acceptodds
Under review as a conference paper at ICLR 2027

A Simulation-Grounded VLM Framework for Agentic Wildfire Monitoring and Reporting

Abstract

Effective wildfire monitoring requires relating visual evidence to physical fire dynamics, yet real videos with synchronized physical annotations are scarce and high-fidelity 3D simulation is costly. We present a simulation-grounded vision-language model (VLM) framework that automatically converts 2D wildfire simulations into labeled video episodes. A fixed Blender mapping produces low-detail 3D proxies aligned with simulator terrain, fuel layout, fire activity, and wind cues; controllable video generation supplies richer appearance. The proxies are intermediate representations rather than finely rendered final scenes. Generated videos and simulator labels form reusable multimodal memory for a training-free multi-agent VLM system that retrieves reference episodes, reconciles visual and memory-based predictions, and produces structured wildfire reports. On held-out generated episodes, video memory achieves 51.5% exact four-tag accuracy, compared with 22.6% for direct VLM querying and 16-17% for text-only memory; the complete system achieves 77.3% accuracy on six simulator-derived report fields. Component ablations, cross-generator tests, and three real-UAV evaluations assess retrieval, reporting, generator changes, and observable monitoring tasks. The framework connects automatic simulation-to-proxy conversion with memory-based VLM reasoning under scarce real-world physical annotations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
wildfire monitoring, 2D simulation-controlled video generation, memory system

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

This paper proposes a simulation-grounded VLM framework that turns 2D wildfire simulations into labeled synthetic videos, then uses video retrieval and a multi-agent system to estimate fire-related physical conditions and generate structured wildfire reports.

Critical strengths
  1. A practical and well-motivated problem.

The paper addresses the lack of real wildfire videos with synchronized physical annotations. Using 2D simulations, lightweight 3D proxies, and video generation to create labeled visual data is a reasonable and potentially scalable solution. This pipeline could be useful for other scientific applications where physically annotated visual data is scarce.

  1. Solid experimental improvements on synthetic benchmarks.

The proposed video-memory approach improves four-tag classification accuracy from 22.6% (direct VLM querying) to 51.5%, while the full multi-agent system reaches 77.3% accuracy on six downstream reporting tasks. The paper also includes component ablations, cross-generator tests, and some real-UAV evaluations, making the empirical results reasonably convincing within the synthetic-data setting.

Critical weaknesses

The paper claims to bridge physical wildfire simulation and real-world visual monitoring. However, its main evaluation only demonstrates the recovery of simulation-derived labels from synthetically generated videos, while the real-world experiments address different, simpler perception tasks. It remains unclear whether the proposed framework learns or retrieves physically meaningful cues that transfer to real wildfire observations.

Questions for authors

The paper’s main motivation is to bridge the gap between physical simulations and real-world wildfire monitoring. However, the current experiments mainly evaluate alignment with synthetic proxy renderings or performance on different real-world perception tasks.

Could the authors clarify whether they have quantitatively evaluated the sim-to-real gap between generated and real wildfire videos, for example, using established video-distribution metrics such as FVD?

More importantly, is there any evidence that the physical cues learned from synthetic videos generalize to real UAV footage? The current real-world experiments focus on smoke/fire detection and fire-area ranking rather than the proposed environmental-driver estimation tasks. Could the authors provide a same-task sim-to-real evaluation or other quantitative evidence supporting this transfer?

3 comments

We thank the reviewer for your recognition of the importance of such a problem, the contribution of our work and the questions for improving. We sincerely hope reviewers will consider the mentioned difficulties of acquiring real data with target settings and consider the experiments we did in order to partially compensate for this problem.

