acceptodds
Under review as a conference paper at ICLR 2027

Thinking Beyond Videos: Unifying Video Reasoning and Deep Research for Open-World Video Agents

Abstract

Open-world video understanding often requires locating sparse visual evidence in a video and acquiring knowledge that is contained neither in the video itself nor in the model's parameters. Thinking-with-Videos actively localizes such evidence, and Deep Research acquires such knowledge over multiple search steps, but the two directions are developed in isolation. To bridge this gap, we introduce VideoRover, a unified Video Deep Research framework that iteratively coordinates video cropping, multimodal search, and webpage browsing. Given a video-question pair, VideoRover conditions each action on the observations accumulated so far, so that localized clips anchor external search, while retrieved evidence triggers further video inspection and verification. Existing data rarely cover such trajectories, so we build an automated pipeline that constructs questions jointly dependent on video and web evidence and synthesizes verified interaction trajectories, yielding 26K SFT examples and 3K hard RL instances. We also introduce VideoRover-Bench, a benchmark stratified by video duration and research difficulty. Experiments on VideoDR and VideoRover-Bench show VideoRover-8B-RL performs on par with proprietary models in direct-answer settings while outperforming some larger models given the same tools. Additional experiments further demonstrate broad generalization and adaptive coordination between video grounding and external retrieval.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Understanding, Deep Research, Agentic RL

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_aWSxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VideoRover trains a video agent to investigate questions through video cropping, image search, text search, and webpage browsing. Its central claim is that video observations guide retrieval while retrieved evidence can redirect subsequent video inspection. The work combines synthesized trajectories, SFT followed by GSPO, and a benchmark stratified by video duration and research difficulty. Experiments show substantial improvements over the backbone equipped with the same tools, but the methodological advance over closely related agents remains unclear, and the interpretation of the tool-use analyses extends beyond what their measurements establish.

Critical strengths
  1. The improvement over the same backbone with tool access is substantial. Table 1 reports approximately 58% aggregate accuracy for VideoRover-RL versus 32% for the backbone in the agentic setting. This comparison supports the practical value of the training recipe more directly than comparisons against proprietary models answering without tools. RL also improves over SFT across every reported duration–difficulty subset.

  2. The work provides a useful training and evaluation setup for an important task. Connecting sparse visual clues to external factual information is a realistic requirement for video agents. The synthesized trajectories and benchmark stratification provide infrastructure for studying this behavior. This supports the value of the work as a data and systems contribution.

Critical weaknesses

1. The methodological novelty is limited and insufficiently articulated.

The introduction characterizes existing approaches as keeping video reasoning and retrieval separate. However, VideoSearcher, already included as a baseline, combines temporal localization, spatial inspection, and multimodal web retrieval within a unified agent trajectory. Its interaction history accumulates both video and web observations, and it also uses SFT followed by RL. Consequently, a shared research state and the integration of visual and search tools are not, by themselves, a clear methodological departure.

Video-DeepResearch, which the manuscript cites, develops a closely related pipeline covering video grounding, visual search, text search, webpage visits, synthesized training trajectories, SFT followed by RL, and a dedicated benchmark. Its trajectory construction uses stage-wise tool unlocking, whereas VideoRover emphasizes flexible interleaving of video and retrieval actions. This is a potential distinction, but the paper needs to explain what specific policy or state-update design makes that interleaving a new mechanism beyond a general agent conditioned on interaction history. As written, I see an incremental combination of established components, with the clearest contributions being the data and task-specific training recipe.

2. The interpretation of the ablation and tool-use analyses should be narrowed.

In Section 6.4, the authors claim that VideoRover "does not follow a fixed tool-use pipeline" and adaptively selects actions based on each question's need for external knowledge. However, Figure 4 only reports average per-example tool-call counts across benchmark types. These aggregate statistics show differences in tool-use frequency but provide no evidence of action ordering or how intermediate observations influence subsequent actions. Therefore, Figure 4 does not substantiate the claim of adaptive tool use.

Appendix F provides qualitative examples of varying tool sequences and evidence-driven revisions, but these examples lack systematic quantitative analysis. They do not establish how frequently adaptive behavior occurs or how much it contributes to performance.

Similarly, Table 3 only evaluates the effect of removing individual tools on accuracy. While these results demonstrate the utility of the tools, they do not establish the flexibility or adaptivity of the tool-use strategy.

Overall, the current experiments lack direct evidence of adaptive tool-use behavior and fail to quantify its contribution to performance. More targeted analyses are needed to substantiate the claims in Section 6.4.

Questions for authors
  1. What specific methodological contribution distinguishes VideoRover from VideoSearcher's unified interaction history and Video-DeepResearch's grounding-and-retrieval pipeline, beyond the tool interface, training data, and training recipe?

