acceptodds
Under review as a conference paper at ICLR 2027

Understanding and Reducing Misalignment in Persuasion Optimization

Abstract

Frontier models are now trained for complex multi-agent tasks, from running a company to engaging in negotiations. These tasks necessitate training models to be persuasive. But what does this optimization pressure mean for model alignment? In this work, we study how reinforcement learning (RL) training for persuasion affects models’ use of misaligned strategies such as evidence fabrication and belief coercion, and what techniques can reduce such misalignment. We find that optimizing sender models solely for receiver belief shifts causes the average number of distinct misaligned strategies used per game to increase by more than 25%, and fabrications per game increase to more than 4 times relative to pre-RL models. We compare various mitigation approaches, including stronger strategy guidance, clean demonstrations, penalizing misaligned strategy usage, and improving the model’s ability to detect fabrications. We find that navigating the persuasiveness-misalignment trade-off in this setting is challenging, with the best trade-offs coming from combining guidance on aligned persuasion strategies with reward penalties on misalignment. Overall, our work provides a testbed for assessing misalignment stemming from RL-based persuasion training and shows concrete paths for reducing misalignment. Importantly, it highlights that without direct intervention, persuasion-involved RL environments are likely to naturally lead to large increases in the use of misaligned strategies.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Persuasion, Reward Hacking, Reinforcement 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_gqHaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the negative impact of RL, especially in persuasion scenarios. Experiments reveal that RL trains LLMs to produce misaligned fabrications as training progresses. The authors then suggest several remedies to fix this misalignment in persuasion optimization, but also show that there is a trade-off.

Critical strengths
  1. RL has become a common post-training method for LLM and this paper reveals RL’s negative impact for social use of LLMs. This seems to be a timely issue for researchers.
  2. The analysis is interesting and shows the RL’s negative impact.
    • RL makes LLMs more deceptive and more likely to make fabricated claims. (Figure 2)
    • Whether receiver is stubborn or not, RL increases the model’s usage of misaligned strategies. (Figure 1)
    • Figure 4 and 5 show that as RL progresses models generate more obvious fabrications (model don’t know whether their claims are fabricated or not) and this becomes more strong when they get deep into the sender’s role.
Critical weaknesses
  1. Settings are quite far from the real world setting.
    • Sender can see the receiver’s inner thoughts, which cannot happen in reality.
    • Receiver’s prompt contains a phrase “the evidence they do reveal is always true.” Here “they” refers to senders, so receivers cannot doubt the fabrication. This is unrealistic in a real-world setting.
  2. The same model serves as both the penalty auditor during training and the evaluation auditor.
    • This can train model to evade this specific auditor (judge) rather tan genuinely becoming more aligned.
Questions for authors
  1. Can you make the sender's and receiver's prompt without unrealistic phrase? and test the tendency is still same? (weakness 1)
  2. Can you set two LLM different (weakness 2) and see what happens?
3 comments

We thank the reviewer's positive, insightful comments and for recognizing our findings in training more aligned LLMs for persuasion and acknowledging the project is timely. We address the feedback and questions below, and have updated the paper accordingly.

Responses to Weaknesses

W1. Realism of the persuasion setting

  • W1.1 Access to receiver’s inner thoughts: The sender never sees a hidden chain-of-thought from the receiver. As shown in Appendix C.4.2 at line 1092, the receiver is explicitly instructed "do NOT reason out loud" and outputs only its visible turn response consisting of "belief", "argument", and "action" blocks. Between rounds, the sender only sees the receiver's stated justification and explanation.
  • W1.2 Receiver's prompt phrase "the evidence they do reveal is always true": In standard Bayesian persuasion from Kamenica & Gentzkow, 2011, the sender is assumed to be truthful and can only strategically choose which evidence to reveal rather than inventing false facts. Under this assumption, the receiver can perform a well-defined, interpretable belief update via Bayes' rule, which is why Cheng et al., 2026 included the phrase "however, the evidence they do reveal is always true" in their receiver prompt. Although our problem formulation does not assume the LLM sender will stay within strict Bayesian constraints, we kept this Bayesian updating template so the receiver's numerical belief shift remains grounded and interpretable rather than an arbitrary score. At the same time, we agree it is important to test whether misalignment is merely an artifact of telling the receiver that revealed evidence is always true. As shown in our new ablation in Response to Q1 below, removing this phrase so the juror can doubt unverified claims still leads to substantial fabrication and misalignment under outcome-only RL.

