acceptodds
Under review as a conference paper at ICLR 2027

ReactionBench: A Human-Anchored Benchmark for Modeling Reactions to Video

Abstract

Understanding what happens in a video is not sufficient for socially responsive multimodal systems: such systems also aim to relate what a person observes to how that person responds. Existing video benchmarks largely evaluate events, temporal reasoning, or people depicted in the video, while affective and viewer-response datasets often focus on emotion labels, aggregate responses, or prediction from the stimulus alone. This leaves an underexplored question: can current multimodal models ground an observed viewer reaction in the stimulus context associated with it? We introduce ReactionBench, a human-anchored benchmark for reaction-specific video understanding. Built from a candidate pool of 31,834 synchronized stimulus–reaction clips spanning 433 hours and eight domains, ReactionBench evaluates three complementary abilities: ranking the intensity of a viewer’s reactions within a video, matching a reaction to its temporally aligned stimulus, and explaining the concrete stimulus cues linked to that reaction. Human annotation provides 14,894 ratings over 3,257 moments, and the frozen evaluation comprises 250 intensity profiles, 112 matching items, and 200 rationale items with 447 blind-written human references. Across seven multimodal models, reaction-specific grounding remains challenging: the best model reaches 50.9% accuracy on four-way stimulus–reaction matching, while GPT-5 shows no detectable loss under reaction substitution. Performance is also sensitive to evaluation protocol.

Primary area:
datasets and benchmarks
Keywords:
Multimodal Large Language Models, Video Understanding, Human Reaction Understanding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ArGsOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the ability of general-purpose multimodal large language models to understand observed viewer reactions and relate them to video stimuli. The authors introduce ReactionBench with three tasks: Intensity Profiling (IP), which ranks reaction intensity; Stimulus–Reaction Matching (SRM); and Grounded Rationale Generation (GRG), which generates stimulus-grounded explanations of reactions. They collect 31,834 synchronized stimulus–reaction candidate clips, obtain 14,894 human intensity ratings, and retain 447 human-written rationale references after quality review. The final evaluation sets comprise 250 IP videos, 112 SRM items, and 200 GRG items. Experiments with seven multimodal large language models show that current models have limited performance overall across the three tasks and that the results are highly sensitive to input format and evaluation protocol.

Critical strengths
  1. The paper focuses on observed viewer reactions and explores models' ability to understand them through three tasks covering intensity ranking, stimulus matching, and grounded explanation. This perspective offers some novelty.
  2. The authors evaluate seven multimodal large language models and examine how input formats and evaluation settings affect the results, including the Joint/Independent comparison for IP, the Stimulus-only and Shuffled-reaction controls for SRM, and experiments with native video inputs. They also report human annotation agreement and sensitivity analyses for some evaluation metrics. These experiments provide a relatively thorough analysis of the evaluation protocol.
Critical weaknesses
  1. The validity of the benchmark has not been sufficiently established. Through the complementary IP, SRM, and GRG tasks, the paper aims to evaluate models' understanding of viewer reactions and their relationship to stimuli. However, there is insufficient evidence that the task scores effectively measure this core capability. IP mainly measures viewer reaction intensity without involving stimulus content. SRM directly evaluates stimulus–reaction matching, but has not been validated against an independent human baseline to establish whether the correct answer can be reliably identified from the reaction and candidate stimuli alone. Moreover, GPT-5's performance differences across the correct-reaction, no-reaction, and substituted-reaction conditions are not statistically significant. GRG primarily evaluates agreement between model explanations and human references, which does not adequately establish whether the model has identified the stimulus that actually elicited the reaction. Although the three tasks address different aspects of reaction understanding, it remains unclear to what extent the benchmark's overall results reflect an understanding of stimulus–reaction relationships.

  2. There are substantial concerns about the reliability of the annotations and evaluation references. First, agreement among human intensity ratings for IP is low (Krippendorff's α = 0.357), and 58% of the test videos contain only two moments to rank. Although the authors apply rater-bias correction and leave-one-rater-out analysis, the stability of the final human-consensus rankings has not been sufficiently established. Second, SRM determines the correct stimulus directly through temporal alignment. Viewer reactions may be delayed, however, and a reaction may have multiple plausible triggers, so temporal alignment does not necessarily identify the actual stimulus–reaction relationship. Finally, GRG references are written after annotators watch continuous stimulus videos with audio and may explain reactions using dialogue or sound, whereas models receive only silent frames sampled at 1 FPS. This information asymmetry may make some references difficult to infer reliably from the model inputs and thereby complicate interpretation of the evaluation results.