  2. Is Section 6.4's conclusion about a non-fixed pipeline supported by evidence beyond Figure 4's average call counts? Please identify that evidence or clarify the intended scope of the conclusion.

3 comments
AuthorsOpenReview ↗

Response to Reviewer aWSx

We sincerely thank the reviewer for the helpful review. We hope our replies below address your key concerns.

W1&Q1: Unclear methodological novelty over VideoSearcher and Video-DeepResearch beyond tool interfaces, training data, and training recipes.

Thank you for raising this important concern. We agree that a shared interaction history and the integration of visual and retrieval tools alone do not distinguish VideoRover. Relative to VideoSearcher and Video-DeepResearch, our contribution centers on a knowledge-anchored data construction and training framework that enables an 8B model to coordinate dense segment-level video inspection with external research.

VideoSearcher constructs training data by adapting existing image–text QA pairs to videos retrieved using their entities. VideoRover instead starts from encyclopedia knowledge, grounds entities in videos, and generates questions linking visual evidence to external facts. While VideoSearcher already supports temporal refinement and multimodal retrieval, VideoRover emphasizes dense segment-level inspection and explicit webpage visitation. Video-DeepResearch uses stage-wise tool unlocking for trajectory construction and operates on at most 20 pre-extracted keyframes. Its frame selection and spatial cropping cannot recover temporal evidence omitted by this initial sampling. VideoRover can instead densely resample selected segments during research to inspect brief events and evidence across neighboring frames, particularly in long videos. We have added a brief discussion of these methods to the Related Work section. The results in Table 1 further show that VideoRover achieves higher average accuracy than both methods.

Beyond question construction, our trajectory synthesis permits flexible tool selection after initial video inspection, allowing intermediate observations to guide further video inspection, search refinement, or webpage verification. This concerns learning effective coordination rather than a new state-update rule. We test its benefit by retaining the same model, tools, and interaction limit while enforcing a repeated crop_video → image_search → text_search → visit cycle with freely selected tool arguments.

MethodVideoDRS-ES-MS-HL-EL-ML-HAvg.
VideoRover-RL56.0080.0064.0040.0070.0052.0042.0057.71
+ Fixed order48.0074.0058.0034.0064.0044.0034.0050.86

Performance decreases on VideoDR and all six VideoRover-Bench subsets, supporting the benefit of flexible coordination over this fixed pipeline, beyond access to the same tools. We have added this experiment to Appendix F.

W2&Q2: Average tool-call counts and tool ablations do not establish adaptive scheduling or quantify its contribution to performance.

Thank you for this important clarification. To examine action ordering directly, we additionally analyze all 300 trajectories from each model on VideoRover-Bench:

MetricVideoRoverVideoSearcherVideo-DeepResearch
Retrieval → video inspection24.3%5.3%N/A
Text search/visit → visual tool20.0%3.3%0.3%
Visit → new text query52.0%N/A21.3%
Accuracy57.33%39.67%47.33%

Each behavioral percentage denotes the proportion of trajectories exhibiting that pattern. Tool interfaces differ across models: Video-DeepResearch combines visual inspection with search, while VideoSearcher lacks independent webpage visitation. The third row counts a previously unused text query as the next action after a webpage visit.

Compared with the other two models, VideoRover more frequently returns to visual inspection after retrieval and, where webpage visitation is available, follows it with new search queries. These patterns are consistent with more flexible tool use according to current research needs, such as revisiting the video to locate relevant evidence or searching for additional information after inspecting retrieved sources. We have added this trajectory analysis to Section 6.5.

As detailed in W1&Q1, enforcing a fixed tool cycle with the same model, tools, and interaction limit reduces performance across both benchmarks. Together, these results support non-fixed tool scheduling and its practical benefit over the tested fixed pipeline.

I thank the authors for their response and have decided to raise my score to 3.

Dear Reviewer aWSx,

Thank you very much for your recognition of our work and for raising your score. We sincerely appreciate your thoughtful review and constructive feedback.

Sincerely,

The Authors

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

This work provides a unified framework that combines thinking with videos i.e., video reasoning, and deep research pipeline, where the framework is given the ability to reason and also search web for more information using text and image web search tools. This facilitates coordination among different tool-calling tasks like video cropping, multimodal search and webpage browsing. The work also provides an automated pipeline for data generation for training such a framework, and provides 26K SFT examples and 3K RL samples, and thereby training a 8B model based on Qwen3-VL-8B with the proposed dataset. To evaluate the performance of existing methods and proposed method, this paper introduces a benchmark, VideoRover-Bench which tests the models capabilities of tool calling and orchestration.