Evaluation on real UAV footage:

This actually is exactly the motivation and reason why we want to propose this framework, especially the first data collection stage and demonstrate it is usable to aid wildfire predictions using our memory system and agentic framework. A direct numerical comparison against arbitrary online UAV videos would also be difficult to interpret because the terrain, viewpoint, sensor characteristics, fire intensity, and temporal evolution would not be matched. Constructing a paired real reference would require recovering these latent conditions from real footage, which is itself a separate and ill-posed inverse problem. To the best of our knowledge, there’s wildfire image dataset and very few UAV wildfire video ones publicly available, e.g. FireMan-UAV-RGBT or Boreal Forest Fire dataset and FLAME dataset are the ones that’s widely used for wildfire evaluations. However, they do not provide real clips paired with matching simulator states and the complete set of environmental and fire-behavior labels required by our evaluation but are designed for smoke and fire detection with many experiment actually performed as controlled burn instead of real wildfire capturing. To be specific, FireMan-UAV-RGBT depicts controlled burns of piled vegetation rather than wildfires. Boreal Forest Fire dataset targets smoke detection and the fire front is often barely visible. FLAME primarily supports fire detection and segmentation. Nevertheless, this is the real-world experiment dataset that we can get to the best of our knowledge and that’s why we evaluated the dataset’s targeted task even they are not directly same as our work. Ideally, evaluation would use UAV footage captured at the physical location represented by our simulation, or an artist-created aerial environment approximating it with turned 2D simulation for propagation, but this is truly beyond our current capable resources.

Such data lacking condition not only hinders the evaluation of proposed framework but also suggest lack of environment for robotic training. This scarcity, particularly for academic teams without access to wildfire agencies, domain experts, or proprietary data, is precisely the motivation for our simulation-grounded framework, i.e. the 2D to 3D proxy to controlled generated video pipeline.

To partially address the problem of testing robustness of our framework under scarce data condition, we also provided an experiment where we switched out the video generator as shown in Table 5 and below:

DirectionOur methodPure GPTPure ClaudePure Gemini
LTX query → VACE memory41.0019.0024.0028.00
VACE query → LTX memory44.4427.2723.2326.26
Here, we use different video generators to perturb the wildfire distribution to test the robustness of the system.

We truly hope reviewer to consider the practical availability of such data, the scope of the paper, and the additional evaluations we provided, even if they did not fully address the reviewer’s expectations.

FVD with real footage:

As mentioned in the previous section, real footage collection of UAV for wildfire videos is really lacking and we don’t have direct access to those videos. However, since the 3D proxy is directly procedurally recovered from 2D simulation automatically, we consider it faithful with simulation. That’s why we choose to evaluate the physics alignment with 3D proxy video (in Appendix D) and added a realistic appearance by querying the existing LLM (GPT, Claude, Gemini) since LLMs should have seen enough realistic wildfire videos implicitly during their training. We provided the physics alignment with proxy video in Appendix D and the appearance evaluation using implicit knowledge of LLMs is shown in Table 14 and below where LLM scores the video with 0-100 for realism:

Fire frontFuelBurn areaTerrainOverall
79.1768.7570.8381.2575.00

Thank you for the detailed response and revisions. After reading the revised paper, I appreciate that the authors have clearly narrowed their claims and explicitly acknowledged the lack of real-world validation for physical-driver estimation.

While the sim-to-real gap remains an open issue, I agree that the current experiments provide reasonable support for the paper’s more focused contribution in simulation-grounded video generation and memory-based VLM inference. It is certainly an interesting and meaningful topics and I'd like to see it advancing.

I therefore consider my main concern sufficiently clarified and am willing to raise my score.

We thank reviewer for considering the real data lacking issue and consider this paper will help advancing this area of research and raise the rating.

Reviewer_o92mOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the scarcity of wildfire videos with synchronized physical annotations. It converts SimFire trajectories over terrain and fuel maps into lightweight Blender proxies, uses controlled video generation to enrich appearance, and stores labeled episodes in a training-free multimodal memory. Retrieval, reranking, complementary reporters, and selective adjudication produce four environmental tags and six binary report fields. Reported exact four-tag accuracy is 51.5%, compared with 22.6% for the strongest query-only baseline; six-field reporting averages 77.3%. Cross-generator and three real-UAV evaluations provide complementary tests, with explicitly narrower perceptual targets on real footage.