Questions for authors
  1. Given the low agreement among human intensity ratings for IP (Krippendorff's α = 0.357) and the fact that 58% of test videos contain only two moments to rank, could the authors report human agreement on relative intensity rankings separately for videos with 2, 3, 4, and 5 moments? Could they also assess how sensitive the final consensus rankings are to the choice of annotators?

  2. The correct answer in SRM is determined by temporal alignment, but viewer reactions may be delayed and different stimuli may elicit similar reactions. Is there further evidence that the correct answer can be reliably identified from the provided reaction and candidate stimuli alone? For example, are independent human matching results available under the same input conditions as those used for the models?

  3. GRG references are written after viewing continuous stimulus videos with audio, whereas models see only silent frames sampled at 1 FPS. How do the authors verify that these references can be inferred from the information actually available to the models? In addition, the human references in Figure 3(b) differ substantially in specificity. How do the authors ensure that reference-based evaluation fairly assesses different plausible explanations produced by the models?

1 comment

Thank you for the detailed review and for the strengths you note: the focus on observed viewer reactions through three tasks, and the analysis of the evaluation protocol (joint vs. independent IP, the SRM stimulus-only and shuffled-reaction controls, native-video inputs, human agreement and metric sensitivity). The new experiments extend these analyses (General Response G1–G4).

W1 / Q2 / W2 (SRM label) — validity, human baseline, temporal alignment.

The SRM label. Reaction and stimulus are synchronized recordings, and the correct option is the stimulus segment that was playing when the reaction occurred (Appendix B.5). Because the label comes from the recording rather than from annotator judgment, we respectfully disagree that a human baseline is needed to establish it; a human score would instead describe how hard the task is for people. Construction adds three checks: the reaction view spans 2 s before to 1 s after the annotated range, following the sub-second onset of spontaneous facial reactions (Dimberg et al., 1998; 2000); option start times differ by at least 6 s (median 11.2 s) and options are description-dissimilar; and a human reviewer replayed every item to confirm the candidates are distinguishable against the reaction. For the camera-ready version, we would be happy to add such a human baseline on the models' inputs as a reference point for readers.

The controls. We see the GPT-5 result you point out as the benchmark's central finding rather than a weakness, and it now replicates across six models (General Response G3): the real reaction adds +2.9 points over stimulus-only and +3.5 over a shuffled reaction, and no single model's contrast is significant. Current models reach most of their SRM score without the observed reaction.

IP and GRG. IP is reaction-only by design: it tests whether a model can read the reaction signal, and SRM and GRG test the link to the stimulus. For GRG, the cue sub-score measures whether the model names the stimulus cue that the references identify; we will report it as the trigger-identification measure.

W2 / Q1 — α = .357, two-moment videos, consensus stability. The consensus rankings used as references are stable: removing any one rater leaves them at Spearman .866 [.847, .884] to the full consensus over 1,073 removals, with the full ordering unchanged in 78% (the mean scores the 117 all-tied remainders as 0, per the paper's ρ₀ convention; .972 without them), flat across 2 / 3 / 4 / 5+ moments (.862 / .867 / .876 / .891). α measures agreement on the absolute 0–3 scale, while IP is scored on within-video order. Stratified rater-vs-others ranking agreement, as requested: .826 [.731, .907] for 2 moments, .490 for 3, .495 for 4 and .312 for 5+ (video-cluster CIs; as in the paper, pairs with an all-tied score vector are dropped, which removes 59 of the 145 two-moment videos). Both tables go into the appendix.