Critical strengths
  1. Combines reasoning with deep research (web search).
  2. Provides new SFT (26K examples) and RL (3K examples) dataset for finetuning base model for the proposed task.
  3. SFT and RL on the base model to achieve better performance in using the tools and achieve similar performance as proprietary models.
  4. Proposes VideoRover-Bench, a new benchmark to evaluate the models capabilities for tool use.
Critical weaknesses
  1. The proposed method does show performance improvement compared to direct answering and agentic tool use compared to other methods, but provides little to no information about the latency this setup introduces.
  2. The work doesn’t provide a tool-call latency vs. accuracy trade-off for the benchmark. This might help us better understand where the model is spending most of its budget.
  3. From Table 3. It can be seen that removing crop_video doesn’t seem to affect much. This indicates that most of the questions are factual and can be obtained by querying the relevant site. The paper does not distinguish between web search for “using the web” to find relevant information vs. “finding the answer from the web”. Is there a setup to ensure that the search doesn’t lead to answer leakage?
  4. What happens to the cases where relevant information on the web page is removed or is taken down? Is there a way in which the correctness of the retrieved information is evaluated, or is it assumed that the search results are always factually correct? If it is the latter, then this design needs to be rethought for benchmarks that might not have factual information on the internet.
  5. The work only cites these works but doesn’t compare the VideoRover-8B model with strong related works such as Video-DeepRsearcher (https://arxiv.org/abs/2608.03979), VideoBrowser (https://arxiv.org/abs/2512.23044). Recommendation: (a) add the performance of these models to the general video understanding benchmark table (Table 2) and (b) evaluate VideoRover-Bench on these models.
Questions for authors
  1. Can the performance of the SFT and RL models be provided as a direct answer (no agentic setup)? To check the zero-shot performance without tool usage?
  2. How does this differ from existing video agentic setups and benchmarks? More emphasis on existing works should be done.
  3. What is the delay/computational overhead because of the tool calls, and how does what is the trade off of the tool call vs accuracy? Hint: please add average latency, output tokens, and tool calls per question.
  4. It would be interesting to see if the training data helps to use the correct tools? It would be interesting to see how the trajectory of the trained model changes compared to the base model.
  5. Are the Table 2 results reproduced? Does the time written below the benchmark name suggest the average duration of videos per benchmark? What is the number of frames, and what is the eval setup?
  6. Is there a possibility that the wrong trajectories are negatively rewarded? Example: if the model gets stuck in a loop for a tool, are any such cases observed?
  7. What if the answer is directly present on the web? Does it find shortcuts in answering or hack the answer instead of using tools to obtain the answer? (for weakness 3)
  8. Visit: For this tool, is it allowed to query/visit any websites? Are there any restrictions? (for weakness 3)
  9. From Table3. It can be seen that removing crop_video doesn’t seem to affect much. Does it mean that only 64 initial frames are enough and the questions are more dependent on the search tool?
  10. It would be interesting to see the performance of the best-performing video reasoning models, ex. VideoZoomer or Video-o3 on VideoRoverBench. This might provide us indication of which components of the video reasoning training are more helpful.
  11. New models such as Qwen3.5 already have tool-calling capabilities; does the provided training data help/improve over such models too?
  12. For tab 1, what is the prompt for the open model evaluation for direct-answer?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Ncqi (1/3)

We sincerely appreciate the reviewer's time and careful feedback. We hope our responses below address your questions.

W1&W2&Q3: Missing latency–accuracy trade-off analysis and cost breakdown, including latency, output tokens, and tool calls.

Thank you for these suggestions. We evaluate the latency–accuracy trade-off of VideoRover-RL by comparing direct answering with its standard tool-enabled inference. Avg. accuracy averages VideoDR and the six VideoRover-Bench subsets equally.

Inference modeVideoDR latency (s/example)VideoRover-Bench latency (s/example)Avg. accuracy (%)
Direct answering18.3022.1111.57
Tool-enabled inference581.66536.3857.71

We further profile latency, tool usage, and output tokens across all 300 VideoRover-Bench examples:

Metriccrop_videoimage_searchtext_searchvisitModel inference
Average latency per invocation (s)1.7618.314.8510.3333.37
Average invocations per question1.741.106.962.7713.51
Average cumulative time per question (s)3.0620.0833.7928.61450.81
Average output tokens per turn211.19177.26269.25284.45260.79

Tool-enabled inference substantially increases latency but also yields a large improvement in accuracy, demonstrating the value of external tools for this compact 8B model. Multi-round model inference accounts for approximately 85.5% and 84.0% of total latency on VideoDR and VideoRover-Bench, respectively, while tool execution accounts for 14.5% and 15.9%. Most of the elapsed time is therefore spent on iterative reasoning over the growing video–web context, rather than tool execution itself.

For evidence-intensive video deep research, we consider the additional inference time justified by the substantial accuracy improvement. We have added these measurements and the timing breakdown to Appendix G of the revised manuscript.

W3&Q7&Q9: Unclear need for video inspection beyond the initial 64 frames and potential web-answer leakage or shortcuts.