Critical strengths
  1. The construction-to-inference pipeline is coherent and inspectable. Simulator-derived labels, reference-only metadata, and explicit disagreement routing connect the generated episodes to a concrete reporting task. The proxy is appropriately presented as an intermediate representation, not as a volumetric physical simulator.
  2. Reference memory provides substantial measured benefit in the studied setting: 51.5% versus 22.6% exact tag accuracy (Table 3), with text-memory and labeled-reference comparisons. Cross-generator results of 41.00%/44.44% exceed the best direct baselines of 28.00%/27.27% (Table 5a), reducing concern about an advantage confined to identical generator appearance.
Critical weaknesses

The real-UAV tests need more explicit sampling and task-adaptation protocols (Appendix L). The paper identifies the three observable tasks and specifies that FLAME fire area is derived from segmentation regions, with ten videos ranked per prompt. However, evaluated sample counts and selection rules, the construction and number of ranking groups, temporal aggregation of fire area, task-specific memory/reporter adaptations, actual prompts, and metric implementation details remain insufficiently specified for reproduction. These results concern smoke/fire detection and visible-area assessment, not physical-driver recovery.

Questions for authors
  1. Please provide per-table nominal/effective sample counts, clarify whether baselines use common query sets, and explain evaluation-time API or format-failure handling. For the generation-screening and resampling procedure already described in Appendix C, please report the numbers screened, retried, and resampled, and explain whether these steps relate to the percentage-grid observations above. Please also provide paired query-level uncertainty for the 4.7-percentage-point gain over majority voting in Table 1 and reconcile the experiment-to-model mapping, including GPT-5 in Table 3 versus GPT-5.2 in Appendix E, using the exact model identifiers or versions used in each experiment.
  2. Please confirm that the claimed gains are for the reported complete configuration, rather than isolated physical-proxy or agent-topology effects. Appendix F retains label-dependent appearance cues and the judge has top-five evidence versus the combined reporter's top-one.
3 comments

We sincerely thank reviewers for recognizing the value of the framework and asks for clarifications on details. We will address the unclear details below and we will open source our code for both 2D to 3D proxy to generation pipeline and the agentic framework if accepted.

Details of experiments in tables:

We realize this is missing in some of the experiments, and we will include these detail in final version:

Table 1, 2, 3, 4, 5a: we use 100 queries from the validation split where for Table 5a, 100 LTX and 100 VACE are used separately. Table 5b: For Boreal, the dataset contains 22 dataset clips with 4 for calibration and 18 for testing. The test clips form nine positive/negative pairs, so 18 clips should not be presented as 18 independent fire incidents. For Fireman-UAV, the evaluated dataset contains 13 source videos with 104 RGB/thermal frame pairs. For FLAME: it’s included in Appendix L.

Appendix A uses 100 queries from validation split. The random/stratified baseline percentages are analytical expectations, not additional random trials. Table 12, 13 uses 400 training set generated/proxy pairs; 12 sampled time points per pair are averaged within each pair. Among those, ; the metric-specific effective counts are 384 measurable paired smoke axes and 357 measurable paired growth-front axes. Table 14 uses the 100 queries from validation split with each video being splitted to 8 frames and submitted to GPT api. For Table 20, 21, 22, we used the nominal 100-query validation split. The 16 reference videos in Table 20 and the varying k in Table 22 do not multiply the number of evaluation queries

For the training set 400 existing videos screened; 316 passed immediately; 84 locations underwent 184 regeneration attempts; 43 locations were recovered; 41 ended below threshold. This pass retained 359 outputs. The audited screening script allows at most three regeneration attempts for a rejected existing output, then marks the location skipped. Replacement was run for the 41 below sampled locations. For Table 1, the measured confidence interval of gain compared from majority voting is [0.69, 9.45]. For LLM versions, we used version gpt-4.1-mini for Table 1 and gpt-5.2 for Table 1, 3 and 21. (we will fix the misleading GPT-5 in Table 3 and 21). We used gpt-5.6-luna for table 23 and 25, claude-sonnet-4-6 for table 23. Otherwise, the recorded GPT, Claude and Gemini version is accurate as shown in Appendix E. We will modify the manuscript to reflect the correct versions.

