acceptodds
Under review as a conference paper at ICLR 2027

SP-GRPO: Shared-Prefix Group Relative Policy Optimization for Multi-Turn Agents

Abstract

Reinforcement learning post-training, notably the critic-free Group Relative Policy Optimization (GRPO), has become a cornerstone for enhancing the reasoning capabilities of large language models. However, extending its group comparison paradigm to long-horizon, multi-turn agent tasks remains challenging, where sparse terminal rewards fail to provide fine-grained, turn-level credit assignment. Existing methods typically attempt isolated fixes by either modifying credit assignment through turn-level advantages calculated across independently rolled-out trajectories, or altering rollout structures via tree expansions. By treating rollout dynamics and credit assignment as decoupled components, these approaches suffer from a core structural misalignment, introducing severe context-confounding bias. To resolve this structural misalignment, we propose Shared-Prefix Group Relative Policy Optimization (SP-GRPO), which seamlessly aligns single-path environment continuation with turn-level counterfactual evaluation. By concentrating the sampling budget on multi-action branching from strictly identical historical states, SP-GRPO evaluates candidate actions under unbiased context, computes exact process advantages via instantaneous feedback, and advances the environment through a single active continuation path. Theoretically, we prove that SP-GRPO strictly preserves the on-policy trajectory distribution, eliminates context confounding bias, and achieves provable variance reduction over cross-trajectory estimators. Empirically, extensive experiments across seven single-hop and multi-hop QA benchmarks show that SP-GRPO significantly outperforms trajectory-level GRPO, tree-based approaches, and existing step-level baselines.

Primary area:
reinforcement learning
Keywords:
Multi-Turn Agent, Turn-Level Advantage, Credit Assignment

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_iyRuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SP-GRPO, which is a search-based agentic RL algorithm enhanced with process advantages. Specifically, SP-GRPO collects the relative advantages of intermediate steps via additional sampling and uses these signals for a later RL framework. Extensive experiments validate the effectiveness.

Critical strengths
  1. This paper introduces process advantages into agentic RL in a simple way.
  2. The experiments in this paper are solid and fully support the claims.
Critical weaknesses
  1. The explanation of the method in this paper lacks sufficient visual clarity. In fact, I spent a significant amount of time trying to grasp the content in Section 3. I believe it would be beneficial to include more illustrative examples and figures to help readers understand the shared-prefix sampling process, as well as how the "advantage" is calculated. The single diagram currently provided (Figure 1) is not enough.
  2. There is already a substantial body of work on search-agent RL (such as https://arxiv.org/pdf/2510.24302, https://arxiv.org/pdf/2507.19849, https://arxiv.org/abs/2602.11767). The primary contribution of this paper lies in synthesizing existing process-based rewards, group-relative advantages, and branch sampling into a coherent framework. Its unique design centers on a combination of "comparing candidates that share the same prefix," "continuing with a single randomly sampled path," and "defining the corresponding scope for updates." Notably, the concepts of process supervision, information-gain rewards, and the PPO/GRPO objectives themselves are not original to this paper. This makes the novelty less significant.
Questions for authors

In my view, the essence of the success of the proposed method lies in trading more computation for more precise supervision signals. Could you further discuss the advantages of these methods under the constraint of a fixed total computational cost?

2 comments

Response to Reviewer iyRu

We thank the reviewer for acknowledging that SP-GRPO introduces process advantages into agentic RL in a simple way, and that our experiments are solid and fully support the claims. We respond to each concern below.

W1: Section 3 lacks visual clarity. Figure 1 alone is insufficient. Please add illustrative examples and figures that explain the shared-prefix sampling process and how the advantage is computed.

We thank the reviewer for this constructive suggestion. In the revised paper, we have added a new method-overview figure (Figure 2) that illustrates the shared-prefix sampling process, the within-group advantage computation for both process and answer groups, and how the collected trajectories enter the loss at the token level. We hope these changes make the method substantially easier to follow, and we are happy to add further illustrations if needed.


W2: The concepts of process supervision, information-gain rewards, and the PPO/GRPO objectives themselves are not original to this paper. This makes the novelty less significant.