W2 / Q3 — GRG audio and reference specificity. A text-only screen flags 13 of 200 references as potentially audio-dependent; without them every model's composite moves by .01–.08 and the larger gaps are unchanged (General Response G4). We will report both sets and mark the flagged items; a human check of the sampled frames remains the direct test of inferability. On specificity, the judge scores cue against at least one of the 1–5 accepted references, so a model matching any reference at that reference's level of specificity receives full credit, and the human anchor (84 pairs, Spearman .593 with the human mean) shows that the judge tracks human judgments under this rule.

We hope these responses clarify the design of the benchmark and address your concerns, and we would be grateful if you would consider them in your final assessment. We are happy to answer any further questions.

Reviewer_9fBaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Aiming to explore whether current multimodal foundation model could ground viewer reactions in the multimodal context, this work proposes REACTIONBENCH, ranking the model's ability in 3 different dimensions. Experiments across 7 different multimodal foundation models shows that current models still struggle with the task.

Critical strengths
  1. This paper designs benchmark targeting a very interesting problem, whether current multimodal foundation model could ground viewer reactions in the multimodal context.

  2. The authors honestly state their proposed method's limitation, such as they report disagreement between human annotators, and do a thorough human check for the judge.

  3. The diagram presentation is very clear and aesthetically pleasing.

Critical weaknesses
  1. The evaluated models are limited, a lot of more advanced models are not evaluated, such as Qwen 3.8 series, Claude series, GPT6.

  2. The test set is small and narrow, which leads to susceptible for overfitting and the unstable of the performance. For example, SRM has 112 items so the Wilson interval is about ±0.09, and models from 0.411 to 0.509 are not really separable.

  3. The benchmark does not show that it measures what it claims about whether model could ground the observed reaction. We would appreciate that if the authors could explain more about the casual relationship between the measured capability and the benchmark score.

Questions for authors
  1. Which GPT-5 SRM number should be the benchmark score, 0.509 or 0.429? If the same model alias gives a 8 point drop, it's very hard for people to compare performance on the benchmark.

  2. I'm wondering if it is possible to rerun the joint IP for all models with the corrected timestamps?

  3. Table 1 lists text, audio and visual for REACTIONBENCH, but all main results use frames only. Is audio part of the benchmark or not?

1 comment

Thank you for the constructive review and for the strengths you note: the problem of grounding an observed viewer reaction in its multimodal context, the explicit reporting of limitations such as annotator disagreement, and the human check of the GRG judge. The new analyses follow the same practice (General Response G1–G4).

W1 — model coverage. The evaluated set spans two proprietary API models (GPT-5, Gemini 3.7 Flash) and open-weight dense and mixture-of-experts models from 9B to 397B. As requested, we added the newest Qwen generation, Qwen3.8-27B (SRM .411; corrected joint IP .230, the highest among open models). The paper's findings hold across every family: no model exceeds .51 on SRM, none of the six models with controls gains reliably from the real reaction, and independent IP scoring leads joint scoring for every model. The harness and frozen sets will be released, so Claude or GPT-6 can be added directly.

W2 — 112 items, Wilson ±.09. Wide intervals are a property of test-set size and are common to human-annotated social-intelligence benchmarks, where every item needs a synchronized stimulus–reaction pair and human review; many benchmarks in this area do not report intervals. We report them, with paired significance tests, so that readers do not over-read small gaps. The set is a frozen evaluation set used for no training or model selection, so its size affects precision, not overfitting. With Holm-corrected exact McNemar tests, 4 of 21 pairs separate; adjacent models in the .41–.51 band do not, and the revision reports them as tied, with Wilson intervals and significance groups in Table 2. Re-running the four open-weight models on identical inputs moves accuracy by at most ±.045. The SRM finding holds at the level the set supports: every model is far from solving the task (best .509 vs. .25 chance) and gains little from the real reaction (General Response G3). We will expand SRM from 205 further match-ready candidates.

W3 — does the benchmark measure reaction grounding? SRM is designed to make this link measurable. Reaction and stimulus are synchronized recordings, and the correct option is the stimulus segment that was playing when the reaction occurred, so the label comes from the recording rather than from annotator judgment. Every item was replayed by a human reviewer to confirm that the candidates are distinguishable against the reaction. The reaction-input controls then measure how much of the score depends on the reaction: across six models, the real reaction adds only +2.9 points over stimulus-only and +3.5 over a shuffled reaction (General Response G3). Current models reach most of their SRM score without the observed reaction; we see this as the benchmark's central finding, and the abstract will state it as such. For the camera-ready version, we would be happy to add a human baseline on the same inputs as a reference point for readers.