Confirm claim gain from complete pipeline

Yes. The reported improvements describe the performance of the complete evaluated configuration: simulator/proxy construction, label-conditioned video generation, multimodal memory and reranking, reporter-specific evidence, and judge-based resolution. The existing comparisons do not separately identify a causal effect of the physical proxy or of the agent topology. The generated video’s useful cues do come from both the physical proxy and the label-conditioned text prompts. However, we do want to clarify that for controlled video generation, proxy only guidance or prompt only guidance can’t give desired results and both simulation guided control and language guidance is needed. We will make this claim more clear in the manuscript in case misunderstanding happens.

The authors have addressed some of my concerns, and I have therefore decided to maintain my original score.

We thank reviewer for your evaluation and the continuous support of our paper.

Reviewer_3MrKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a pipeline that converts 2D SimFire trajectories into intermediate 3D Blender proxies representing simulator-derived terrain, fuel, fire activity, and wind cues and generates videos, creating a labeled memory bank for environmental driver-tag prediction and multi-VLM reports as inference, resolving any conflicting outputs through a VLM judge. This pipeline evaluates synthetic episodes using the same and different generators, as well as real-world UAV captures.

Critical strengths
  1. The converter pipeline is reusable and provides pairs of videos and simulator-derived labels in scenarios where physical ground truths are scarce. Detailed label definitions and construction rules support resource contribution.

  2. The complete system achieves 77.3% accuracy on six report fields, while reporter majority voting without judge scores 72.6% accuracy (Table 1), and the average per-field majority baseline scores approx. 58.9% (Appendix A). A comparison on Fireman against the nearest strongest baseline is 0.9625 vs 0.8625 balanced accuracy. However, this enhancement supports perception claims, not that of the environment driver's recovery.

Critical weaknesses
  1. Baseline comparisons do not fully establish the gains due to memory-based inference over simpler alternatives with the same label access. Four-tag predictions use the top-ranked retrieved episode’s tags generated by the simulator (Appendix C). The values in Table 3 demonstrate reranking benefits, with four-tag accuracy improving from 45.5% to 51.5% over embedding-only retrieval. However, simple classifiers (e.g., linear probe on frozen features) using the same 400 labeled memory episodes are not evaluated. Domain VLMs in Table 4 fail to exceed majority rates (Appendix A) mostly, making them less informative as baselines.

  2. Although Section 7 states that this study does not isolate the contribution of simulator-derived proxy structure from metadata-conditioned prompting, establishing the simulator-to-proxy stage's contribution requires prompt-only and neutralized-tag-prompt controls.

Questions for authors
  1. How do tuned KNN voting or a linear probe on frozen features, using the same episodes, compare with the no reranking and reranking baselines? Please report paired confidence interval differences.

  2. Can prompt-only and neutralized-tag-prompt controls isolate the contribution of proxy structure, particularly for moisture (its accuracy can come from two distinct routes)? Will the episodes, labels, and converter be released?

  3. Please report the actual number of evaluated queries for your experiment and whether accuracies are computed over a single set or averaged across multiple runs.

3 comments

We thank the reviewer for suggestions and recognition of the value of the proposed pipeline when truth data is scarce. Our main contribution is memory-based inference and structured wildfire reporting without task-specific weight fitting. The system establishes a benefit of the complete reporting procedure within this setting.

Tuned KNN and linear probes

We have added the requested tuned KNN and linear-probe comparisons. Retrieval and KNN use episodes at inference time without fitting weights. The linear probes fit a relevance score for query–memory pairs. The fitted linear approach is a supervised adaptation baseline, even though their pretrained encoders remain frozen. All memory-based methods below use the same 400 usable memory episodes, the same frozen embedded video features, and the same 100 validation queries. KNN voting settings and linear probe parameters are selected only within the memory split. Differences are in percentage points, with paired 95% confidence intervals where compared with the referenced method under that column.

MethodAccuracyMethod − dense [95% CI]Method − reranker [95% CI]
Training-memory majority7.1%-38.4 [-49.5, -27.3]-44.4 [-55.6, -33.3]
Top-one retrieval, no reranking45.5%Reference-6.1 [-15.2, 3.0]
Tuned KNN, independent votes46.5%+1.0 [-11.1, 13.1]-5.1 [-16.2, 6.1]
Tuned KNN, joint-bundle votes38.4%-7.1 [-16.2, 2.0]-13.1 [-24.2, -2.0]
Trained linear reranker41.4%-4.0 [-15.2, 7.1]-10.1 [-21.2, 1.0]
Memory retrieval + reranking (ours)51.5%+6.1 [-3.0, 15.2]Reference

We show experiments done for Table 4 where accuracy columns are percentages and the final column reports our full system minus the row method, in percentage points.