W2. Using the same model for training penalties and evaluation auditing

We agree that sharing Qwen3.5-35B-A3B between training-time penalties and test-time evaluation could in principle allow the sender to evade that specific judge. Here, we additionally evaluate with GPT-5.4-mini as an independent auditor. As shown in Response to Q2 below, the alignment gains hold under the independent judge.

Responses to Questions

Q1. Removing the unrealistic phrase from the prompt and testing whether the tendency persists

We removed "however, the evidence they do reveal is always true" from the receiver prompt (allowing the juror to doubt uncorroborated claims) and re-trained Qwen3-4B under outcome-only RL. Outcome-only RL still drives substantial misalignment: misaligned strategies per game increase from 2.8 to 3.4, and fabricated claims per game increase from 1.24 to 6.13, consistent with the one that contains the phrase. This indicates that the phrase "however, the evidence they do reveal is always true" does not play a critical role in our experiment.

Q2. Setting the training auditor and evaluation auditor to two different LLMs

We ran another round of strategy audit using GPT-5.4-mini as the auditor for both checkpoints before and after RL. Compared with checkpoints trained without penalty, training with penalty reduced misaligned strategies per game from 3.35 to 1.24 on 4B and from 3.14 to 1.15 on 8B, closely matching Qwen3.5-35B-A3B. See table below:

SenderBelief ShiftMisaligned per Game (Qwen Judge)Misaligned per Game (GPT Judge)
4B, Full strategy prompt, no penalty, step 00.412.723.44
8B, Full strategy prompt, no penalty, step 00.341.582.30
4B, Full strategy prompt, no penalty, step 1000.622.943.35
8B, Full strategy prompt, no penalty, step 1000.842.673.14
4B, Full strategy prompt + GPT-5.4-mini Distill + penalty, step 2000.550.681.24
8B, Full strategy prompt + GPT-5.4-mini Distill + penalty, step 2000.500.491.15
Reviewer_gqHaOpenReview ↗

Response to authors' Rebuttal

Thank you for the clear and focused response. I appreciate the clarification on the receiver's outputs, the new ablation removing "the evidence they do reveal is always true," and the evaluation with an independent GPT-5.4-mini judge.

Although these responses have addressed my main concerns to a reasonable extent, I believe my current score remains appropriate when considering the overall results presented in the paper. Therefore, I will maintain my current score.

Reviewer_RyF6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how persuasion-oriented RL changes language models’ use of misaligned strategies. Using Old Bailey as the main training environment, the authors track fabrication and coercion, examine their association with receiver belief shifts, and compare interventions based on prompting, clean demonstrations, reward penalties, and fabrication-detection training. The results suggest that outcome-only optimization increases misaligned behavior, while combining strategy guidance, demonstrations, and penalties improves the measured persuasion–alignment trade-off.

Critical strengths
  • The evidence-grounded setting makes unsupported claims easier to identify than in open-ended persuasion tasks. Tracking behavior throughout RL training, together with transfer evaluations on House Showing and Nutrition Advice, provides a useful empirical picture.

  • The intervention comparisons reveal a meaningful gap between capability and behavior: improving fabrication detection does not reliably reduce misaligned persuasion, and clean demonstrations alone offer limited protection under continued outcome-only RL.

Critical weaknesses
  • The central message needs sharper qualification. The experiments demonstrate increased misuse under a specific objective that rewards an LLM receiver’s reported belief shift. It remains unclear how much reflects general persuasion risks versus exploitation of this simulated receiver and reward. Similarly, counting distinct misaligned strategies treats behaviors with different frequencies and severity equally, limiting the interpretation of the reported “25% increase.”

  • Section 3.4 does not convincingly distinguish deception from hallucination. Direct-ask and in-role probes show that fabrication recognition depends on context, but neither establishes what the sender recognized when originally generating the claim. The findings support changes in verification behavior more directly than the proposed explanation of why fabrication occurs.

