Q2. Are there experiments using more realistic data?

Yes. We use PRM800K candidate mathematical reasoning steps with human ratings. For each problem, a text-kernel ridge evaluator is refitted from the task's offline human labels under policy-dependent weights, and the actor runs over four candidate steps. This gives a controlled experiment on real reasoning text rather than LLM retraining. We include all 37 eligible official-test problems. Before any actor run, the initial evaluator misranks the optimum on 21 of them, so the correct-ranking analysis focuses on the other 16. The table reports mean optimal-action mass (not success rates) after 100,000 logical scalar score calls:

ProblemsUFIBQ
All 37 eligible problems0.0280.9190.657
16 initially correctly ranked0.0631.0000.688

On all 128 U trajectories in the 16-problem group, the evaluator ranks the optimal step first at every visited policy. Yet on 15 of these 16 problems, none of eight U runs reaches p≥0.9p\ge0.9. The diagnostics show both effects: the median score margin falls from 0.0132 to 0.00060, the distinct-pair comparison probability of the optimal step from 0.0198 to 0.00113, and its conditional logit drift from 3.3×10−43.3\times10^{-4} to 3.2×10−73.2\times10^{-7} (Figure 2c–d; Appendix H.10).

Q3. How typical is failure at or above the threshold?

Failure is not confined to the original critical evaluator or a single initial policy, but at finite budgets it varies sharply with initialization. Both new families rank the optimal action first at every interior policy; one is at the threshold and one slightly above it. The table counts runs reaching p≥0.9p\ge0.9 after two million rounds. The selected initial policies are (0.03,0.77,0.1,0.1)(0.03,0.77,0.1,0.1) for quadratic compression and (0.1,0.7,0.1,0.1)(0.1,0.7,0.1,0.1) for conformity.

Evaluator (ratio)Selected startDirichlet(1,1,1,1) starts
Quadratic, a=1 (1.035)0/6499/128
Conformity, a=1 (1.000)0/64105/128

Most sampled Dirichlet initializations reach the target within this budget, while the selected starts stall. For both families, Proposition 4 also proves asymptotic wrong-vertex convergence from nonempty open sets of initializations (Appendix H.7).

Related work. We expanded the comparison with fixed-reward softmax optimization, REBEL, online DPO, and performative learning (Introduction; Appendix G.7).

We hope the new results address your concerns. We welcome any further questions or discussion.

Reviewer_ELX5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how policy-dependent evaluators affect sampled pairwise softmax updates. It derives a bias threshold for recovering the optimal action. At the threshold with positive sensitivity, recovery is guaranteed for two or three actions, but can fail with four or more, even when the evaluator always ranks the optimal action first. The paper explains this failure mechanism and uses full-information updates and balanced queries to distinguish failure of the update from insufficient information in the evaluator.

Critical strengths
  1. The theoretical characterization is fairly complete. Beyond sufficient conditions for recovery, the paper constructs failures at and above the threshold. The distinction between three and four actions comes with both a proof and a clear explanation.

  2. The paper carefully distinguishes eventual convergence from finite-time behavior and examines the roles of noise, step sizes, and feedback formats. Repeated simulations and parameter sweeps support the behavior of the constructed examples.

Critical weaknesses
  1. The main threshold experiments remain concentrated on one evaluator family. Appendix G.5 varies bias, sensitivity, action count, and noise, but uses the same evaluator family and initialization pattern. This illustrates the counterexample well, but leaves open how often similar failures occur with other evaluators and initializations. Some additional experiments would help establish the broader relevance of the phenomenon.

  2. The alternative updates lack an empirical comparison. Full-information updates and balanced-query updates are important parts of the paper, but the simulations mainly evaluate the original update U. In particular, BQ scales its updates by the minimum action probability, which the authors acknowledge can slow recovery. It is therefore unclear whether its wider theoretical recovery range translates into an advantage under a finite query budget.

3。 The asymptotic guarantees provide limited guidance on how long recovery takes. For example, in Appendix G.5, the six-action setting below the threshold still has a median optimal-action probability of about 0.00168 after two million rounds. This does not contradict the theorem, but shows that eventual recovery can be far from useful finite-budget performance. The finite-round guarantee applies to identification through uniform balanced queries, rather than the recovery time of U or BQ.

Questions for authors
  1. Do similar failures occur with different initial probabilities or evaluator families beyond the critical counterexample construction? Could the authors provide a few representative results?

  2. How do U, the full-information update, and BQ compare under similar evaluator-call budgets? In particular, how much does minimum-probability scaling slow BQ as the number of actions increases?

  3. For settings below the threshold where recovery is slow, could the authors clarify how recovery time depends on the distance to the threshold? Results across training budgets, or a time bound under stronger noise assumptions, would be helpful.