MethodFast spreadPeak areaBurnscarCanopyRapid + largeSevereMeanFull system − method [95% CI]
Training-memory majority56.759.860.855.760.859.858.9+18.4 [11.0, 25.8]
Trained linear reranker67.064.966.083.566.064.968.7+8.6 [0.5, 16.7]
Tuned KNN, independent tag votes70.161.964.989.764.963.969.2+8.1 [-0.3, 16.5]
Tuned KNN, joint-bundle votes69.160.861.977.361.962.965.6+11.7 [3.1, 20.3]
Combined memory reporter71.174.276.361.976.375.372.5+4.8 [0.9, 9.1]
Full reporting system (ours)75.377.380.473.279.478.477.3Reference

What does this establish about proxy structure, and will code be released?

The reported gains are for the complete simulator-to-proxy, generation, memory, and reporting configuration and we don’t claim independent section contribution as discussed in Sec. 7 and admit that the gain is systemwise.

Prompt-only and neutralized-tag-prompt controls could help separate these routes, but those controls are outside the present evaluation because neutralized-tag-prompt will give inconsistent ground truth appearance and stored tags/simulation parameters prompt only will destruct the simulation guided guarantee and make video generation purely rely on generators’ internal distribution. We therefore attribute the measured gains to the complete configuration. The reporting gain similarly includes the judge's additional evidence and computation; it is not an isolated estimate of agent-topology effects. Yes, we will release both converter and agentic pipeline code upon acceptance.

Actual number of evaluated queries

We realize this is missing in some of the experiments and we will include these detail in final version if accepted:

Table 1,2,3,4,5a: we use 100 queries from the validation split where for Table 5a, 100 LTX and 100 VACE are used separately. Table 5b: For Boreal, the dataset contains 22 dataset clips with 4 for calibration and 18 for testing. The test clips form nine positive/negative pairs, so 18 clips should not be presented as 18 independent fire incidents. For Fireman-UAV, the evaluated dataset contains 13 source videos with 104 RGB/thermal frame pairs. For FLAME: it’s included in Appendix L.

Appendix A uses 100 queries from validation split. The random/stratified baseline percentages are analytical expectations, not additional random trials. Table 12, 13 uses 400 training set generated/proxy pairs; 12 sampled time points per pair are averaged within each pair. Among those, the metric-specific effective counts are 384 measurable paired smoke axes and 357 measurable paired growth-front axes. Table 14 uses the 100 queries from validation split with each video being splitted to 8 frames and submitted to GPT API. For Table 20, 21, 22, we used the nominal 100-query validation split. The 16 reference videos in Table 20 and the varying k in Table 22 do not multiply the number of evaluation queries

All evaluations are not run through multiple turns but a single time evaluation.

Thank you for evaluating your method against simpler baselines using the same memory bank. These comparisons address my baseline request. The first table in the rebuttal shows that the advantage of the proposed system over tuned kNN and dense retrieval without reranking remains uncertain on exact four-tag prediction, while the second table provides stronger evidence for the full reporting system on the six fields. My concern about the incremental gains by simulator proxy component remains unresolved, so my overall assessment remains unchanged. Hence, I will maintain my rating.

We truly thank reviewer for the follow-up and for confirming that the additional comparisons address your baseline request. We would like to clarify our interpretation of the four-tag results and their implications for the overall assessment.

For exact four-tag prediction, our method achieves 51.5%, compared with 46.5% for tuned kNN, 45.5% for dense retrieval, and 41.4% for the linear reranker. Although the paired 95% confidence intervals include zero, they are centered on positive gains, and all point estimates favor our method. For example, the improvement over dense retrieval is +6.1 points [−3.0, 15.2], and over the linear reranker is +10.1 points [−1.0, 21.2]. We therefore view these results as suggestive, though not statistically conclusive, evidence of an advantage.