  • The same auditor model supplies misalignment penalties and evaluation labels, leaving potential evaluator-specific optimization unresolved. Appendix D.1 provides helpful human verification, but only on 30 trajectories and through review of existing auditor verdicts. Stronger independent validation is needed to establish that the interventions reduce the underlying behaviors rather than primarily their detection by this auditor.

Questions for authors
  1. Which findings go beyond the expected effects of an underspecified reward? Would frequency- or severity-sensitive measures support the same conclusions as distinct strategy counts?

  2. Can the deception–hallucination distinction be supported more directly, or should Section 3.4 be reframed around context-dependent fabrication recognition?

  3. Do the mitigation gains persist under an independent auditor and blind human annotation, particularly for penalty-trained models?

3 comments

We thank the reviewer's insightful comments and for recognizing our findings in training more aligned LLMs for persuasion. We address the feedback and questions below, and have updated the paper accordingly.

Responses to Weaknesses

W1. Separating general persuasion risks and exploitation in a specific environment.

In real-world persuasion settings such as courtroom trials, property sales, and health consultations, persuaders are also rewarded for shifting a listener's belief. That same outcome-only incentive is precisely what drives distortive and coercive behaviors in practice. Because our tasks are real historical cases and realistic multi-turn scenarios with explicit evidence sets, optimizing for receiver belief shift captures this core incentive structure rather than an unrelated proxy. Consistent with this, the misaligned behaviors we observe transfer to House Showing and Nutrition Advice, and hold when replacing the receiver with stronger models (DeepSeek-V4-Flash), showing that they are not just artifacts of a single environment or receiver model. That said, we agree on keeping this scope clear, which is why in Appendix A we explicitly note that our receivers are LLM simulators rather than human participants, so our quantitative estimates should be interpreted within an LLM-simulated setting even though the underlying incentive structure reflects real-world persuasion.

Regarding the distinct-strategy count, we agree that the trajectory-level binary metric serves as a conservative lower bound, and this is why we pair it with instance-level fabricated claim counts (which are frequency- and severity-sensitive) in Figure 2 and category decompositions in Figure 7.

W2. Distinguishing deception from hallucination in Section 3.4

The two prompt templates in Section 3.4 are designed to bridge neutral verification (when the model does not have persuasive pressure) and generation-time persuasive pressure:

  1. Direct Ask evaluates the model as an outside, third-party verifier given only the case background and evidence record.
  2. In-Role Prompting replays the exact sender-receiver conversation up to the turn immediately before the fabricated claim was produced, placing the model back into the exact persuasive state and optimization pressure under which it generated the claim.

Our results in Figure 5 show that when placed back under this in-role persuasive pressure, the model fails to recognize fabrications far more often than when acting as a neutral third-party verifier, and this degradation becomes more severe as RL training progresses. According to our definition of deception and hallucination, this shows that under persuasive pressure, models will become more hallucinated and less able to recognize that they are fabricating. That said, we agree that behavioral prompting probes remain an indirect proxy for internal reasoning states, and we have added this limitation in Appendix A.

W3. Shared auditor model for training penalties and evaluation

We agree that using Qwen3.5-35B-A3B for both reward penalties and evaluation may lead to potential judge-specific evasion. We conducted new evaluations with an independent frontier LLM auditor (GPT-5.4-mini) , which confirms that the alignment gains persist. See Response to Q3 for full results.

Responses to Questions

Q1. Findings beyond an underspecified reward and frequency-/severity-sensitive measures

  • Beyond showing that outcome-only RL increases misalignment, our study reveals the following non-obvious findings:
    1. Asymmetric suppression: Under penalty RL, coercive strategies are mostly eliminated, whereas distortive strategies remain harder to suppress, with the same pattern holding out-of-domain.
    2. Reasoning-action dissociation: Even with explicit forbidden-strategy prompts and clean SFT initialization, RL models state in their "thinking" block that they must avoid forbidden tactics while still fabricating evidence in their "argument" block, as shown in Section 4.1 and Appendix E.3.
    3. Verification-generation dissociation: Raising in-role fabrication detection accuracy from 0.68 to 0.95 via an auxiliary loss does not effectively mitigate fabrication generation, whereas reward penalties reduce misalignment without improving detection accuracy.
  • Frequency- and severity-sensitive measures show even stronger effects than distinct strategy counts. Counting fabricated claim instances per game reveals a 4.4× to 7.8× surge on Old Bailey, rising from 1.4 to 6.1 on 4B and from 0.8 to 6.2 on 8B. The increase in fabricated claims also generalizes across different domains that are not used for training.

