acceptodds
Under review as a conference paper at ICLR 2027

Value-Free Policy Optimization via Reward Partitioning

Abstract

Single-trajectory preference optimization methods learn from datasets of ((prompt, response, reward)) tuples, offering a practical alternative to pairwise preference learning by directly leveraging scalar feedback. Existing approaches such as Direct Reward Optimization (DRO) have demonstrated promising results but rely on value function estimation, introducing additional variance, optimization complexity, and sensitivity to off-policy data. We introduce Reward Partition Optimization (RPO), a simple and scalable reward-driven objective that eliminates the need for value function learning. RPO normalizes rewards through a partition-based formulation estimated directly from prompt-level reward distributions, yielding a stable supervised optimization objective without auxiliary models or reinforcement learning loops. We evaluate RPO across multiple encoder-decoder and decoder-only language models using automatic metrics, LLM-as-a-judge evaluations, and optimization stability analyses. Experimental results show that RPO consistently outperforms strong baselines, including SFT, KTO, and DRO, while producing more aligned, diverse, and less toxic generations.

Primary area:
reinforcement learning
Keywords:
Large Language Models, Preference Optimization, Single-Trajectory Learning, Value-Free Reinforcement Learning, Reward Partitioning, Empirical Normalization, Offline Policy Optimization, Direct Reward Optimization, LLM Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sCKMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a single-trajectory offline policy optimization method for LLM alignment. The method leverages the closed-form optimal policy for the KL-regularized RL objective to construct a policy-fitting objective that utilizes pre-existing offline completions present in a training dataset to provide a single-trajectory learning signal. The method is supported by an empirical evaluation on the UltraFeedback dataset.

Critical strengths

None

Critical weaknesses
  1. Lack of novelty. The paper doesn't mention or acknowledge numerous papers that tackle exactly same issue and use very similar or sometimes an almost identical approach, e.g. A*-PO paper [1] that also approximates the value function V(x)=τlog⁡Z(x)V(x) = \tau \log Z(x) from offline completions and uses the same squared loss, SPO [2], QRPO [3], AGRO [4], TBA [5], etc that also optimize the same squared loss while handling the normalization term in different ways. Therefore, the core regression objective and precomputed-normalizer approach are already present in prior work, including a very similar methods like A*-PO.
  2. Error in the theory. The proposed estimator for the partition function in eq. 13, namely Z^(x)=∑j∈Ixπref(yj∣x)exp⁡(rjτ)\hat{Z}(x) = \sum_{j \in \mathcal{I}_x} \pi_{ref} (y_j \mid x)\exp\left(\frac{r_j}{\tau}\right), is not a correct monte-carlo estimate for the partition function in eq. 11. A simple argument demonstrating the error is that using that estimator the total probability mass of the completions from Ix\mathcal{I}_x under the policy obtained by exactly fitting the proposed regression targets is 1, thereby assigning zero probability to all completions outside Ix\mathcal{I}_x. A correct one would be Z^(x)=1∣Ix∣∑j∈Ixπref(yj∣x)μ(yj∣x)exp⁡(rjτ)\hat{Z}(x) = \frac{1}{|\mathcal I_x|} \sum_{j \in \mathcal{I}_x} \frac{\pi_{ref} (y_j \mid x)}{\mu (y_j \mid x) } \exp\left(\frac{r_j}{\tau}\right), where μ\mu is the probability distribution from which the completions in Ix\mathcal{I}_x were sampled. The fact that μ\mu is generally unknown in offline data makes the application of the proposed method very limited, therefore some other methods mentioned above set μ\mu to a known policy, e.g. μ=πref\mu = \pi_{ref} as in A*-PO [1] that uses reference model completions to estimate ZZ.
  3. Level of noise in Z^(x)\hat{Z}(x). There is no analysis of the noise amplitude in the proposed estimate of Z^\hat{Z}, however, the QRPO paper [3] provides such a theoretical analysis and shows that in case the reward distribution has a right tail the noise level is catastrophic. The current paper evaluates the method in a setting with bounded rewards, which may not reveal the severity of this problem. However, the method presentation is not restricted to bounded rewards, and the absence of such analysis leaves an important limitation unaddressed.