We thank the reviewer for raising this point, which allows us to clarify the nature of our contribution. Our goal is indeed not to introduce yet another reward signal or objective variant, but to identify a structural failure mode that arises when existing components are combined in the usual way, and to show that resolving it requires redesigning the rollout itself. Prior methods estimate turn-level advantages from evidence collected at the trajectory level. We identify this granularity mismatch as the unifying failure mode of chain- and tree-based designs, and show it is not benign: aligning turns by index with an identical process reward loses 11.3 points (Step-GRPO), and exact post-hoc state matching starves the comparison groups (Figure 3 in the revised manuscript, GiGPO −5.1). To address this limitation, we propose, to our knowledge, the first framework that obtains a complete same-state comparison group at every turn while keeping rollout cost linear in the horizon, through dense same-prefix branching, instantaneous within-state comparison, and uniform single-path continuation. The components interact non-trivially: dense branching with greedy or IG-weighted continuation collapses to 17.2/46.0 (Table 3), because content-dependent continuation breaks the on-policy property (Proposition 1). The uniform rule is therefore a load-bearing, non-obvious part of the design, and it is what makes the formal guarantees (Proposition 1, Theorems 1–2) possible at all, which prior combinations lack. The practical value of resolving the mismatch is demonstrated empirically: SP-GRPO achieves the best average performance across seven single- and multi-hop benchmarks at both model scales, outperforming trajectory-level, step-level, and tree-based baselines. As for the works cited by the reviewer, they branch only at selected points, triggered by token entropy or a learned score, and score candidates by simulated or trajectory-level returns, often with content-dependent selection, so they do not form a complete same-state comparison group at every turn. We have added these works to the related work of the revised manuscript for a complete comparison.

Thanks for your response. I think the clarification and the new figure have largely improved the illustration of this paper. It is an empirically solid paper, and I still maintain my positive recommendation.

Reviewer_Uc8rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SP-GRPO, a critic-free RL method for turn-level credit assignment in multi-turn search agents. It argues that existing step-level methods compare actions under mismatched histories (chain-based) or depend on costly, high-variance tree returns (tree-based). At each turn, SP-GRPO samples G candidates from an identical prefix and scores search queries by information gain toward the gold answer to obtain within-turn group-relative advantages. Final answers receive GRPO-style outcome advantages. The rollout then continues along a single, uniformly chosen query, which keeps cost linear in the number of turns. The authors provide theory showing that this removes cross-state baseline contamination without introducing selection bias. On seven QA benchmarks with Qwen2.5-3B/7B, SP-GRPO achieves the best weighted average exact match over chain-based and tree-based baselines.

Critical strengths
  1. Simple and intuitive idea. Comparing candidate actions under an identical prefix and continuing a single path is easy to understand. It is also easy to implement on top of GRPO, with no critic or learned reward model.

  2. Theoretical support. The analysis shows that shared-prefix grouping removes the state-mismatch residual of cross-trajectory baselines, and that uniform continuation introduces no selection bias.

Critical weaknesses
  1. The method does not really address credit assignment, and the setting is too short to test it. The advantages of search queries are computed only from an information-gain score based on the gold answer, and the final outcome reward is never used for them. This is closer to reward shaping than to credit assignment, so the method avoids the credit assignment problem instead of solving it. In addition, there are at most four retrieval turns. The experiments therefore do not test the "long-horizon" setting the paper claims, where delayed credit and variance growing over the horizon really matter.

  2. The efficiency analysis is incomplete. Token cost is only compared with IGRPO, and time per step is only compared with GiGPO. The paper does not report the total training time needed to reach similar performance. Based on the reported numbers, SP-GRPO actually needs more time than GiGPO (about 41 hours vs. 33 hours).

Questions for authors

see weakness

3 comments

Response to Reviewer Uc8r [1/2]

We are grateful to the reviewer for recognizing the simplicity and practicality of our design that comparing candidate actions under an identical prefix while continuing a single path is intuitive and easy to implement on top of GRPO, with no critic or learned reward model. We address each concern in detail below.

W1: The method does not really address credit assignment. Query advantages come only from a gold-answer-based information-gain score, so this is closer to reward shaping than to credit assignment, and the outcome reward is never used for them. Moreover, with at most four retrieval turns, the setting is too short to test the claimed long-horizon regime where delayed credit and horizon-growing variance matter.