Q2. Supporting the deception-hallucination analysis and clarifying Section 3.4

Beyond the In-Role vs. Direct-Ask gap, explained in the response to W2, our qualitative examples provide more direct evidence that both mechanisms occur during persuasion:

  1. Deliberate fabrication: In our newly added Appendix E.4, we provide a qualitative example where the model explicitly mentioned that it would fabricate new evidence in order to increase the persuasiveness. This directly indicates that the model is trying to be deceptive, since it’s aware that it is fabricating.
  2. Unrecognized fabrication under pressure: In other trajectories such as Appendix E.3, the model explicitly commits in its "thinking" trace to avoid all false claims and stick to verifiable facts, yet immediately invents an unsupported date and archive record in its "argument" block, consistent with the elevated error rates under In-Role probing.

Q3. Do mitigation gains persist under an independent auditor?

Yes. We ran another round of strategy audit using GPT-5.4-mini as the auditor for both checkpoints before and after RL. Compared with checkpoints trained without penalty, training with penalty can reduce the number of misaligned strategies per game on 4B from 3.35 to 1.24 and on 8B from 3.14 to 1.15, closely matching the auditing results from Qwen3.5-35B-A3B. See table below:

SenderBelief ShiftMisaligned per Game (Qwen Judge)Misaligned per Game (GPT Judge)
4B, Full strategy prompt, no penalty, step 00.412.723.44
8B, Full strategy prompt, no penalty, step 00.341.582.30
4B, Full strategy prompt, no penalty, step 1000.622.943.35
8B, Full strategy prompt, no penalty, step 1000.842.673.14
4B, Full strategy prompt + GPT-5.4-mini Distill + penalty, step 2000.550.681.24
8B, Full strategy prompt + GPT-5.4-mini Distill + penalty, step 2000.500.491.15

Regarding blind human annotation, while full blind human labeling across all 11 strategies and multi-turn transcripts is resource-intensive within the discussion window, our new blind human check on extracted fabrications achieves a 0.93 agreement rate with our strategy auditor (see Appendix D.1), showing that our auditing pipeline can reliably identify and extract fabricated claims.

Dear reviewer, we hope that our response to your comments has addressed any concerns or questions. If any aspects of our rebuttal remain unclear or require further clarification, please feel free to let us know. We would be happy to provide additional information or elaborate as needed. Thanks!

Reviewer_1hcTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors study the effect of optimizing LLMs (Qwen3) on persuasion with RL (GRPO) and its subsequent misalignment in terms of Deception and Hallucination. The authors report the results on the Old Bailey trial dataset, with different Qwen3 models as Prosecutor and Juror. Qwen is again used to classify the trajectories into a taxonomy of persuasion strategies extending prior work. The key finding is that outcome only RL, makes the prosecutor deceive the Juror with fabricated claims (purposefully deceiving as well as hallucination) and misaligned strategies. The authors also compare mitigations by tuning the prompt, SFT on clean distillations, and a penalty. Better prompts and GPT-5.4-mini distillation plus the penalty gives the best trade-off, close to GPT-5.4-mini and GPT-5.5.

Critical strengths

It's a clever idea and a good environment to test such findings The writing was clear and easy to understand

Critical weaknesses
  1. My primary concerns are against the toy nature of the task and validity of the experiments. While the authors present strategy guidance already showing good performance improvement, I would argue that prompting the prosecutor to be faithful and not fabricate details should be the baseline. Further, the juror, is a very weak and brittle judge. Note that the prosecutor’s goal is to advocate for conviction, so they may reveal only part of the evidence; however, the evidence they do reveal is always true. This makes it very natural to reward fabrication, and really question the environment design. Reward hacking needs to be studied in far more calibrated envs, in real life any discriminator, persuadee, or Juror needs to have sufficient means to not be hacked so easily. This is why simple guidance prompts mitigate the issue as well. Fig. 6 also shows this. For Qwen3-4B, the best penalized checkpoint ends at about the same belief shift (~0.53) as the untrained base-prompt model, against 0.84 after outcome-only RL, so almost all of RL's gain here is the misaligned part.