Thank you for raising this important question. Removing crop_video leaves the initial video input of up to 64 uniformly sampled frames intact, so this ablation measures the benefit of additional temporal inspection, not whether video evidence is needed. It reduces overall accuracy from 57.71% to 49.57% and VideoDR accuracy from 56% to 45%, showing a substantial contribution. Initial frames can suffice when relevant entities are clearly visible, whereas crop_video enables closer inspection of brief or fine-grained evidence, particularly in long videos.

Regarding answer leakage, finding relevant facts or even the final answer on a webpage is not itself a shortcut if identifying the relevant information requires first recognizing the target from the video. During data construction, we withhold the target identity from the question and refer to it indirectly through video content. For benchmark examples, we additionally verify joint dependence on video evidence and external knowledge through manual checking. To test whether questions alone enable web-based shortcuts, we evaluate VideoRover-RL with only the textual question and access to text_search and visit, removing all video input. It answers only 1/100 questions on VideoDR and 1/300 on VideoRover-Bench correctly. These results strongly suggest that question-only web search is insufficient for most examples and support the importance of identifying relevant entities from video evidence.

W4: Handling unavailable or removed webpages and verifying the factual accuracy of retrieved information.

Thank you for raising this important question. We do not assume that search results are always available or factually correct. During benchmark construction, we retain the supporting passage, source URL, and reference answer, followed by independent verification and human review. Therefore, the ground truth remains valid even if the original webpage is later modified or removed.

At inference time, search results are treated as candidate evidence. If a page is inaccessible or insufficient, the agent can refine its query and search for alternative sources. As also noted by Vision-DeepResearch, real-world web environments are inherently noisy and unstable, making iterative evidence collection necessary. However, webpage browsing enables rather than guarantees faithful verification. The Limitations section already acknowledges our dependence on the availability and quality of external search results.

AuthorsOpenReview ↗

Response to Reviewer Ncqi (2/3)

W5: Missing comparisons with Video-DeepResearch and Video-Browser.

Thank you for this suggestion. Video-DeepResearch operates on a pre-extracted set of up to 20 keyframes. Its visual tool selects and spatially crops these existing frames for visual search, but cannot densely resample a temporal segment to recover evidence missed by the initial sampling. VideoRover instead uses crop_video to inspect selected segments at higher temporal density, supporting finer-grained temporal grounding, particularly for brief events in long videos. Video-DeepResearch uses Qwen3-VL-30B-A3B and Qwen3.5-35B-A3B backbones, whereas VideoRover uses Qwen3-VL-8B. We did not include this comparison originally because the base models differ in scale and architecture.

We have now evaluated Video-DeepResearch-30B-A3B on VideoRover-Bench and added the results to Table 1 of the revised manuscript:

MethodVideoDRS-ES-MS-HL-EL-ML-HAvg.
VideoRover-RL56.0080.0064.0040.0070.0052.0042.0057.71
Video-DeepResearch-30B-A3B62.00†68.0054.0026.0060.0044.0032.0049.43

Here, † denotes the VideoDR result reported in the original paper. Although Video-DeepResearch reports higher accuracy on VideoDR, VideoRover-RL performs better across all six VideoRover-Bench subsets and achieves a higher overall score.

Video-Browser addresses a different setting. Starting from a natural-language query, it searches for, screens, and inspects candidate videos, with its hardest questions requiring reasoning across multiple sources. VideoRover instead receives a specific input video and combines temporal grounding with external knowledge retrieval. Video-Browser’s original setup is therefore not directly applicable to the tasks evaluated in our paper and is unsuitable for a direct comparison. We have added a brief discussion of both methods to the Related Work section.

Q1: Direct-answer performance of the SFT and RL models without tools.

Thank you for this suggestion. We additionally evaluate both VideoRover-SFT and VideoRover-RL in a direct-answer setting, retaining the video and question inputs but disabling all tools.

ModelSettingVideoDRS-ES-MS-HL-EL-ML-HAvg.
VideoRover-SFTAgentic39.0058.0044.0030.0054.0032.0018.0039.29
VideoRover-SFTDirect answer10.0018.004.006.0010.004.004.008.00
VideoRover-RLAgentic56.0080.0064.0040.0070.0052.0042.0057.71
VideoRover-RLDirect answer7.0022.0010.006.0016.0012.008.0011.57

Both models perform substantially better with tools than with direct answering, suggesting that their strong performance primarily relies on tool-enabled interaction rather than direct-answer capabilities alone.

Q2: Insufficient distinction from existing video agents and benchmarks.

Thank you for this question. VideoRover extends video inspection with image_search, text_search, and visit, enabling the model to alternate between examining video segments and retrieving external evidence. While VideoSearcher and Video-DeepResearch also combine visual and retrieval tools, our contribution emphasizes task-specific data construction and training for coordinated video deep research with an 8B model. We have added a brief discussion of these methods and the VideoDR benchmark to the Related Work section.