2 comments
AuthorsOpenReview ↗

Additional Experiments and Recovery-Time Bounds

Thank you for the concrete questions. Our original threshold experiments focused on one evaluator family and did not compare the updates at matched budgets. We have now added two evaluator families, matched-budget comparisons, and a first-hitting theorem (Theorem 2).

Q1. Other evaluators and initializations.

Yes. We tested quadratic score-margin compression and a single-action conformity bonus. Both have analytically computed bias-dispersion and sensitivity constants and rank the optimal action first at every interior policy. At amplitude a=1a=1, conformity is at the threshold and quadratic compression slightly above it (ratio 1.035). After two million sampling rounds, 0/64 quadratic runs from (0.03,0.77,0.1,0.1)(0.03,0.77,0.1,0.1) and 0/64 conformity runs from (0.1,0.7,0.1,0.1)(0.1,0.7,0.1,0.1) reach p≥0.9p\ge0.9. The outcome depends strongly on initialization, though: from 128 independent Dirichlet(1,1,1,1) initializations, 99 quadratic and 105 conformity runs reach the target (maps in Appendix H.7). These counts describe finite-budget behavior. Proposition 4 gives the asymptotic result: for both evaluators, the update can converge to a wrong vertex with high probability from suitable open sets. The proof uses a local score tie and strictly worse auxiliary actions, not the original evaluator formula (Section 3.1).

Q2. U, FI, and BQ at matched budgets.

To match budgets, we count two scalar score calls per distinct-pair contrast, KK per full-information (FI) step, and zero for duplicate draws. With K=4K=4, initialization (0.1,0.7,0.1,0.1)(0.1,0.7,0.1,0.1), and 200,000 logical scalar calls, the numbers of runs reaching p≥0.9p\ge0.9 are:

EvaluatorUFIBQ
Quadratic compression (a=1)28/64Recovered (det.)59/64
Critical family (ratio 1)2/64Recovered (det.)0/64
Critical family (ratio 0.95)64/64Recovered (det.)0/64
Conformity (a=1)0/64Recovered (det.)0/64

FI is deterministic in these synthetic settings, so its column reports a single trajectory. BQ helps on quadratic compression, but on the below-threshold critical instance (ratio 0.95) U recovers in all 64 runs and BQ in none.

The slowdown comes from scaling BQ's inverse-probability correction by the global πmin⁡\pi_{\min} to bound the step. Even one rare suboptimal auxiliary action matters: initialized at 10−610^{-6} with fixed correct scores, it leaves BQ's median optimal mass at 0.101 after 100,000 calls, while U and FI reach 0.9 in all 64 runs. On the critical family at the threshold (ratio 1), BQ's median optimal mass after 200,000 calls falls from 0.793 to 0.722 to 0.645 as KK increases from 4 to 6 to 10.

Matching calls does not match actor work: on quadratic compression, 200,000 calls take a median of 2.5×1092.5\times10^9 actor rounds for U, 5.7×1055.7\times10^5 for BQ, and 5×1045\times10^4 for FI. Since the three methods also have different oracle access, Appendix H.8 reports equal-round controls as well.

Q3. Recovery time and distance to the threshold.

Let κ=Δ−ΨL−1(ε)>0\kappa=\Delta-\Psi_L^{-1}(\varepsilon)>0. No stronger noise assumption is needed: under the original conditional finite-variance assumption, the new Theorem 2 bounds the first-hitting budget of U in cumulative step size by O(κ−(K−1))O(\kappa^{-(K-1)}) with high probability. For each fixed K≥4K\ge4, a correctly ranked subcritical family attains the matching lower-bound order from a fixed open set of initializations. For K≤3K\le3 and fixed L>0L>0, one finite first-hitting budget suffices uniformly up to the threshold.

We extended the six-action experiment to two billion sampling rounds (32 runs per threshold ratio):

Threshold ratioMedian rounds to first reach p≥0.9Reached within 2 billion
0.901.08 million32/32
0.95305 million32/32
0.98>2 billion (censored)0/32

Moving from ratio 0.90 to 0.95 increases the median first-hitting time by more than two orders of magnitude. At ratio 0.95, the median optimal mass is still about 0.00165 after two million rounds, but all 32 runs later reach the target. The 0.98 runs are right-censored at this budget. As in Theorem 2, these are first-hitting times, not times after which pp stays above 0.9 (Section 2.3; Figure 2b; Appendix H.9).

We hope these responses address your concerns, and we would be happy to clarify any remaining questions.

Thank you for addressing my questions. The new results resolve my main concerns, and I’ll maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.