[1] Kianté Brantley, Mingyu Chen, Zhaolin Gao, Jason D. Lee, Wen Sun, Wenhao Zhan, and Xuezhou Zhang. Accelerating RL for LLM Reasoning with Optimal Advantage Regression. NeurIPS, 2025.

[2] Taco Cohen, David W. Zhang, Kunhao Zheng, Yunhao Tang, Rémi Munos, and Gabriel Synnaeve. Soft Policy Optimization: Online Off-Policy RL for Sequence Models. arXiv preprint arXiv:2503.05453, 2025.

[3] Simon Matrenok, Skander Moalla, and Caglar Gulcehre. Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions. NeurIPS, 2025.

[4] Yunhao Tang, Taco Cohen, David W. Zhang, Gabriel Synnaeve, and Rémi Munos. RL-finetuning LLMs from On- and Off-policy Data with a Single Algorithm. AISTATS, 2026.

[5] Brian R. Bartoldson, Siddarth Venkatraman, James Diffenderfer, Moksh Jain, Tal Ben-Nun, Seanie Lee, Minsu Kim, Johan Obando-Ceron, Yoshua Bengio, and Bhavya Kailkhura. Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training. NeurIPS, 2025.

Questions for authors
  1. Why does the KL divergence in figure 2 start from a non-zero value given that the paper states that the policy starts from the reference checkpoint? Why do different methods have different starting points?
  2. Table 1 contains a column with GPT-4o and Claude-3.5 win-rates of RPO checkpoint against various methods, and table 5 contains a similar column with each RPO density variant’s win rates against SFT. Those columns allow us to compare the win-rates between SFT and RPO and observe that 75% density setting has higher reported judge win rates than 100% for every model, despite the claim that performance improves monotonically with density. Can the authors explain that? Likewise, why does the noise in rewards (with σ=0.1\sigma=0.1) (table 6) improve those win rates for every model compared with the corresponding clean-reward RPO checkpoints?
  3. Which hyperparameter search ranges were used for each method?
2 comments

We sincerely thank the reviewer for the careful and technically substantive assessment. The concerns about prior work, partition estimation, and empirical consistency are important. We agree that RPO's relationship to related approaches and the theoretical interpretation of its empirical partition require better qualification. We address each point below.

  1. Novelty and relationship to prior work

We agree that the manuscript should have discussed A*-PO, SPO, QRPO, AGRO, and TBA. A*-PO and QRPO are particularly related to RPO's core ingredients: regression toward reward-derived policy targets and normalization. We will substantially revise the related work and narrow our novelty claims.

Our contribution is not squared-error policy regression or offline normalization in general. Rather, RPO studies an empirical construction using same-prompt candidates from scalar-feedback datasets to define a prompt-dependent normalization, without training a separate value network. We acknowledge that this distinction alone does not establish broad methodological novelty. We will compare the assumptions, sampling requirements, normalization, and optimization settings of these methods, positioning RPO as a specific empirical variant rather than the first method to introduce this principle.

  1. Partition estimator and theoretical scope

We appreciate the reviewer identifying this issue. The exact partition is

Z(x)=Ey∼πref[exp⁡(r(x,y)τ)].Z(x)=\mathbb{E}_{y\sim\pi_{\mathrm{ref}}}\left[\exp\left(\frac{r(x,y)}{\tau}\right)\right].

If completions are sampled from q(y∣x)q(y|x), an unbiased Monte-Carlo estimator generally requires

Z^(x)=1M∑j=1Mπref(yj∣x)q(yj∣x)exp⁡(rj/τ).\widehat Z(x)=\frac{1}{M}\sum_{j=1}^{M}\frac{\pi_{\rm ref}(y_j|x)}{q(y_j|x)}\exp(r_j/\tau).

RPO's current estimator omits this correction. Unless candidates come from the reference policy or suitable sampling weights are available, it is therefore not an unbiased estimator of the full reference-policy partition. We agree this limits the current theoretical presentation.

