acceptodds
Under review as a conference paper at ICLR 2027

Can Frontier Auto Researchers Reconstruct Human Research Decisions in AI?

Abstract

Scientific research has long been driven by substantial human effort, from reviewing prior work and identifying research gaps to designing specific solutions. Recent advances in harness-based agent systems have given rise to increasingly capable auto researchers, which can autonomously carry out these tasks and, increasingly, entire end-to-end research workflows. Yet despite their rapidly growing capabilities, how closely the research decisions and outputs produced by auto researchers align with those of human experts remains underexplored. We investigate this gap through the lens of research reconstruction and study: Given comparable research context, how similar are the research decisions and outputs produced by auto researchers to those of human researchers in AI research? To answer this question, we introduce AutoResearcher-Human Research Similarity (AutoHRS), a stage-wise evaluation framework for assessing how faithfully auto researchers reconstruct human research artifacts across different stages of the scientific workflow. Using 152 ICLR oral papers as our corpus, we generate 456 research trajectories spanning different stages of the research process. Applying AutoHRS to these trajectories, we find that current auto researchers frequently fail to recover key research decisions and their resulting artifacts across multiple stages of the research process.

Primary area:
datasets and benchmarks
Keywords:
Auto Researcher, Harness System, LLM4AI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_129yOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces AutoResearcher-HumanResearch Similarity (AutoHRS), which is a framework for evaluating how well AutoResearchers can reconstruct human researchers across different steps in AI/ML-research workflows. Stated insights include that AutoResearchers tend to rely on recombining elements from relevant references provided to them when being tasked to generate research ideas, and that they favor established models/metrics rather than more problem-specific ones when being tasked to generate experiment plans.

Critical strengths
  1. Despite the lack of novelty, the qualitative analysis of the AutoResearcher outputs and its related findings and its related findings are sound and agree with previous works.
Critical weaknesses
  1. Lack of generalizability and novelty of claimed statements. The experiment is only performed with ClaudeCode / Sonnet 4.5. First of all, recent scientific research benchmarks show that this model does not serve as a strong or reliable scientific auto-researchers (some relevant evaluation works are also not mentioned in the relevant work section). There were many other, much stronger models available well before the submission deadline and the ICLR 2026 knowledge cutoff, and analysis based on those would have made many of the claimed findings stronger, more relevant, and less obsolete. Many of these statements are also not new findings: the fact that these weaker models e.g. rely on established models/metrics rather than more problem-specific ones is not a new insight. While acknowledged in the limitations, the analysis only includes ICLR papers without considering any other scientific domains, giving the findings even less credibility. Moreover, the paper suggests that a subset of the experiments were replicated with Gemini CLI and Codex, pointing to Appendix B.1 and B.2. No appendix is available in the submitted PDF, so none of the appendix pointers help understand the work's full methodology, results, and conclusions.

  2. Evaluation Design: The checklist evaluation pipeline is limited to measuring reconstruction only via simple LLM judge labels "yes", "no", "partially" compared to one human-reference. This raises the following concerns: (i) the evaluation measures whether reconstruction of one example reference by human researchers. It does not take into account how other human researchers would e.g. design an experiment plan given a research idea, i.e. here the verification is over-constrained to one specific way of how a single human researcher or research team would write the experiment plan - this is a classic failure mode of AI evaluation works measuring scientific research capabilities. Giving the same task to another human researcher/research team would likely also result in a different approach, so it's not obvious that this evaluation setup is measuring something that applied to AutoResearchers only; (ii) even if the judge labels the AutoResearchers' approach as "no", there is no measure of the direction of the "failed" reconstruction: did the AutoResearcher suggest an approach that is worse or better than the human reference? What if the AutoResearcher comes up with a better research idea or experiment plan than the human?

Questions for authors

see questions posed in the "critical weaknesses" section.