Regarding benchmark design, VideoRover-Bench requires identifying indirectly referenced targets from video and retrieving external knowledge to answer questions. Compared with VideoDR and related benchmarks, it provides a balanced breakdown across two video-length groups and three difficulty levels, with 50 examples per subset and 300 examples across nine knowledge domains. All examples are manually verified for answer correctness and joint dependence on video and external evidence.

Q4: Impact of training on tool selection and trajectories relative to the base model.

Thank you for this suggestion. Our training data teaches the model the tool-call format, how to use relevant tools for visual grounding and external evidence gathering, and how to reason over the resulting video and web observations. Comparing Qwen3-VL-8B with VideoRover-RL on 300 VideoRover-Bench trajectories, the proportion using image_search increases from 19.33% to 65.00%, using visit from 51.33% to 99.00%, and returning to crop_video after retrieval from 3.67% to 24.33%. Meanwhile, exact duplicate requests decrease from 26.23% to 2.76% of executed calls, and accuracy increases from 32.33% to 58.00%. These results indicate more purposeful evidence gathering and richer video–web coordination, alongside substantially improved video deep-research performance.

AuthorsOpenReview ↗

Response to Reviewer Ncqi (3/3)

Q5: Clarify Table 2 result reproduction, duration labels, frame counts, and evaluation settings.

Thank you for requesting this clarification. All results in Table 2 were reproduced by us. For a consistent comparison, every model receives the same initial video input, sampled at 1 FPS with at most 256 frames and a maximum frame resolution of 224 × 224. For all methods equipped with temporal cropping, each cropped segment follows the same configuration as VideoRover: 2 FPS, at most 32 frames, and a maximum frame resolution of 448 × 448. Other model-specific inference settings and evaluation protocols follow the official implementations. We have clarified in Appendix C.2 that these input and cropping configurations are shared across methods.

The times below the benchmark names denote their average video durations in seconds. We have clarified this in the Table 2 caption and explicitly stated that all results were reproduced by us.

Q6: How erroneous or looping tool-use trajectories are rewarded and whether such cases occur.

Thank you for this question. During RL training, we observe a small number of rollouts that repeatedly invoke tools or fail to reach an answer before the interaction limit. Trajectories with missing final answers, formatting errors, or invalid tool calls are filtered out. These trajectories are excluded from optimization. For retained trajectories, correct and incorrect final answers receive rewards of 1 and 0, respectively.

Our reward is based on final-answer correctness, so a rewarded trajectory may still contain redundant tool calls or incorrect intermediate reasoning. We acknowledge that this outcome-based reward does not guarantee process correctness. Nevertheless, our experiments show that this simple training objective effectively improves video deep-research performance.

Q8: Website access restrictions for the visit tool.

Thank you for this question. The visit tool retrieves detailed webpage content from URLs returned by the search tools and is not restricted to a predefined set of websites. However, access remains subject to login requirements, paywalls, regional restrictions, anti-crawling mechanisms, and page availability. When access fails, the tool returns a failure message, allowing the model to plan its next action based on the current research state, such as refining the query, selecting another result, or visiting alternative sources.

Q10: Evaluate VideoZoomer and Video-o3 on VideoRover-Bench to assess the contribution of video-reasoning capabilities.

Thank you for this suggestion. We additionally evaluate VideoZoomer and Video-o3 on VideoDR and VideoRover-Bench:

MethodVideoDRS-ES-MS-HL-EL-ML-HAvg.
VideoRover-RL56.0080.0064.0040.0070.0052.0042.0057.71
VideoZoomer5.0010.004.004.006.006.000.005.00
Video-o311.008.004.006.0010.008.008.007.86

These results suggest that strong video-reasoning capabilities alone do not readily transfer to video deep research, highlighting the importance of combining visual grounding with external knowledge retrieval and task-specific training.

Q11: Whether the training data also benefits newer models with existing tool-calling capabilities, such as Qwen3.5.

Thank you for this question. Qwen3.5 already has strong tool-calling capabilities, but task-specific coordination remains important. On VideoRover-Bench, VideoRover-RL more frequently uses image_search (65.0% vs. 8.0%), returns to crop_video after retrieval (24.3% vs. 3.7%), and issues new queries after webpage visits (52.0% vs. 26.0%) than Qwen3.5-27B. These behaviors align with the visual grounding and iterative evidence gathering emphasized in our training data, which are central to the task we study. Moreover, Video-DeepResearch reports substantial gains from task-specific training of Qwen3.5-35B-A3B, demonstrating that such training remains effective even for models with existing tool-calling capabilities. Limited training resources and the rebuttal timeline prevent us from conducting this additional training, but we believe our data could also benefit newer models by strengthening task-specific video–web coordination.