Q1 — GPT-5 SRM: .509 or .429? We apologize for the confusion. The benchmark score is .509 (Table 2). The two numbers are not two measurements of the same setting: .429 comes from the separate controls run, whose stimulus option clips contained one to seven more frames than the benchmark input on 43 of the 112 items. The gap sits on those items (21 vs. 14 correct, p = .04); on the 69 identical items the runs score .522 and .493 (p = .77), in line with the ±.045 retest movement of the open-weight models on identical inputs. The revision will state the source of every number and release input hashes with every prediction file.

Q2 — corrected timestamps for all models. Done for all accessible models, plus a retraining of the joint adapter (General Response G1, G2). Qwen3.6-27B is no longer served and Gemini 3.7 Flash was not accessible to us during the discussion period; their rows will be marked as original-prompt only.

Q3 — audio. Audio is part of the dataset release, not of the evaluated condition: all main tables use frames only, because only one evaluated model accepts native video with audio. Table 1 will separate "released modalities" from "evaluated input", and General Response G4 reports the audio screening of GRG.

We hope these responses address your concerns, and we would be grateful if you would consider them in your final assessment. We are happy to clarify any remaining questions.

Reviewer_Hd9dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ReactionBench, a benchmark for evaluating how multimodal models interpret observed viewer reactions and relate them to video stimuli. The dataset combines synchronized stimulus–reaction videos with human intensity ratings and human-written explanations, and includes training and validation data for supervised adaptation alongside fixed evaluation sets. It covers three tasks: within-video reaction intensity ranking, stimulus–reaction matching, and grounded rationale generation. The authors evaluate seven multimodal models and investigate annotation agreement, scoring protocols, reaction removal and substitution, rationale judging, and supervised adaptation.

Critical strengths
  1. Evaluating an observed viewer reaction, rather than predicting a typical response from the stimulus alone, provides a useful perspective on multimodal understanding.

  2. The experiments are thorough. The authors compare different scoring setups and examine parsing failures, helping explain whether higher scores reflect better task performance or simply changes in how answers are produced and scored.

  3. The 14,894 intensity ratings and 447 accepted human-written explanations provide a valuable empirical foundation. The annotation and scoring procedures are documented in detail.

Critical weaknesses
  1. In joint training, frames are subsampled to meet the input limit, but the prompt still states that frame N corresponds to second N. This mismatch affects 838/856 joint-training rows and may cause the model to associate the rated moments with the wrong frames. Could the authors rerun a representative adaptation experiment with corrected timestamps and check whether the output collapse persists? This would help distinguish the effect of the joint training format from that of temporal misalignment.

  2. The evaluation population is narrow. IP and SRM cover only seven creators from five channels. Combined with selection toward strong reactions, this limits the benchmark’s coverage of ordinary viewing behavior and subtle responses.

I would be willing to raise my score if the authors can satisfactorily address the concerns above.

Questions for authors

See above

1 comment

Thank you for the careful review and for the strengths you note. We agree these are central to the submission: the benchmark evaluates an observed viewer reaction rather than a typical response predicted from the stimulus; the evaluation separates task performance from how answers are produced and scored (scoring protocols, parse-failure analysis); and the 14,894 intensity ratings and 447 accepted human-written explanations, with documented annotation and scoring procedures, are its empirical base and will be released. You also pinpointed a real bug in the timestamp statement, and we ran the experiment you asked for.

W1 — timestamp mismatch in joint training (838/856 rows).

Retraining (please refer to G2 in the General Response). We rebuilt the 856 joint SFT rows with the true per-frame timestamps and retrained the joint-only adapter with the paper's recipe (Qwen3.5-9B, LoRA r=16, 5 epochs, seed 0), then evaluated it with the corrected prompt on the 250 IP videos.