4 comments
AuthorsOpenReview ↗

Author Response (1/2)

We thank you for the evaluation of our paper and the feedback. We would like to clarify that, in the original submission, the appendix was included in the supplementary material PDF. We have now incorporated the appendix into the main PDF to make it easier to access through the hyperlinks in the main text. And we would like to address your other concerns and questions as follows:

W1.1: The experiment is only performed with ClaudeCode / Sonnet 4.5. First of all, recent scientific research benchmarks show that this model does not serve as a strong or reliable scientific auto-researchers (some relevant evaluation works are also not mentioned in the relevant work section). There were many other, much stronger models available well before the submission deadline and the ICLR 2026 knowledge cutoff, and analysis based on those would have made many of the claimed findings stronger, more relevant, and less obsolete Many of these statements are also not new findings: the fact that these weaker models e.g. rely on established models/metrics rather than more problem-specific ones is not a new insight.

We thank the reviewer for raising this concern. We would like to clarify that our evaluation is not limited to Claude Sonnet 4.5. In addition to our main experiments, we also evaluated Gemini CLI (Gemini 3.5 Flash) and Codex (GPT-5.2), with the corresponding results already reported in Appendix B.2 in the supplementary material pdf file. Although ICLR allows authors to include the appendix in the supplementary materials, we acknowledge that its location was not clearly indicated in the main paper, which may have caused confusion.

Furthermore, we took your concern about the availability of more powerful models with knowledge cutoffs before the ICLR 2026 submission deadline seriously. We additionally evaluated Claude Opus 4.6 and Codex GPT-5.4 on the same test set to further examine the generalizability of our results. Here are the full results across different harness-model configs on the test set:

Harness-Model ConfigIdea Generation (best candidate)Experiment Design (best candidate)Paper Synthesis (best candidate)
Claude Code (Opus 4.6)0.330.480.74
Codex (GPT-5.4)0.320.500.72
Claude Code (Sonnet 4.5)0.250.410.68
Gemini CLI (Gemini 3.5 Flash)0.260.440.71
Codex (GPT-5.2)0.280.500.69

From these results, although we observe some improvements with Opus 4.6 and GPT-5.4 as the backbone models, the overall performance patterns remain consistent across different models and harnesses. In particular, idea generation consistently achieves the lowest scores, while paper synthesis achieves the highest scores. These findings further support the generalizability of our main conclusions across the evaluated models. To control the computational costs of running the AutoResearcher pipelines, we selected Claude Sonnet 4.5 for the main experiments. We believe these results help address the generalization concern.

W1.2: While acknowledged in the limitations, the analysis only includes ICLR papers without considering any other scientific domains, giving the findings even less credibility.

Thank you for pointing this out. We respectfully disagree that limiting our evaluation to the AI field undermines the credibility of our findings, as our work explicitly focuses on AI research, as already stated in both the title and the main text. Extending our evaluation to other scientific domains is beyond the scope and resources of this study, particularly because our human annotators have expertise primarily in AI research. We have clarified the scope in Section Abstract (line 011) and Introduction (line 056-057) of the revised manuscript.

W1.3: Moreover, the paper suggests that a subset of the experiments were replicated with Gemini CLI and Codex, pointing to Appendix B.1 and B.2. No appendix is available in the submitted PDF, so none of the appendix pointers help understand the work's full methodology, results, and conclusions.

Thank you for pointing this out. As mentioned earlier, in the original submission, the appendix was included in the supplementary material PDF. Appendices B.1 and B.2 provide additional experimental details and cross-harness evaluation results for Gemini CLI and Codex. Although ICLR allows appendices to be submitted separately as supplementary material, we have now incorporated the appendix into the main PDF of the revised paper for easier access and a better reading experience.

AuthorsOpenReview ↗

Author Response (2/2)

W2.1: The evaluation measures whether reconstruction of one example reference by human researchers. It does not take into account how other human researchers would e.g. design an experiment plan given a research idea, i.e. here the verification is over-constrained to one specific way of how a single human researcher or research team would write the experiment plan - this is a classic failure mode of AI evaluation works measuring scientific research capabilities

We thank the reviewer for raising this concern. The reviewer is concerned that using a single human reference for each research idea may not capture the diversity of experimental plans that different human researchers could develop.