Q12: Specify the direct-answer evaluation prompt for open models in Table 1.

Thank you for this suggestion. We have added the answer-evaluation prompt to Appendix I.4, with the full system and user prompts shown on the last page of the revised manuscript.

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

This paper introduces VideoRover, a video research agent that iteratively combines video inspection, image/text search, and webpage browsing to answer questions requiring both visual evidence and external knowledge. The authors construct synthetic training trajectories and a 300-question benchmark, and train a Qwen3-VL-8B agent through SFT followed by GSPO. Experiments report improvements over the backbone and competitive performance against larger tool-using models, with additional evaluations on general video benchmarks.

Critical strengths
  1. The framework demonstrates strong empirical effectiveness. With an 8B open-source backbone, VideoRover achieves aggregate performance close to proprietary models evaluated in the direct-answer setting, while substantially improving over the same backbone with tool access (Table 1). These results highlight the value of task-specific training and tool use for extending the capabilities of video understanding agents.

  2. The paper is clearly written and well organized. Figures 1–3 effectively illustrate the overall framework, data construction pipelines, and benchmark creation process, making the multi-stage workflow easy to follow.

  3. The supplementary code provides additional methodological details, and the demo videos help illustrate the agent's workflow.

Critical weaknesses
  1. Potential selection bias in video collection. Section 4.1 retains only the top three YouTube search results, without analyzing ranking-induced sampling bias or clarifying whether personalization or recommendation signals affect retrieval. A sensitivity analysis using larger candidate pools would help assess the impact on data coverage and reported performance.

  2. Potential error propagation in data synthesis (Figure 2). Incorrect video evidence extracted by Qwen during Visual Grounding and Verification could yield faulty visual anchors, while ambiguous or visually unsupported candidate questions produced during QA Generation could introduce unreliable training supervision. Although verification and filtering are applied, the paper does not quantify how often such errors survive these checks or assess their impact on training.

  3. Unquantified contribution of the larger evidence extractor. The webpage tool uses Qwen3-VL-30B-A3B to produce goal-conditioned evidence and summaries (Appendix B), so the deployed system relies on an additional model beyond the 8B backbone. Although agentic baselines share the tools, no ablation isolates this extractor's contribution. Comparing raw webpage text or smaller extractors would clarify how much performance depends on the larger model and help contextualize the small-model claims.

  4. Insufficient justification of retrieval dependence. Section 4.1 claims that failure by Qwen3-VL-4B without retrieval ensures dependence on external knowledge. However, such failure may reflect perception or reasoning limitations, while stronger models may answer from parametric knowledge. The authors should qualify this claim and analyze retrieval dependence using the reported stronger direct-answer baselines.

  5. Limited access to fine-grained spatial evidence. The reported 224×224/448×448 frame-resolution limits (Appendix C.1) may obscure small text or objects, while the tools offer temporal cropping without targeted spatial cropping or zooming. Given the introduction's emphasis on fine-grained visual evidence (Section 1, first paragraph, p. 1, lines 34–36), the authors should assess the impact of downsampling, particularly on OCR and small-object questions, or explicitly discuss this restriction as a limitation.

Questions for authors

see weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer rUvC (1/2)

We sincerely thank the reviewer for the detailed comments. We hope the responses below resolve your concerns.

W1: Potential ranking and personalization bias in top-3 video selection, without sensitivity analysis using larger candidate pools.

Thank you for raising this concern. We used a dedicated account for video retrieval and downloading that had no prior personal usage history. Queries were generated from Wikipedia-derived content rather than personal interests. These choices reduce the influence of pre-existing user preferences, but do not eliminate platform-level ranking bias or potential personalization during collection. We have noted this limitation in the Limitations section.

To examine sensitivity to candidate-pool size, we sampled 2,000 entries from our original collection and retrieved the top 20 YouTube results for each query. We compared the nested top-3, top-5, top-10, and top-20 pools:

Candidate poolRetrieved video entriesUnique videosUnique-video growth over Top-3Duplicate rate
Top-36,0005,8190.00%3.02%
Top-510,0009,62265.35%3.78%
Top-1020,00018,894224.69%5.53%
Top-2040,00036,803532.46%7.99%

Larger candidate pools substantially increase unique-video coverage, although duplication also increases. Restricting collection to top-3 therefore excludes many distinct lower-ranked candidates. Our top-3 choice prioritizes the highest-ranked results, which are expected to be more relevant to the query, while balancing coverage against processing costs. We also observed that lower-ranked videos often had limited relevance to the query, making them less suitable as candidates. Expanding to top-20 would require screening approximately 6.7 times as many candidates, substantially increasing downloading, video processing, annotation, and verification costs. We therefore consider top-3 a reasonable choice that balances candidate relevance and processing costs.