For reporting, the evidence is stronger: the full system achieves 77.3% mean accuracy, improving over the linear reranker by 8.6 points (95% CI [0.5, 16.7]) and the combined memory reporter by 4.8 points ([0.9, 9.1]).

Our main contribution is the full inference and reporting pipeline without task-specific weight fitting. As explained in our rebuttal, isolating the proxy’s benefit is not straightforward. Without the proxy, generation relies on language prompts alone and loses the direct visual guidance from the simulated fire dynamics. Neutralizing tag prompts can also cause the generated appearance to disagree with the stored simulation labels. Because these changes affect the correspondence between the videos and their labels, we cannot cleanly isolate the proxy’s contribution within the current evaluation. We therefore attribute the measured gains to the complete pipeline and explicitly acknowledged the proxy evaluation limitation. We hope the reviewer can take into account both these evaluation difficulties and the improvements shown in our additional comparisons with the baseline methods.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We have incorporated the additional experiments and clarifications from our rebuttal into the revised manuscript. We've marked our changes in red.

  1. Added retrieval baselines (Section 4; Appendix N.1, Table 26). We added comparisons with training-memory majority, tuned KNN with independent-tag and joint-bundle voting, and a trained linear reranker. These comparisons use the same 400 memory episodes and frozen 2,048-dimensional Qwen video features, with hyperparameters selected only within the memory split. Retrieval with reranking achieves 51.5% exact four-tag accuracy, compared with 46.5% and 38.4% for the two KNN variants and 41.4% for the linear reranker.
  2. Expanded downstream reporting comparisons (Section 5.4; Appendix N.2, Table 27). We added per-field and mean accuracies for the additional baselines across all six reporting tasks. In this comparison, the full system achieves 77.3% mean accuracy versus 72.5% for the combined-memory reporter.
  3. Added uncertainty estimates and clarified evaluation units (Section 4; Appendices D and N, Table 28). We report paired confidence intervals for the added baseline comparisons and add the confidence interval for the full system's gain over majority-vote reporting. We also document the nominal validation-query counts and single-run evaluation protocol. For proxy alignment, we specify 400 generated/proxy pairs, 12 sampled time points averaged within each pair, and effective counts of 384 smoke-axis pairs and 357 growth-front-axis pairs.
  4. Clarified the training-free setting and the scope of the gains (Sections 3.2 and 6; Appendix N.4). We define training-free inference as using frozen pretrained models without fitting task-specific weights, while retaining labeled memory as inference-time evidence. The fitted linear reranker is identified as a supervised adaptation baseline. We also clarify that the measured gains concern the complete evaluated pipeline: generated cues depend on both proxy structure and label-conditioned prompts, and judge-based reporting uses additional evidence and computation. The comparisons therefore do not isolate causal effects of proxy structure or agent topology.
  5. Documented generation screening and replacement (Appendix C). We added the screening audit for 400 existing training videos: 316 passed immediately, 43 locations were recovered through regeneration, and 41 remained below threshold. We report the 184 regeneration attempts across 84 locations, the limit of three additional attempts per rejected output, and the separate replacement step.
  6. Clarified the existing generalization and real-UAV evaluations (Section 5.5; Appendices D.1 and L). We specify separate nominal 100-query sets for the two cross-generator directions. For Boreal, we clarify that four clips are used for calibration and 18 test clips form nine positive/negative pairs; for FireMan-UAV-RGBT, we specify 104 RGB/thermal frame pairs from 13 videos. We explain the available annotations and dataset limitations, and clarify that these tests assess observable smoke/fire and fire-area tasks. Proxy alignment and VLM cue ratings are distinguished from calibrated physical fidelity or realism against matched real footage.
  7. Corrected model identifiers and clarified code availability (Appendices E and N.4). We corrected the model-version labels, including GPT-5.2 for the tag-prediction and reranker comparisons, and documented the exact model identifiers used in the reporting, Boreal, and FLAME evaluations. We also explicitly commit to releasing both the 2D-to-3D proxy converter and the agentic memory and reporting pipeline upon acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.