However, while this concern is conceptually valid, addressing it through multiple human references is challenging in practice. Unlike tasks with standardized inputs and multiple acceptable answers, scientific research is inherently open-ended, and collecting multiple independently developed human research papers addressing exactly the same idea under comparable research contexts is generally infeasible. Consequently, prior benchmarks evaluating AI research capabilities have also adopted published human research papers as references. For example, PaperBench (ICML 2025) evaluates AI agents' ability to replicate published AI research, while FIRE-Bench (ICML 2026) evaluates whether AI agents can rediscover scientific findings from human research papers. These established benchmarks demonstrate the practicality of using published research as a reference for evaluating AI research agents.

In our work, we further mitigate this concern through two design choices.

First, we select high-quality and diverse human research references. Our evaluation covers 152 ICLR oral papers from different research teams, capturing a broad range of research decisions and experimental designs across the AI field. We specifically select ICLR oral papers, rather than poster papers, papers from less selective venues, or rejected papers, because oral papers undergo rigorous peer review and represent highly regarded research contributions. We therefore consider their experimental designs to be strong reference examples reflecting the decisions of experienced human researchers. Importantly, we use these papers as references for reconstruction, rather than assuming that their experimental designs are the only scientifically valid solutions.

Second, our weighting scheme emphasizes relatively well-constrained aspects of experimental design. Specifically, Baselines and Comparisons (0.30) and Evaluation Protocol (0.25) together account for 55% of the Experiment Design score. These dimensions are more strongly constrained by the research problem and established experimental practices than other aspects of experimental design. For example, a new method should generally be compared against relevant existing approaches and evaluated using appropriate protocols. Although alternative valid choices may exist, we expect less variation in these dimensions across well-designed experimental plans addressing the same research idea.

W2.2: even if the judge labels the AutoResearchers' approach as "no", there is no measure of the direction of the "failed" reconstruction: did the AutoResearcher suggest an approach that is worse or better than the human reference? What if the AutoResearcher comes up with a better research idea or experiment plan than the human?

We thank the reviewer for raising this concern. However, we respectfully disagree with this criticism, as our work does not aim to evaluate whether AutoResearchers produce better or worse research than human researchers. Instead, our goal is to investigate whether AutoResearchers can reconstruct research decisions similar to those made by human researchers under comparable research contexts. Therefore, a "No" label indicates a divergence from the human reference, rather than suggesting that the AutoResearcher's approach is necessarily inferior. We acknowledge that an alternative approach may be equally valid or even better than the human reference, but evaluating such differences in research quality is beyond the scope of AutoHRS.

Thank you again for your time and attention in helping us to improve our paper. We welcome any further feedback or discussion that you are able to provide.

AuthorsOpenReview ↗

Kind Request to Review Our Rebuttal

Dear Reviewer,

As the author–reviewer discussion comes to a close, we would be grateful if you could share any remaining concerns or points you'd like to discuss. We understand this is a busy time, and we truly value the time and effort you've spent to reviewing our work and engaging with our responses.

Thank you!

AuthorsOpenReview ↗

Kind Request to Review Our Rebuttal. Thank you!

Dear Reviewer,

As the author–reviewer discussion will ends tomorrow, we would be grateful if you could share any remaining concerns or points you'd like to discuss. We truly value the time and effort you've spent to reviewing our work and engaging with our responses.

Thank you!

Reviewer_vKCpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces AutoHRS (AutoResearcher–Human Research Similarity), a framework for evaluating whether automated research systems can reconstruct the decisions and outputs made by human researchers at different stages of the scientific process. The authors divide the research process into three tasks: generating research ideas from the literature, designing experiments based on those ideas, and synthesizing a paper from experimental results. The evaluation corpus consists of 152 ICLR 2026 oral papers, yielding 456 stage-specific research trajectories. Using manually reviewed checklists, the authors compare agent outputs with human references from the papers. They find that current agents achieve relatively low reconstruction scores for idea generation and experimental design, and higher scores for paper synthesis. The paper also analyzes differences between agent outputs and human research, and examines how research-area granularity and research type affect the results.

Critical strengths
  1. The research question is important and addresses a gap in automated research evaluation. Evaluating only whether a final task is completed may hide differences in research decisions, such as problem formulation, baseline selection, and evaluation design. This paper’s attempt to evaluate such decisions stage by stage is insightful.