We thank the reviewer for raising this fundamental question. We would like to clarify that our contribution actually directly targets at credit assignment instead of reward shaping. The credit-assignment question SP-GRPO addresses is: given several candidate actions at the same state, which one should be reinforced, and how confidently? Prior step-level methods answer this question with evidence pooled across diverged histories, which Theorem 1 shows contaminates the turn-level comparison with a state-mismatch residual Δ(s) and extra between-state variance σ_b². SP-GRPO's contribution is to make this per-turn comparison exact — same prefix, i.i.d. candidates, provably unbiased (up to O(1/G)) and with the between-state variance removed — and these estimator properties hold regardless of which per-turn reward is used. The information-gain reward is one instantiation (in the IGPO family); the framework would equally accommodate a learned process reward. Our ablations isolate precisely this estimator contribution: aligning turns by index with the same IG reward collapses performance by 11.3 points (Step-GRPO, Section 4.3.1), showing that the gain comes from same-state comparison, not from the reward alone. We also note that the terminal outcome is not discarded, since every answer candidate collected along the rollout is trained with the question-level outcome advantage (Section 3.2, Algorithm 1). Finally, the information gain is not an additive shaping term: it is the state potential itself, so the group-centered instantaneous increment estimates the local advantage rather than re-specifying the objective.

As for the long-horizon setting, we set the maximum number of retrieval turns to four in order to match the baselines. We have additionally trained all four methods with a limit of ten retrieval turns under an identical protocol. Under the extended horizon, SP-GRPO reaches 37.3% average EM against 31.5% for GiGPO, 24.2% for IGPO and 32.1% for IGRPO (estimated at the same step), demonstrating that the advantage of same-state turn-level credit grows with the horizon, while return-based credit degrades as terminal outcomes are propagated through an increasing number of turns. We have included this analysis in our revised manuscript in Sec 4.3.2.

Response to Reviewer Uc8r [2/2]

W2: The efficiency analysis is incomplete. Token cost is only compared with IGRPO, and time per step is only compared with GiGPO. The paper does not report the total training time needed to reach similar performance. Based on the reported numbers, SP-GRPO actually needs more time than GiGPO (about 41 hours vs. 33 hours).

We agree that the efficiency analysis was incomplete, and we thank the reviewer for requesting the missing numbers. We have completed the analysis in RQ3 of the revised manuscript. Token cost per question is 1.9×10³ for SP-GRPO against 1.1×10³ for GiGPO, 1.2×10³ for IGPO and 6.0×10³ for IGRPO, and the time per step is 1,041 s for SP-GRPO, 673 s for GiGPO, 868 s for IGPO and 1,246 s for IGRPO. The additional tokens over a chain rollout come from the G candidate queries scored at each turn, which is the fixed cost of turn-level comparison, and the per-step time remains lower than IGRPO's. Whether this overhead pays off is what the matched-budget comparison below addresses.

Table R1. Validation EM versus cumulative GPU-hours at matched checkpoints, where each cell reports GPU-hours / validation EM.

StepGiGPOOurs (SP-GRPO)IGPOIGRPO
5046.6 / 35.746.8 / 38.841.9 / 30.4143.1 / 35.3
10081.7 / 41.5101.5 / 42.184.4 / 33.9190.1 / 41.1
150116.0 / 42.4164.2 / 44.5137.9 / 35.9240.1 / 42.3
200151.4 / 43.6228.0 / 46.1187.4 / 39.8289.9 / 43.6

To further compare the results under the same budget, we also report validation EM against cumulative GPU-hours at identical checkpoints (Table R1). At identical training progress SP-GRPO attains the highest validation EM, and along the cost axis it matches GiGPO's final EM at a slightly lower budget (about 141 vs 151 GPU-hours), while IGRPO requires 290 GPU-hours and IGPO never reaches it. At any given budget SP-GRPO's validation EM is comparable to or above every baseline, and it is the only method whose accuracy keeps improving beyond that point, showing that the per-turn scoring overhead is bounded, does not inflate the cost of reaching a given accuracy, and is offset by a higher accuracy at any given budget. These comparisons are included in RQ.3 of the revised manuscript.