Instead, the construction defines a normalization over the observed candidate set. Thus, the exact-fitting argument applies to this empirical support, not necessarily the full response space. We will correct the derivation and distinguish this finite-support objective from the full KL-regularized optimal policy. Recovering the latter requires knowledge of the sampling distribution and appropriate importance weighting, or another principled estimator. This qualification materially narrows our theoretical claim.

  1. Noise in the partition estimate

We agree that bounded rewards do not eliminate the variance caused by exponentiation. QRPO's analysis of this issue should have been discussed. Our bounded, standardized UltraFeedback rewards demonstrate behavior only in that setting, not robustness to arbitrary or right-tailed reward distributions. We will discuss how estimation noise depends on reward dispersion, temperature, and candidate count, and limit our robustness claims to the tested regime.

  1. Questions and empirical inconsistencies

Q1. Non-zero initial KL. Identical initial policy and reference checkpoints must yield zero KL. We will clarify when the first measurement is taken and how KL is estimated. If the plotted values correspond to step zero, this is an inconsistency that must be corrected; identical initialization cannot explain non-zero values.

Q2. Density and reward-noise results. We agree that the reported judge win rates contradict the stated monotonic trend: 75% density outperforms 100%, while adding reward noise improves win rates across models. These results cannot support claims of monotonic improvement or straightforward robustness. We will audit checkpoint selection, evaluation, and table construction, report relevant variability, and revise the claims to match verified results. We will not attribute these differences to a mechanism without evidence.

Q3. Hyperparameter search. The manuscript reports a shared learning rate of 10−410^{-4}, 150 warmup steps, and default τ=1.0\tau=1.0, with an RPO temperature ablation over {0.1,1.0,5.0}\{0.1,1.0,5.0\}. However, it does not provide complete method-specific search ranges. We will document the configurations actually tested for each baseline and distinguish fixed settings from searched hyperparameters, without implying a broader search.

We thank the reviewer again. These issues require substantive corrections, not merely clarification. We will narrow the theoretical claims to what the estimator supports, acknowledge prior work, and revise experimental conclusions wherever the evidence does not justify the current wording.

Thank you for the detailed response. At this points I would maintain my score.

Reviewer_rL8VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Reward Partition Optimization (RPO), which eliminates the value model required by Direct Rewrad Optimization (DRO) for preference optimization. RPO computes each prompt’s normalization term directly from the rewards of that prompt’s candidate responses instead of learning it. It then trains the policy with a simple regression loss toward this reward-derived target, giving a fully offline supervised objective. Experiments on UltraFeedback with both encoder-decoder and decoder-only models show RPO outperforming SFT, KTO, and DRO on automatic metrics and LLM-as-a-judge win rates. The gains reportedly carry over to out-of-distribution benchmarks, and ablations support the method’s main design choices.

Critical strengths
  1. The motivation is clear and the formulation is simple. RPO estimates the value term offline from rewards within each prompt’s group of candidate responses, which removes the separate value model DRO requires and lowers training cost.
  2. The experiments cover six models across both encoder-decoder and decoder-only architectures. Ablations study temperature, per-prompt normalization, candidate density, and reward noise.
Critical weaknesses
  1. Outdated models. The experiments use fairly old backbones. The newest is Qwen2.5-7B, released in September 2024, about two years before this submission. But the paper cites the Qwen3 report but does not evaluate on Qwen3. It's unclear whether the method still works on advanced models.
  2. Missing comparison with other offline alignment methods. The baselines are limited to SFT, KTO, and DRO. UltraFeedback provides several scored responses per prompt, so it can easily be turned into preference pairs, yet the paper does not compare against standard offline methods such as DPO.
  3. Missing comparison with group-based RL methods. RPO's core idea is to replace a learned value model with statistics computed over a group of responses to the same prompt. GRPO also does't have the critic model and use the group average reward as baseline. The paper neither discusses nor compares against them, for example an offline version of GRPO trained on the same candidate groups.
  4. Limited novelty. Estimating the soft value as a log-mean-exp of rewards over sampled responses, and then regressing the policy's log-ratio toward the resulting target, closely matches existing work that the paper does not cite. For example, A*-PO (Brantley et al., 2025, NeurIPS 2025) uses the same two steps. It first estimates the optimal value offline as a log-mean-exp of rewards over responses sampled from the reference policy, then trains the policy with a least-squares regression loss.
Questions for authors
  1. Single-candidate case. When a prompt has only one candidate (e.g., the 25% density setting), the target appears to reduce to the negative reference log-probability, which no longer depends on the reward. How does RPO behave in this case, and why does it still clearly outperform SFT in Table 5?
  2. KL at initialization. All policies are initialized from the reference model, so the KL divergence at step 0 should be zero. Figure 2, however, shows initial values of roughly 0.85–1.6. How is the KL computed, and over which distribution?
  3. Relation to prior work. How does RPO differ, both conceptually and empirically, from A*-PO? Can the authors add comparisons with it, as well as with DPO and an offline GRPO-style group baseline?
  4. Recent models and standard metrics. Can the authors report results on a recent backbone such as Qwen3 or Qwen3.5?
2 comments