  2. The stage-wise evaluation is more diagnostic than a single end-to-end metric. The paper distinguishes among research idea generation, experimental design, and paper synthesis, and further breaks each stage into specific dimensions. The results show clear differences in agent performance across stages.

  3. The authors attempt to validate the agreement between automatic evaluation and human judgment. The authors manually evaluate 20 examples and report the correlation between AutoHRS and human judgments. This is more convincing than reporting unvalidated LLM-judge scores alone.

  4. The paper analyzes patterns in how agent outputs differ from human work. For example, agents tend to recombine elements from the provided reference papers, while their experimental designs often rely on common models and standard metrics. These observations can help motivate future research questions.

Critical weaknesses
  1. The relationship between reconstruction similarity, research quality, and originality is unclear.
    AutoHRS uses research decisions in published papers as references and checks whether an agent reproduces the corresponding elements. This can measure how similar an output is to a specific human research trajectory, but it does not directly establish whether an agent’s idea is valuable, its experiment is rigorous, or its research is better than that in the reference paper. A reasonable and innovative alternative may receive a low score because it does not reproduce the original paper’s specific choices. Conversely, a high reconstruction score does not necessarily indicate high research quality. I encourage the authors to frame the metric more explicitly as “research-trajectory reconstruction similarity” and avoid interpreting it as general research capability or research quality.

  2. Controls against leakage of information from target papers need to be described more transparently.
    The authors state that they selected the corpus based on the underlying models’ knowledge cutoffs and the papers’ initial public release dates to reduce the risk that agents could directly reproduce the target papers. However, the paper should also report the range of initial public release dates in the final corpus, the versions of the models used and their corresponding knowledge cutoffs, and the model configurations used in the cross-harness experiments. The paper mentions partial replications using Gemini CLI and Codex, but the main text does not specify which underlying models were used in those experiments.

    In addition, the paper does not clearly state whether network access, browsers, or external search tools were enabled when running the agents. If internet access was allowed, agents may have retrieved the target papers directly. If it was disabled, the corresponding execution restrictions should also be reported. I encourage the authors to clearly specify the network permissions, available tools, and input materials for each experiment so readers can assess whether the leakage controls were sufficient.

  3. Coverage of model backbones is limited, leaving the generalizability of the conclusions unclear.
    The main experiments use only Claude Sonnet 4.5. The behavior of automated research agents may depend on the capabilities and training characteristics of their underlying models, so it is unclear whether the reconstruction performance and bias patterns observed in this paper apply to other model families. Although the authors mention partial replications using Gemini CLI and Codex, the main text does not provide enough detail about the model configurations and results to assess consistency across models. I encourage the authors to conduct a more systematic comparison across different model families to determine whether the main findings of AutoHRS generalize.

Questions for authors

Please see the Critical Weaknesses section.

5 comments

We thank you for the evaluation of our paper and the feedback. We would like to clarify that, in the original submission, the appendix was included in the supplementary material PDF. We have now incorporated the appendix into the main PDF to make it easier to access through the hyperlinks in the main text. And we would like to address your other concerns and questions as follows:

W1: The relationship between reconstruction similarity, research quality, and originality is unclear.

We thank the reviewer for pointing this out. As we mentioned in the abstract and Lines 67–72, our work focuses on how closely the research decisions of AutoResearchers align with those of human researchers given the same research context, rather than on how good these decisions are. But we still appreciate the reviewer's suggestion and have further emphasized this distinction in the Conclusion Section of the revised paper.

We have also added the suggested framing of the metric as “research-trajectory reconstruction similarity” in Section 3.1 (metrics, line 242-244) of the revised paper. Thank you for this suggestion!

W2.1: However, the paper should also report the range of initial public release dates in the final corpus, the versions of the models used and their corresponding knowledge cutoffs, and the model configurations used in the cross-harness experiments.

We thank the reviewer for raising this concern. We chose September 2025 as the cutoff date for the initial public release of the ICLR papers. In addition to collecting these papers using an automated script, human annotators manually screened them and filtered out unsuitable papers to ensure data quality.

Although we had already included the harness configurations in the appendix (provided in the supplementary material) of the original submission, we have now conducted additional experiments with GPT-5.4 and Claude Opus 4.6. Accordingly, we have updated the harness configurations in Appendix B.2 (Lines 1003–1012) of the revised PDF. The configurations used in our experiments are summarized below.