2.⁠ ⁠All datasets consist of very few samples, which make the reward hacking even more inevitable.

Questions for authors

1.⁠ ⁠If we fix the juror prompt "the evidence they do reveal is always true," and the juror gets the evidence list to check or a different belief update system closer to real life, does outcome-only RL still raise fabrications by a similar amount? 2.⁠ ⁠When a human checks the step-0 claims extracted as fabricated, how many are actually supported by the evidence set? This decides whether the Fig. 5 trend reflects hallucination or the extractor's precision.

5 comments

We thank the reviewer's insightful comments and for recognizing the clarity of our paper and the contributions of our evaluation framework. We address the feedback and questions below, and have updated the paper accordingly.

Responses to Weaknesses

W1. Task design and experimental validity

  • W1.1 Prompting the prosecutor to be faithful as the baseline: We agree that instructing the sender to be honest is an essential baseline, and our paper already includes this at two levels:
    1. Base Prompt already instructs the model to be honest: As shown in Appendix C.4.1, even our Base Prompt specifies negative prosecutor utility for "wrongful convictions or dismissed cases due to improper methods" (line 937) and explicitly instructs the sender: "Be honest about what you present" (line 947). Outcome-only RL still overrides this instruction.
    2. Full strategy guidance as the Section 4 baseline: In our Full Guidance Prompt, we explicitly forbid all 11 misaligned strategies, and use this prompt as the direct baseline for comparing SFT initializations and reward design choices. As shown in Section 4.1 and Appendix E.3, full strategy guidance lowers Step-0 misalignment, but is still brittle under outcome-only RL: models explicitly note in their "thinking" block that fabrication is forbidden while still inventing evidence in their "argument" block.
  • W1.2 Receiver prompt phrase ("always true") and judge design: In standard Bayesian persuasion from Kamenica & Gentzkow, 2011, the sender is assumed to be truthful and can only strategically choose which evidence to reveal rather than inventing false facts. Under this assumption, the receiver can perform a well-defined, interpretable belief update via Bayes' rule, which is why Cheng et al., 2026 included the phrase "however, the evidence they do reveal is always true" in their receiver prompt. Although our problem formulation does not assume the LLM sender will stay within strict Bayesian constraints, we kept this Bayesian updating template so the receiver's numerical belief shift remains interpretable. At the same time, we agree it is important to test whether misalignment is merely an artifact of telling the receiver that revealed evidence is always true. As shown in our new ablation in Response to Q1 below, removing this phrase so the juror can doubt unverified claims still leads to substantial fabrication and misalignment under outcome-only RL.
  • W1.3 Receiver calibration and interpreting Figure 6:
    1. Calibration across stronger receivers and non-courtroom domains: Appendix D.4 (Figure 11) replaces the receiver with DeepSeek-V4-Flash and observes the same emergence of misalignment; likewise, in non-courtroom environments like House Showing and Nutrition Advice, the model still exhibits increases in fabricated claims and misaligned strategies.
    2. Interpreting Figure 6: Comparing the penalized RL checkpoint directly against the untrained Base Prompt overlooks that the untrained Base Prompt is already misaligned: at Step 0, Qwen3-4B with Base Prompt uses ~2.7 misaligned strategies per game, meaning part of its belief shift comes from unpenalized misaligned strategies. When we instead fix the sender prompt with full strategy guidance and cleaned initialization, the receiver’s belief shift starts at ~0.41 (4B) and ~0.38 (8B) at step 0. RL with penalties then increases belief shift by +31.0% on Qwen3-4B and +48.5% on Qwen3-8B, while reducing misaligned strategies by 72.1% (4B) and 60.5% (8B). Thus, aligned RL achieves substantial persuasive gains over aligned baselines without sacrificing safety.