We sincerely thank the reviewer for the constructive assessment and for recognizing the simplicity, motivation, and empirical scope of RPO. We appreciate the suggestions concerning stronger baselines, recent model families, and related work. We address each point below.

  1. Evaluation on recent backbones

We agree that evaluating RPO only on models up to Qwen2.5-7B limits our generalization claims. The current experiments demonstrate its applicability across encoder-decoder and decoder-only architectures, but not to newer families such as Qwen3. We will narrow our claims accordingly and, subject to additional experiments, extend the evaluation to a recent Qwen3 checkpoint using the same protocol. We will also report standard benchmark metrics to facilitate comparison with existing methods.

  1. Comparison with offline alignment methods

We agree that DPO is an important missing baseline. UltraFeedback's multiple scored responses can be converted into preference pairs, enabling comparison under the same prompts and candidate pool. We will include DPO with a clearly specified pair-construction strategy and matched base checkpoints and evaluation protocol. This will help determine whether RPO's gains come from scalar-reward normalization or simply from additional supervision through grouped candidates.

  1. Relation to group-based methods such as GRPO

RPO and GRPO share the intuition that responses to the same prompt provide useful relative information without a separate critic. However, GRPO typically uses newly sampled on-policy groups and group-relative advantages, whereas RPO uses fixed offline candidates and reward-derived targets. We will clarify this distinction.

An offline group-relative baseline using the same UltraFeedback groups would provide an informative comparison. We will add it if feasible and clarify that avoiding a learned critic is not, by itself, RPO's distinguishing contribution.

  1. Novelty and relationship to A*-PO

We agree that A*-PO is closely related and should have been discussed. Both methods use offline responses to estimate a reward-dependent normalization and fit the policy through squared regression. We therefore do not claim that either component is novel in isolation.

RPO's intended distinction is its empirical partition construction from observed same-prompt candidate sets in scalar-feedback datasets, rather than requiring candidates sampled from the reference policy. This also entails different statistical assumptions and limitations, particularly concerning the candidate sampling distribution. We will add a conceptual comparison with A*-PO and revise our novelty claims. An empirical comparison would further clarify their practical differences.

  1. Questions

Q1. Single-candidate case.

The reviewer is correct. If a prompt has exactly one candidate, our definition gives Z^(x)=πref(y∣x)exp⁡(r/τ),\widehat Z(x)=\pi_{\mathrm{ref}}(y|x)\exp(r/\tau), and hence V^(x)=r+τlog⁡πref(y∣x).\widehat V(x)=r+\tau\log\pi_{\mathrm{ref}}(y|x). The resulting target is

=-\log\pi_{\mathrm{ref}}(y|x).$$ Thus, the reward cancels and the target contains no reward information. In this degenerate case, RPO instead encourages fitting the observed completion. We will clarify that meaningful reward partitioning requires multiple candidates per prompt and investigate the candidate counts at each density. The 25% density result should not be taken as evidence that a single candidate is sufficient. Q2. KL at initialization. We agree that identical policy and reference models must have zero KL at initialization under a consistent computation. The current figure does not adequately document this measurement. We will recompute the curves from actual training checkpoints, specifying the evaluation prompts, sampling procedure, and KL estimator. We will not interpret the current non-zero initial values as evidence of policy divergence. Q3. Additional comparisons. Comparisons with A*-PO, DPO, and an offline group-relative baseline would strengthen the evaluation. We will prioritize these comparisons using matched data, base models, and evaluation settings. Q4. Recent models and standard metrics. We agree that a recent backbone and standard benchmarks would strengthen the paper. We will clarify the scope of the current results and extend the evaluation where feasible. We thank the reviewer again. These suggestions will help us better establish RPO's scope, distinguish it from related approaches, and assess its practical value more rigorously.