Thank you for your response. It was posted so close to the end of the discussion period that I did not have enough time to review it.

Reviewer_9o4ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SP-GRPO, a reinforcement learning framework for multi-turn search agents that aims to improve turn-level credit assignment. Instead of sampling multiple independent trajectories as in standard GRPO, SP-GRPO samples multiple candidate actions under the same historical prefix at each turn. The candidate search actions are evaluated using an information-gain-based process reward, and their advantages are calculated through within-group normalization. Only one randomly selected search action is continued, while the remaining branches are pruned. Final-answer candidates are collected and optimized separately using outcome-level rewards. The authors argue that this shared-prefix design avoids the context mismatch introduced by comparing actions from independent trajectories. Experiments on seven search QA benchmarks show improvements over several chain-based and tree-based RL baselines.

Critical strengths
  1. The authors identify a reasonable problem in existing turn-level credit assignment methods, where actions from different historical contexts may not be directly comparable. The proposed shared-prefix rollout provides a straightforward way to address this issue.

  2. The authors compare against several relevant baselines, including Search-R1, GiGPO, IGPO, Tree-GRPO, and IGRPO, across seven benchmarks and two model scales. The paper also provides useful ablations on reward formulations, continuation strategies, and rollout structures.

  3. By maintaining a single active trajectory while sampling multiple candidate actions at each turn, the method avoids the rapidly increasing cost of full tree expansion. The ablation comparing random, softmax, and greedy continuation is also interesting, particularly the significant performance degradation observed with greedy selection.

Critical weaknesses
  1. I would encourage the authors to discuss more closely related branching methods, such as BranPO, and clarify the differences.

  2. The proposed method uses the change in the model's probability of generating the gold answer as an information-gain reward, which is closely related to IGPO. However, this signal may not reliably reflect the actual contribution of an intermediate search action. In multi-hop reasoning, some retrieval steps may be necessary for subsequent searches without immediately increasing the probability of the final answer. Such actions could receive negative advantages despite being useful for the overall task. I think the authors should provide more analysis of whether the proposed credit assignment remains effective when the process reward is imperfect or only weakly correlated with eventual success.

  3. Although SP-GRPO achieves the best average performance, there are some significant exceptions. In particular, on Bamboogle, SP-GRPO achieves only 33.6% and 36.8% with the 3B and 7B models, respectively, compared with 65.7% and 71.8% for IGRPO. This is a substantial degradation, especially considering that the paper aims to improve credit assignment for multi-turn search agents. I would like to see more analysis of why the method performs so poorly on this benchmark and whether this is related to the limitations of immediate information-gain rewards or the single-path rollout structure.

  4. While the paper argues that single-path continuation reduces rollout complexity compared with tree-based methods, the method still requires evaluating multiple candidate queries and computing their process rewards at every turn. The reported wall-clock time per training step is also higher than that of some chain-based baselines. Therefore, a comparison under matched GPU-hour or wall-clock budgets would be useful. In addition, the theoretical analysis of variance reduction is primarily established for a linear mean-centered estimator, while the actual implementation uses softmax-transformed and normalized advantages. I think the connection between the theoretical results and the optimization objective used in practice could be explained more clearly.

Questions for authors

See the weakness part above

2 comments

Response to Reviewer 9o4Z [1/2]

We sincerely thank the reviewer for the positive assessment of our work. We are glad the reviewer finds that we identify a reasonable problem in turn-level credit assignment and that the shared-prefix rollout provides a straightforward solution. Below we respond to each concern point by point.

W1: I would encourage the authors to discuss more closely related branching methods, such as BranPO, and clarify the differences.

We thank the reviewer for pointing us to BranPO, and we agree that an explicit comparison is warranted. While branching methods are related, they still have some limitations that SP-GRPO addresses. Specifically, BranPO branches only at adaptively selected high-entropy decision points along a backbone trajectory, and its per-step advantages are still backed up from downstream discounted returns of the sibling subtrees. SP-GRPO differs in two key aspects: (i) it branches a full group of candidates at every turn under a strictly shared prefix, rather than only at selected decision points; and (ii) it scores candidates with a turn-local process reward, avoiding the compounding variance of downstream returns (cf. Figure 4(b) in the revised manuscript, where return-based credit collapses once a third retrieval turn is required). Branching-based methods were already discussed in our original manuscript, and we have revised Sec. 2 to explicitly highlight representative work such as BranPO.