W2. Dataset scale

Most of the existing persuasion benchmarks are open-ended, where fabrication cannot be grounded, and relatively small-scale (SOTOPIA: 90 examples; PersuSafety: 572 examples; makemesay: 189 examples). Old Bailey provides 1,221 real historical trials, making it, to our best knowledge, the largest multi-turn persuasion environment with closed, verifiable evidence sets. For out-of-domain evaluation, House Showing and Nutrition Advice serve as a compact supplemental generalization check, and the total size of these two environments (339 games) is still comparable to other existing persuasion datasets.

Responses to Questions

Q1. Removing "the evidence they do reveal is always true":

We tested whether outcome-only RL still raises fabrications when the truthfulness assumption is removed while keeping the prior-to-posterior belief update format intact:

  1. RL without the "always true" phrase: We removed "however, the evidence they do reveal is always true" from the receiver prompt and re-trained Qwen3-4B under outcome-only RL. Outcome-only RL still drives substantial misalignment: misaligned strategies per game increase from 2.8 to 3.4, and fabricated claims per game increase from 1.24 to 6.13.
  2. On giving the receiver the evidence list to check: We intentionally do not give the receiver direct access to the ground-truth evidence list because information asymmetry is a core feature of real-world persuasion. In practice, such as a seller pitching a property to a buyer or a doctor advising a patient, the receiver rarely has enough context or a ground-truth record to verify every claim on the spot. Moreover, if the receiver already held the full evidence list, the sender would have no private information left to reveal. Fabrication and manipulation pose the greatest real-world risk in exactly these asymmetric-information settings, which is why our work focuses on aligning the sender rather than assuming the receiver can independently verify every statement.

Q2. Human precision check on Step-0 claims extracted as fabricated

Thank you for suggesting this. To validate the reliability of the LLM-based strategy auditor in identifying and extracting fabricated evidence, we randomly sampled 30 evidence items flagged as fabricated by Qwen3.5-35B-A3B across two initial checkpoints: Qwen3-4B Base (step 0) and Qwen3-8B Base (step 0). Human annotators independently assessed the correctness of these classifications, achieving a 0.93 agreement rate with the auditor. This indicates that our pipeline is reliable in detecting and extracting fabricated evidence.

Dear reviewer, we hope that our response to your comments has addressed any concerns or questions. If any aspects of our rebuttal remain unclear or require further clarification, please feel free to let us know. We would be happy to provide additional information or elaborate as needed. Thanks!

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback and for recognizing the timeliness of our study, the clarity of our presentation, and the value of our evidence-grounded persuasion testbed. To address the shared questions across reviews, we added several new experiments and updated our manuscript accordingly:

  1. Removing "the evidence they do reveal is always true" from the receiver prompt (1hcT, gqHa): Re-training Qwen3-4B with this phrase removed still yields substantial misalignment under outcome-only RL: misaligned strategies per game rise from 2.80 to 3.40, and fabricated claims surge from 1.24 to 6.13, confirming that reward hacking is not an artifact of this specific phrase. We added these results in Appendix D.5.
  2. Auditing penalty-trained models with an independent LLM judge (RyF6, gqHa): We re-evaluated our checkpoints with GPT-5.4-mini instead of Qwen3.5-35B-A3B, Compared with checkpoints trained without penalty, training with penalty can reduce the number of misaligned strategies per game on 4B from 3.35 to 1.24 and on 8B from 3.14 to 1.15, closely matching the auditing results from Qwen3.5-35B-A3B. We added these results in Appendix D.1
  3. Human verification on Step-0 extracted fabrications (1hcT): We sampled 30 evidence items flagged as fabricated by Qwen3.5-35B-A3B across two initial checkpoints: Qwen3-4B Base (step 0) and Qwen3-8B Base (step 0). Human annotators independently assessed the correctness of these classifications, achieving a 0.93 agreement rate with the auditor. We added these results in Appendix D.1.
  4. Direct evidence of deliberate fabrication (RyF6): We provide a new qualitative example in Appendix E.4 to show the case when the LLMs directly mention they need to fabricate evidence in their CoTs.
  5. Added limitations on our approach in understanding the source of fabrication at the end of Appendix A (RyF6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.