  • The output collapse persists. The retrained adapter produces an all-tied profile on 121/250 videos, against 122/250 for the paper's adapter (joint ρ₀ .096 vs. .100; difference −.005 [−.088, +.079]).
  • The timestamp text is not its sole cause. Correcting it does not remove the collapse, which points to the joint training format rather than the temporal misalignment, the distinction you asked about (single seed, one recipe).
  • Table 3 changes. The corrected prompt also lifts the local base (.024 → .055), so the joint-only adapter's gain over base is no longer significant (+.040 [−.075, +.152]). Table 3 will be re-reported under the corrected prompt.

Evaluation (please refer to G1 in the General Response).

  • The five misaligned evaluation videos change 3 of the 25 model–video values, ±.008 each on the 250-video mean.
  • The explicit per-frame timestamp list improves the two Qwen models in same-day controls (+.069 [+.010, +.131] and +.097 [+.017, +.176]); GPT-5 and Kimi-K2.6 show no detectable change.
  • Independent scoring keeps the higher point estimate for all five paired models; the gap is no longer significant for those two Qwen models.
  • Table 2 and §5 will report the corrected-prompt numbers as primary, and Appendix D.5 will document the bug, the affected rows and all reruns.

W2 — narrow evaluation population (7 creators, 5 channels; selection toward strong reactions).

The focus on strong reactions is a deliberate design choice that the paper states: the evaluated sets "represent selected creator reactions, not ordinary viewing behavior" (Appendix A.2), and the limitations describe the creator population (Appendix F). The reason is measurability.

  • Strong reactions give reliable labels. They are clearly visible in the reaction video and tend to be shared across viewers (a punchline, a goal, a jump scare), so the trigger can be located in the stimulus and the intensity can be rated.
  • Subtle responses do not. They vary from person to person and often leave no visible trace, so their labels would reflect annotator disagreement as much as the reaction itself.
  • Strong reactions are already hard for current models. The best SRM accuracy is .509 against .25 chance, and models make little measurable use of the real reaction (General Response G3).

In the revision we will:

  • move this scope statement and its rationale from the appendix into §3;
  • summarise the per-channel sensitivity table of Appendix D.4 in the main text (matched human pairwise accuracy .567–.769 per channel);
  • release the full candidate pool, including 205 further match-ready candidates, so that the set can be extended to new creators.

We hope the retraining answers W1, and would be grateful if you would reconsider your score in light of G1 and G2 in the General Response. We are glad to answer any follow-up.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (1/2)

We thank all three reviewers and the AC. Before the details, we briefly restate what the paper contributes, since the new experiments strengthen each part.

Contributions.

  1. ReactionBench, to our knowledge the first large-scale benchmark for a question that existing video benchmarks leave open: can a multimodal model ground an observed viewer reaction in the stimulus that elicited it? It is built from 31,834 synchronized stimulus–reaction clips (433 hours, eight domains) and anchored in human data: 14,894 intensity ratings over 3,257 moments and 447 accepted human-written explanations. Three tasks cover intensity ranking (IP, 250 videos), stimulus–reaction matching (SRM, 112 items) and grounded explanation (GRG, 200 items).
  2. A seven-model evaluation showing substantial headroom: the best SRM accuracy is 50.9% against 25% chance, and models gain little from the real reaction. The rebuttal extends the reaction controls from GPT-5 to six models (G3) and confirms this as the paper's central finding.
  3. Diagnostic analyses of how the evaluation protocol shapes measured performance: joint vs. independent scoring, reaction removal and substitution, judge validation against humans, and supervised adaptation.

We thank the reviewers for recognising these points: the focus on observed reactions (Hd9d, 9fBa, ArGs), the thorough analysis of the evaluation protocol (Hd9d, ArGs), the human data and documented procedures (Hd9d), and the transparent reporting of limitations (9fBa).

We have addressed every concern raised. The General Response covers the concerns raised by more than one reviewer, with new experiments: the timestamp fix, evaluated for all accessible models and with a retrained adapter (G1, G2; Hd9d, 9fBa), reaction controls for six models (G3; 9fBa, ArGs), and an audio screening of GRG (G4; 9fBa, ArGs). Points raised by a single reviewer (population scope for Hd9d, score stability and the two GPT-5 figures for 9fBa, annotation reliability for ArGs) are answered in full, with new analyses, in the individual replies. We would be grateful if the reviewers and the AC could take these results into account in the final assessment. All runs use the frozen item sets and the paper's scoring code.