HarnessUnderlying ModelKnowledge Cutoff
Claude CodeClaude Sonnet 4.5Jan 2025
Claude CodeClaude Opus 4.6May 2025
Gemini CLIGemini 3.5 FlashJan 2025
CodexGPT-5.2Aug 2025
CodexGPT-5.4Aug 2025

W2.2: In addition, the paper does not clearly state whether network access, browsers, or external search tools were enabled when running the agents.

Thank you for raising this concern. We had already described the network access restrictions in Appendix B.1 of the supplementary material in our original submission, but we acknowledge that the description was not sufficiently detailed. Regarding network access, we partially disabled internet access during the experiments. We disabled web search and applied strict date filters to APIs used for literature reviews. To further ensure that the target research trajectory was not leaked, after each round of experiments, we asked the model to self-check the logs and also conducted manual checks to identify potential knowledge leakage. We have already added further details regarding network access and knowledge leakage prevention in Appendix B.1 (Lines 985–990) of the revised paper.

W3: Coverage of model backbones is limited, leaving the generalizability of the conclusions unclear.

We thank the reviewer for raising this concern. As we mentioned at the beginning of our response, the appendix is included in the supplementary material PDF file, in which we compared the performance of Claude (Sonnet 4.5), Gemini CLI (Gemini 3.5 Flash), and Codex (GPT-5.2) on a randomly sampled test set of 20 papers, with sampling proportional to the distribution of paper types (Empirical, Method/Algorithm, and Evaluation). To further address your concern about generalizability, we additionally evaluated Claude Opus 4.6 and Codex GPT-5.4 on the same test set to further examine the generalizability of our results. Here are the full results across different configs on the test set:

Harness-Model ConfigIdea Generation (best candidate)Experiment Design (best candidate)Paper Synthesis (best candidate)
Claude Code (Opus 4.6)0.330.480.74
Codex (GPT-5.4)0.320.500.72
Claude Code (Sonnet 4.5)0.250.410.68
Gemini CLI (Gemini 3.5 Flash)0.260.440.71
Codex (GPT-5.2)0.280.500.69

Since all the models showed consistent performance patterns on the test set, and considering the substantial computational cost of running AutoResearchers, we selected Claude Sonnet 4.5 for the full evaluation across the entire dataset. We believe these results help address the generalization concern.

Thank you again for your time and attention in helping us to improve our paper. We welcome any further feedback or discussion that you are able to provide.

AuthorsOpenReview ↗

Kind Request to Review Our Rebuttal

Dear Reviewer,

As the author–reviewer discussion comes to a close, we would be grateful if you could share any remaining concerns or points you'd like to discuss. We understand this is a busy time, and we truly value the time and effort you've spent to reviewing our work and engaging with our responses.

Thank you!

Thank you for the authors’ responses. The additional details on leakage controls and results across different backbones have addressed my concerns on these points. I will raise my Soundness score from 2 to 3.

Regarding the first weakness, I remain uncertain about the practical significance of research-trajectory reconstruction similarity. The paper’s main findings is that agents have low similarity to human research trajectories, but I do not see what practical guidance this result provides. Research trajectories may be highly variable: different human researchers can also follow very different paths, and such differences do not necessarily indicate differences in ability or quality. It therefore remains unclear what we can infer when an agent’s trajectory differs from a particular human trajectory. For this reason, I will not raise my Overall Rating.

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

The paper investigates a really interesting question: how closely do the research decisions and outputs produced by auto researchers align with those of human experts? The paper operationalizes the study of this question by creating an evaluation framework for assessing how faithfully auto researchers reconstruct human research artifacts across three stages of research:

  1. idea generation GIVEN a set of reference papers and a research direction

  2. experiment design GIVEN a given research idea

  3. paper synthesis GIVEN a research context and experimental findings

For each stage, the auto researcher is provided with the ground truth input and is asked to produce output in a specifically structured format. Output is automatically judged according to its similarity to the output of human researchers in AI using automatically generated checklists.