Thanks for the authors' response, and hope the authors can further polish the manuscript when submitting it to other venues in the future.

Reviewer_egQ2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces the Reward Partition Optimization (RPO) for LLM alignment from preference feedback.

Setting: we are given some datapoints indicating preferences for certain LLM responses over others. For example, this data could be of these forms:

  • (prompt, response_w, response_l), where a user indicated they prefer response_w to response_l when shown these two responses for the prompt.
  • (prompt, response, score), where the score is a number indicating preference. We seek to train an LLM to align better with these preferences. The paper concerns the second setting when richer, multiple preference data is available: suppose given a single prompt, there are multiple (response, score) pairs. The papers argues that this can be used to improve upon existing methods. They introduce an algorithm called Reward Partition Optimization (RPO).

They build upon the existing Direct Reward Optimization (DRO) method, which essentially learns the LLM weights as well as an "effective" reward model/value function simultaneously as a joint optimization problem (see Section 2.3). The paper argues that the value function corresponds to a normalization term, and can be computed empirically when multiple preference data mentioned above is available.

In their experiments, they compare their method RPO with SFT (supervised fine-tuning), DRO, and KTO (Kahneman–Tversky Optimization) on the UltraFeedback dataset. They present comparisons across a range of metrics for both in-domain and out-of-distribution data. Their method has a small but consistent advantage throughout for the set of results presented in the main body.

Critical strengths

The paper makes a clear and useful observation about how to handle the value term that arises in KL-regularized reward optimization. In prior work such as DRO, the value function is introduced as an additional learned object and is optimized jointly with the policy. The authors point out that, in this setting, the value function is not an arbitrary auxiliary quantity: at the optimum it is exactly the "log-partition" term. When multiple rewarded responses are available for the same prompt, this normalization term can therefore be estimated directly from the empirical reward distribution instead of being represented by a separate parameterized model. This leads to a particularly simple modification of DRO: replace learned value estimation with an explicit empirical partition computation, and optimize only the policy.

This observation is conceptually clean and practically appealing. It removes one source of function-approximation error and eliminates the coupling between policy and value-function optimization, while preserving the same KL-regularized optimal-policy structure that motivates DRO. The resulting RPO objective is fully supervised, requires only a single trainable model, and is straightforward to implement. The paper also provides empirical evidence that this simplification can be useful in practice: RPO trains faster than DRO and KTO in their experiments, and the ablation removing prompt-level normalization consistently degrades performance, supporting the claim that explicitly estimating the normalization term is an important component of the method.

Critical weaknesses
  1. I have some concerns about the experiments.
  • There are no error bars/standard deviations for the numbers presented. Given that the reported gains are often modest (see Table 1 for example), it is unclear if these gains are within the margin of error.

  • The gains against baselines (DRO, SFT, KTO) are only presented for one set of parameters. For example, the best temperature parameter for RPO is chosen. For other temperatures, only RPO's numbers are reported in the appendix. This does not rule out that DRO might (for example) be better at a temperature that is sub-optimal for RPO. The paper states that the ablations are to support the design choices, in Appendix A.5: "Across the four ablations, the results support the main design choices of RPO."

  • There is substantial instability in the RPO policy as it trains (Figure 2), indicating that training has not perhaps completed.

  1. Datasets where multiple (response, score) pairs are available for the same prompt seem uncommon. The paper itself only considers a single dataset. While training itself is a bit cheaper for their method (RPO), it is unclear if generating such datasets makes the whole process significantly more expensive.

(Minor) There are also some typos/writing issues: the citations never use parentheses, there is a missing reference in the main algorithm, etc. These issues made the paper slightly more difficult to read.