G1. Corrected-timestamp joint IP (250 videos)

The joint prompt said "frame N corresponds to second N", which is wrong after subsampling for the 5 evaluation videos longer than 64 s. The corrected prompt lists the true timestamp of every frame. We re-ran every model still accessible (Qwen3.6-27B is no longer served; Gemini 3.7 Flash was not accessible to us during the discussion period) and added Qwen3.8-27B.

Modelpaper joint ρ₀corrected joint ρ₀indep ρ₀indep − corrected [95% CI]
GPT-5.026.020.177+.157 [+.024, +.292]
Kimi-K2.6.086.100.233+.133 [+.017, +.244]
Qwen3.5-9B.029.088.255+.168 [+.031, +.303]
Qwen3.6-35B-A3B.043.113.210+.097 [−.040, +.231]
Qwen3.5-397B-A17B.065.189.267+.065 [−.053, +.183]†
Qwen3.8-27B (new)—.230——

Paired video bootstrap, 10,000 resamples. † 226 paired videos (24 unparsed in the independent run).

The five misaligned videos account for at most ±.008 of any model's change. The larger Qwen gains come from the explicit per-frame timestamp list: same-day reruns with the original prompt show no detectable drift from the paper's numbers ( +.001 and +.028, CIs include 0), and the corrected prompt improves on them by +.069 [+.010, +.131] and +.097 [+.017, +.176]. Independent scoring stays ahead of joint scoring for all five paired models, significantly for three. Table 2 and §5 will report the corrected prompt as primary.

AuthorsOpenReview ↗

General Response (2/2)

G2. Retraining the joint adapter on corrected rows (Qwen3.5-9B, Table 3)

838 of the 856 joint SFT rows carried the same statement. We rebuilt them with the true timestamps and retrained with the paper's recipe (LoRA r=16, 5 epochs, seed 0), then evaluated with the corrected prompt:

Armjoint ρ₀ [95% CI]all-tied profiles
Base.055 [−.053, +.166]1/250
Paper adapter (old rows).100 [+.026, +.176]122/250
Retrained adapter (corrected rows).096 [+.020, +.170]121/250

Retrained − paper adapter: −.005 [−.088, +.079]. Fixing the timestamps does not remove the output collapse, so the timestamp text is not its sole cause in this single-seed replication. Because the corrected prompt raises the local base (.024 → .055), the joint-only adapter's gain over base becomes +.04 with a CI that includes 0. Table 3 will be re-reported under the corrected prompt, and Appendix D.5 will document the bug and all reruns.

G3. Reaction-input controls for six models (SRM, 109 shared items)

The paper ran the stimulus-only and shuffled-reaction controls for GPT-5; we ran them for five more models. Δ = normal − control accuracy (source-group cluster bootstrap).

ModelNormalΔ vs. stimulus-onlyΔ vs. shuffled reaction
GPT-5.422+.046+.037
Qwen3.5-9B*.312−.064−.009
Qwen3.6-35B-A3B.431+.055+.055
Qwen3.8-27B.404+.101+.055
Qwen3.5-397B-A17B.459+.037+.046
Kimi-K2.6.367+.000+.028
Pooled, 6 models+.029 [−.018, +.077]+.035 [−.004, +.076]

* About 20 normal and shuffled responses are truncated before the answer letter (scored incorrect); without this model the pooled Δ is +.048 [−.003, +.098] and +.044 [+.004, +.087]. No single model's contrast reaches p < .05 (exact McNemar).

This extends the paper's GPT-5 finding to six models: current models reach most of their SRM score without the observed reaction. Stimulus-only accuracy mixes stimulus cues with answer-position priors (always answering D scores .330), which we will report next to chance. The SRM label itself is the recorded, temporally aligned stimulus segment.

G4. GRG references and audio

A text-only screen of the 200 GRG references flags 13 items as potentially audio-dependent (12 dialogue or a punchline). Without them every model's composite moves by .01–.08, and the larger gaps between models are unchanged. We will report both sets and mark the flagged items.

We hope these results, together with the individual replies, address the reviewers' concerns. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.