The corpus used for the evaluation is 152 ICLR oral papers. The paper describes generally how this corpus (i.e. the inputs and outputs for each stage for each paper) is constructed and validated but the details, which are critical to get right, are relegated to the appendix. The framework itself is validated wrt human judgments and shows strong alignment.

Results show that current harness-based auto researcher agent systems are extremely limited.

Critical strengths

The research question addressed by the paper is really interesting and, of course, quite timely. It is also an extremely vague question, which is both conceptually and practically hard to operationalize. The paper does a good job along both of these dimensions --- first, by defining and then focusing on three specific subtasks; then by putting together a multitude of carefully decisions for (a) how to create the ground-truth artifacts and (b) evaluate similarity to human researchers. Validation of the checklist-based scores with respect to human judgments was also key.

Critical weaknesses

The weighting scheme could have been described a bit more in the main paper. It currently reads to me as though determination of the weights requires a new set of human judgments for each auto researcher, which would not be feasible in practice.

Questions for authors

No specific questions.

1 comment

We sincerely thank you for the thoughtful evaluation of our paper and the constructive feedback. We would like to clarify that the weights are initially determined through a grid search maximizing correlation with human judgments on one AutoResearcher configuration. We further verified that the same weights achieve the highest correlation across four additional configurations, supporting their generalizability to different AutoResearchers. Thank you for this helpful suggestion. We have clarified this procedure in Section 3.1 (Lines 254–257) of the revised paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Rebuttal and Revisions

Dear AC and Reviewers,

We sincerely thank all reviewers for their time and effort in reviewing our paper. We also appreciate the constructive feedback and engagement during the discussion, as well as the recognition of the motivation and significance of our work. In particular, Reviewer ASrdr described our work as "really interesting and, of course, quite timely," while Reviewer vKCp acknowledged that "The research question is important and addresses a gap in automated research evaluation."

Throughout the rebuttal period, we have carefully considered the reviewers' feedback, provided detailed responses to their concerns, and revised our manuscript accordingly. All revisions are highlighted in blue in the updated manuscript for ease of reference. We would like to summarize the main updates and clarifications below.

  • Appendix Accessibility (vKCp, 129y). Reviewers vKCp and 129y raised concerns about the missing appendix in the main PDF. Although the appendix was originally included in the supplementary materials, which is permitted under the ICLR submission guidelines, we recognize that its absence from the main PDF may have affected the accessibility and clarity of our results. To address this, we have incorporated the complete appendix into the latest version of the main PDF.

  • Generalizability Across Models and Harnesses (vKCp, 129y). Both reviewers questioned the generalizability of our findings. Some of these concerns may have arisen because the results involving additional harnesses, which were briefly mentioned in the main text, were detailed in the previously separate appendix. With the appendix now integrated into the main PDF, these results are more accessible. Furthermore, we have conducted additional experiments with stronger models, including Claude Opus 4.6 and GPT-5.4, to further examine the generalizability of our findings. We are pleased that Reviewer vKCp acknowledged that our additional experiments and clarifications addressed their concerns regarding generalizability.

  • Clarifications of Experimental Settings and Evaluation Design (vKCp, ASrdr). We have addressed the reviewers' questions regarding several methodological details, including the specific experimental settings and the rationale behind our evaluation criteria and weighting scheme. We are pleased that Reviewer vKCp acknowledged our additional clarifications. These clarifications have been incorporated into the revised main text and are highlighted in blue for ease of reference.

  • Practical Significance of Research-Trajectory Reconstruction (vKCp). In response to Reviewer vKCp's concern, we have further clarified the motivation and significance of evaluating AutoResearchers' ability to reconstruct human research decisions under comparable research contexts. We emphasize that this provides a meaningful and complementary evaluation dimension rather than a measure of overall scientific quality. We have also incorporated this clarification into the Conclusion section of our revised manuscript. We hope Reviewer vKCp will find our further explanation helpful in addressing their remaining concern.

As the discussion period comes to a close soon, we hope that our revisions and responses have adequately addressed the reviewers' concerns. We would be grateful if the reviewers and Area Chair could take these updates and clarifications into consideration in their final assessment. Thank you again for your time and constructive feedback.

Best Regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.