W2: Unquantified propagation of visual grounding and QA generation errors into training despite verification and filtering.

Thank you for raising this concern. We reuse the same Qwen model to maintain consistency in entity localization between visual grounding and trajectory construction, improving the likelihood of constructing coherent trajectories. However, we acknowledge that this can propagate correlated errors. Independent verification by GPT-5.4-mini helps mitigate this risk, while benchmark construction additionally includes Gemini-based checks and manual review.

To quantify candidate quality, we reviewed the human annotation logs from benchmark construction. Among 614 candidate examples, 121 (19.71%) were discarded for issues including ambiguous question wording and inaccurate visual-entity identification. The remaining 493 (80.29%) proceeded to subsequent balancing and sampling. This retention rate indicates that a substantial majority of the audited candidates passed manual quality checks, providing encouraging evidence of candidate quality. Because the training and benchmark candidates originate from the same upstream synthesis procedure, this audit provides a useful diagnostic of candidate quality.

W3: Unquantified contribution of the larger webpage evidence extractor, requiring ablations to contextualize the small-model claims.

Thank you for raising this concern. Following the auxiliary webpage-summarization approach used in Vision-DeepResearch[1], we use Qwen3-VL-30B-A3B to extract goal-relevant evidence from retrieved webpages. To assess dependence on this larger extractor, we replaced it with Qwen3-VL-8B while keeping the VideoRover-RL backbone unchanged.

Evidence extractorVideoDRS-ES-MS-HL-EL-ML-HAvg.
Qwen3-VL-30B-A3B56.0080.0064.0040.0070.0052.0042.0057.71
Qwen3-VL-8B54.0086.0066.0040.0070.0056.0040.0058.86

The smaller extractor maintains comparable performance, with a slightly higher overall score in this run. These results suggest that the observed performance does not critically depend on the larger extractor. Its role is to select and summarize relevant information from supplied webpage content, rather than independently conduct the full video–web research process, and the 8B extractor appears sufficient for this role in our evaluation.

Reference

[1] Huang, Wenxuan, et al. "Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models." ICML 2026.

AuthorsOpenReview ↗

Response to Reviewer rUvC (2/2)

W4: Failure of a weak model does not establish retrieval dependence; stronger direct-answer baselines need analysis.

Thank you for this clarification. We use Qwen3-VL-4B as a lightweight screening model because it enables efficient filtering of large candidate pools. We agree that failure without retrieval may reflect perception or reasoning limitations rather than missing knowledge. We have therefore removed the claim that this filtering "ensures dependence on external knowledge" and clarified the screening procedure in Section 4.1.

Retrieval requirements vary with model architecture, scale, training, and prior knowledge, making it difficult to guarantee that every question requires external retrieval for every model. Indeed, several proprietary direct-answer baselines in Table 1 achieve strong performance using the video and their internal knowledge without retrieval. As model capabilities advance, future models may likewise answer more of these questions directly. Similarly, VideoDR assesses video–web dependency through human annotation and video-only/web-only checks, rather than requiring a specific model to fail without retrieval. Its construction targets information beyond the video, not whether that information already exists in a particular model’s parameters.

Our use of Qwen3-VL-4B provides a scalable screening step that removes questions it can already answer correctly without retrieval. Our contribution lies in the broader data-construction and training pipeline. Training with the resulting data and framework improves the 8B agent’s overall score from 31.86% to 57.71%, demonstrating its practical value for developing video deep-research capabilities.

W5: Low frame resolutions and missing spatial cropping may limit OCR and small-object recognition, requiring evaluation or explicit acknowledgment.

Thank you for raising this concern. Unlike single-image tasks, video understanding requires processing many frames. Agentic video reasoning further accumulates cropped frames, tool observations, and reasoning history, placing additional pressure on the context budget. Our resolution choices follow the budget-conscious practice of prior video-understanding work, such as LongVT[2], which defaults to a per-frame pixel budget of 224×224, and VideoTemp-o3[3], which uses 224×224 or 448×448 depending on the evaluation benchmark.

To assess resolution sensitivity, we evaluated Qwen3-VL-8B on 200 TextVQA[4] examples and V*Bench[5] at different resolutions to assess OCR and small-object recognition, respectively:

Dataset112×112224×224336×336448×448560×560672×672784×784896×896
TextVQA28.0066.9582.9592.4595.4595.9596.9596.45
V*Bench32.9842.9347.1256.5460.2162.8368.5969.63

Increasing resolution generally improves performance on both datasets, indicating better OCR and fine-grained visual perception. At higher resolutions, marginal gains diminish, particularly on TextVQA. We consider our current resolution settings sufficient for the visual evidence required by our tasks. These results motivate our coarse-to-fine design: using 224×224 frames for an efficient global overview, followed by 448×448 frames to inspect key temporal segments in greater detail while managing the context budget.