(Minor) Both tables say "win rate against RPO", followed by stating that numbers larger than 50 favor RPO. This is technically correct, but slightly confusing; please change the phrase "win rate against RPO".

Questions for authors
  1. Are the gains against the baseline sustained for other parameter choices (e.g., those reported in the appendix)?

  2. Do you know why RPO produces longer answers? Is this true for the preference dataset (that longer answers are preferred)?

  3. How is reward standardization combined with the partition computation? The paper states that UltraFeedback rewards are standardized to zero mean and unit variance before training. Is Z^(x)\widehat{Z}(x) computed using these standardized rewards or the original 0-10 rewards?

  4. How is the log-ratio in eqn (16) implemented?

1 comment

We sincerely thank the reviewer for the thoughtful and constructive feedback. We particularly appreciate the recognition of RPO's central insight: replacing learned value estimation with an empirical reward partition provides a simple and principled alternative to DRO. We address the concerns and questions below.

  1. Experimental reliability and sensitivity

Error bars and statistical significance. We agree that reporting variability is important, particularly when performance differences are modest. We will strengthen the experimental analysis by reporting mean and standard deviation across independent runs with different random seeds. This will help establish whether the observed improvements are statistically robust rather than artifacts of a particular run.

Fairness of hyperparameter comparisons. We agree that comparing methods at a single temperature could obscure differences in their respective optimal configurations. Importantly, the temperature controls the reward-to-policy scaling and its optimal value need not be shared across objectives. A fair comparison should therefore tune each method independently under the same validation protocol and computational budget. We will clarify this distinction and strengthen the comparison accordingly.

Training stability (Figure 2). We appreciate this observation. The KL curves illustrate policy divergence, but KL stabilization alone does not establish convergence of the optimization objective. We will clarify the distinction and provide additional training-loss and convergence analyses to better characterize RPO's optimization dynamics.

  1. Availability and cost of grouped preference data

We agree that obtaining multiple responses per prompt introduces an additional data-generation consideration. However, RPO does not require additional human annotations: it exploits multiple scored candidates already available in datasets such as UltraFeedback. Importantly, these candidates can be generated and scored offline, independently of policy optimization. Thus, the relevant comparison should distinguish dataset construction cost from training cost. RPO's main advantage is that it eliminates the additional value-network optimization required by DRO, rather than claiming that collecting grouped data is cost-free.

  1. Questions

Q1. Are the gains sustained across other parameter choices?

The temperature ablation indicates that RPO remains effective at τ=5.0\tau=5.0, although performance degrades at τ=0.1\tau=0.1. We agree that the strongest evidence would come from comparing all methods across independently tuned hyperparameters. We will clarify the scope of the current results and strengthen this comparison.

Q2. Why does RPO produce longer answers?

A plausible explanation is that RPO directly fits reward-derived targets through sequence-level log-probabilities. If higher-reward responses tend to contain more detailed information, this supervision can favor more complete generations. However, we agree that response length alone does not establish improved quality. We will clarify this interpretation and distinguish verbosity from actual alignment improvements.

Q3. How is reward standardization combined with the partition computation?

We use standardized rewards throughout the RPO objective, including the empirical partition:

w_j(x)\exp(r_j^{\mathrm{std}}/\tau).$$ Consequently, $\widehat V(x)$ and the normalized targets are computed using the same standardized reward scale. This ensures consistency between reward normalization and policy supervision. Q4. How is the log-ratio in Equation (16) implemented? It is computed using the sequence-level conditional log-probability: $$\log\frac{\pi_\theta(y|x)}{\pi_{\mathrm{ref}}(y|x)}=\sum_{t=1}^{T}\left[\log\pi_\theta(y_t|x,y_{<t})-\log\pi_{\mathrm{ref}}(y_t|x,y_{<t})\right].$$ Both models are evaluated on the same input-response sequence, with padding tokens excluded from the sum. The reference model is frozen throughout training. Finally, we thank the reviewer for pointing out the presentation issues. We will correct the terminology surrounding win rates, improve citation formatting, and fix the missing algorithm reference. We hope these clarifications address the concerns and demonstrate that RPO offers a practically useful and conceptually principled approach to single-trajectory reward optimization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.