W2: Provide more analysis of whether the proposed credit assignment remains effective when the process reward is imperfect or only weakly correlated with eventual success.

We appreciate this question, as it goes to the core of how SP-GRPO assigns credit. We agree that some retrieval steps may be necessary for subsequent searches without immediately improving P_ans. We therefore examine whether the immediate-only IG signal can mislead our credit assignment in two respects: whether it affects the decision to keep retrieving or to answer, and whether preferring a candidate with a higher immediate IG actually improves final success.

On the first aspect, the stop-or-continue decision is not decided by IG. Answer candidates are scored by the exact-match outcome reward, so answering prematurely and incorrectly is already penalized. In addition, since the process advantages are mean-centered within the process group, the process-group update leaves the probability of searching versus answering unchanged (proof added in Appendix A.4 of the revised manuscript). The relevant comparison is therefore among retrieval candidates issued at the same state.

On the second aspect, we ranked each executed query against the other candidates sampled at the same state by IG, partitioned all executed queries into quartiles by this within-state rank, and measured the final success rate of the trajectories containing them. Success rises from 34.7% for the bottom IG quartile to 42.1% for the top quartile (95% CI [4.8, 10.1]); an uninformative signal would yield identical success rates across quartiles. Since the executed query is randomly sampled independently of its IG, the gap is not an artifact of IG-based selection. The within-state ranking is thus aligned with success rather than inverted.

It should be noted that this analysis is only about IG-based credit assignment. The sensitivity to the type of process reward is reported in the "Process Reward Formulation" analysis of Sec. 4, where an answer-agnostic reward costs 6.4 EM points (47.4 → 41.0), showing that the mechanism tolerates a weakly informative reward but not a wrong one. The quartile analysis is newly added in Sec. 4.3.2 of the revised manuscript, complementing the process-reward formulation study already reported in the original submission.

Response to Reviewer 9o4Z [2/2]

W3: On Bamboogle, SP-GRPO trails IGRPO substantially (33.6 vs. 65.7 at 3B, 36.8 vs. 71.8 at 7B), despite the best average performance. Please analyze why the method performs so poorly on this benchmark, and clarify whether this is related to the limitations of immediate information-gain rewards or of the single-path rollout structure.

We thank the reviewer for prompting a closer examination of the Bamboogle gap. We would first note that the gap is not indicative of a systematic failure of turn-level credit on multi-hop search: on the two larger multi-hop benchmarks, HotpotQA and 2Wiki, SP-GRPO consistently outperforms IGRPO (41.8 vs. 38.9 and 42.9 vs. 40.2 at 3B; 45.2 vs. 44.6 and 45.0 vs. 43.8 at 7B), and remains competitive on MuSiQue. These datasets contain 7,405 / 12,576 / 2,417 test questions respectively, whereas Bamboogle contains only 125, so each question accounts for 0.8 EM points and the 3B gap corresponds to roughly 40 questions.

To understand the source of the gap, we conducted a failure breakdown of all incorrectly answered Bamboogle questions at 3B. The failures decompose into retrieval misses (35.2%), strict-EM formatting artifacts (18.4%), and reasoning errors with the gold evidence already in context (12.8%). Notably, the formatting and reasoning-with-evidence failures together account for 31.2% of the errors and are orthogonal to search and credit assignment: the former are answers that are correct in content but rejected by exact match, which alone would be worth about 12.5 EM points under a content-based scorer, and the latter occur with the required evidence already retrieved and would affect any method with the same backbone. A substantial part of the apparent gap therefore reflects measurement and backbone reasoning rather than our credit assignment. We have included this analysis in Appendix C.7 in our paper.


W4: Per-step cost is higher than chain-based baselines since multiple candidate queries are evaluated each turn, so a comparison under matched GPU-hour budgets is needed. The variance-reduction theory also assumes a linear mean-centered estimator while the implementation uses softmax-normalized advantages, and their connection could be clarified.

We thank the reviewer for this constructive comment. We have added a matched-budget comparison to RQ3 and a discussion of the connection between our theory and the implemented objective, which we have promoted to a formal corollary in Appendix A.2.

Matched GPU-hour budgets. We report validation EM at identical checkpoints (steps 50/100/150/200) together with the cumulative GPU-hours of each run. Each cell reports GPU-hours / EM.

Table R1. Validation EM versus cumulative GPU-hours at matched checkpoints.

StepGiGPOOurs (SP-GRPO)IGPOIGRPO
5046.6 / 35.746.8 / 38.841.9 / 30.4143.1 / 35.3
10081.7 / 41.5101.5 / 42.184.4 / 33.9190.1 / 41.1
150116.0 / 42.4164.2 / 44.5137.9 / 35.9240.1 / 42.3
200151.4 / 43.6228.0 / 46.1187.4 / 39.8289.9 / 43.6

According to the results in Table R1 (Figure 5 in RQ3 of our revised manuscript), along the cost axis SP-GRPO matches GiGPO's final EM at a slightly lower budget (approximately 141 vs. 151 GPU-hours, by linear interpolation between checkpoints), while IGRPO requires 290 GPU-hours to reach the same accuracy and IGPO never reaches it. At any given GPU-hour budget, SP-GRPO's validation EM is comparable to or above that of every baseline, and strictly above IGPO and IGRPO at every budget. We further note that the per-step overhead of SP-GRPO is dominated by the per-candidate probe forwards, which can be batched within the rollout engine, and that our token cost per question is 3.2× lower than IGRPO's, so the comparison against IGRPO is even more favorable in token terms than in GPU-hours.

Theory and practice. In SP-GRPO, all G rewards within a group are collected from the same prefix, so the state-mismatch residual satisfies Δ(s_t) ≡ 0 and the cross-state variance term σ_b² is removed. The softmax–z-score of Eq. 2 is a group-wise monotone transform of these same-prefix rewards, computed under the same shared prefix, so it reintroduces neither Δ(s_t) nor σ_b². Theorem 2 relies only on mean-centering, which holds exactly for the implemented z-score. We have promoted this connection to a formal corollary in Appendix A.2.

Comments on the paper

1 comment

General Response

We thank all reviewers for their constructive comments and suggestions. We have revised the paper to address the concerns, with all modifications highlighted in blue in the revised PDF, and we reply to each reviewer's questions separately below. The major revisions are:

  1. Added a within-state quartile analysis of IG-based credit assignment in Sec. 4.3.2, showing that executed queries with higher IG reach higher final success (42.1% vs. 34.7% between top and bottom quartiles, 95% CI [4.8, 10.1]). Together with the process-reward formulation study already present in the original manuscript (47.4% to 41.0% with an answer-agnostic reward), this indicates that the mechanism tolerates a weakly informative reward but not a wrong one. (Reviewer 9o4Z)
  2. Completed the efficiency analysis in Sec. 4.3.3 (RQ3): token cost and per-step time for all baselines, and a matched GPU-hour budget comparison (Figure 5), showing that SP-GRPO matches GiGPO's final EM at a lower budget (about 141 vs. 151 GPU-hours) and stays strictly above IGPO and IGRPO at every budget. (Reviewer Uc8r, 9o4Z, iyRu)
  3. Added a Bamboogle failure analysis in Appendix C.7, decomposing all incorrectly answered questions at 3B into retrieval misses, strict-EM formatting artifacts, and reasoning errors with the gold evidence in context. (Reviewer 9o4Z)
  4. Added a long-horizon experiment with the retrieval limit raised from four to ten turns under an identical protocol in Sec. 4.3.2, where SP-GRPO reaches 37.3% average EM against 31.5%/24.2%/32.1% for GiGPO/IGPO/IGRPO, showing that the advantage of same-state turn-level credit grows with the horizon. (Reviewer Uc8r)
  5. Added a new method-overview figure (Figure 2) that illustrates the shared-prefix sampling process, the within-group advantage computation for both process and answer groups, and how the collected trajectories enter the loss at the token level. (Reviewer iyRu)

We appreciate the reviewers' valuable feedback, which has substantially improved the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.