Reference

[2] Yang, Zuhao, et al. "Longvt: Incentivizing" thinking with long videos" via native tool calling." CVPR 2026.

[3] Liu, Wenqi, et al. "VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos." ICML 2026.

[4] Singh, Amanpreet, et al. "Towards vqa models that can read." CVPR 2019.

[5] Wu, Penghao, and Saining Xie. "V*: Guided visual search as a core mechanism in multimodal llms." CVPR 2024.

Dear Reviewer rUvC,

With less than 48 hours remaining in the discussion period, we would like to kindly follow up on our rebuttal. We have made every effort to respond thoroughly to your comments and would greatly appreciate any further feedback. If any concerns remain, please let us know so that we can provide additional clarification before the deadline.

Thank you again for your time and consideration. Even a brief update would be very helpful.

Sincerely,

The Authors

Reviewer_iKaxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary
  1. The authors introduce VideoRover, and agentic framework that can identify entities in videos by searching world knowledge and then answer questions about the entity.
  2. Each trajectory rollout involves tool use of different kinds including temporal cropping, image search, text search, and web browsing until the answer is produced.
  3. To this end, the authors curate a benchmark and synthetic training trajectories via an automated pipeline using Qwen3.5-27B and DeepSeek-V4-Flash and then SFT and RL train Qwen3-VL-8B on this data.
  4. The authors evaluate the SFT and RL trained versions of VideoRover on their benchmark and some other standard benchmarks like VideoDR, Video-MMMU etc.
Critical strengths
  1. The authors create a new benchmark that requires video evidence and external knowledge to identify an entity in a video. This is a common workflow people use when doing video based research so it is definitely very useful to hill-climb on this and the authors have provided a balanced dataset spanning video durations and difficulty with broad topic coverage.
  2. The authors also provide an automated pipeline for curating samples for such a task and for generating the training trajectories to finetune LLM agents which presents a scalable way of improving on this task.
Critical weaknesses
  1. Soundness : The proposed task seems to have 2 stages, identify the entity in the video and answer questions about it. The second part does not seem to require visual evidence at all and can likely be done by a stand-alone LLM that can do web search. The authors need to run a staged experiment by providing the oracle entity from stage 1 and then evaluating stage 2 separately to see if these two stages need to be combined at all.
  2. Novelty : The critical components involved in stage 1 seem to already exist in VideoDR (video evidence and external knowledge) and videosearcher-8b demonstrates many of the same skills (search, key frame selection, cropping). The paper does not sufficiently distinguish the benchmark or method from VideoDR and video-searcher respectively. It would also be really useful to see which of these components/skills make the task difficult by providing the oracle solution for each skill separately during evaluation.
Questions for authors

please see weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer iKax

We sincerely appreciate the reviewer's thoughtful suggestions. We hope our replies below address your questions.

W1: Unclear need for integrated video reasoning and web search; an oracle-entity two-stage evaluation is needed.

Thank you for this suggestion. We evaluate an explicit two-stage decomposition using both Qwen3-VL-8B and VideoRover-RL. In Stage 1, the model receives the video and question and identifies relevant entities using crop_video and image_search. In Stage 2, a separate model context receives only the predicted entity names and the original question, using text_search and visit to answer without access to the video or preceding interaction history.

To directly address the requested oracle experiment, we additionally replace the predicted entities with the annotated key visual entity names from VideoRover-Bench. Both models then answer using only these names, the question, and the two web tools. Below, Two-stage uses predicted entities, Oracle entity uses annotated entities in the same staged setting, and Avg. is the arithmetic mean of all reported accuracy values in each row, excluding missing entries.

MethodVideoDRS-ES-MS-HL-EL-ML-HAvg.
Qwen3-VL-8B29.0058.0038.0016.0042.0030.0010.0031.86
+ Two-stage10.0028.0016.004.0022.0018.006.0014.86
+ Oracle entity—60.0050.0034.0064.0050.0050.0051.33
VideoRover-RL56.0080.0064.0040.0070.0052.0042.0057.71
+ Two-stage44.0064.0034.0032.0048.0046.0036.0043.43
+ Oracle entity—84.0068.0054.0080.0070.0074.0071.67

The two-stage pipeline using predicted entities performs worse than integrated inference for both models across both benchmarks and all six VideoRover-Bench subsets, supporting the effectiveness of jointly coordinating video inspection and external retrieval. Replacing predicted entities with oracle identities substantially improves performance, highlighting visual entity recognition as a key bottleneck. Given the same oracle entities, VideoRover-RL also outperforms Qwen3-VL-8B, suggesting that our training strengthens downstream tool-assisted retrieval and reasoning. We have added the two-stage results for VideoRover-RL to Appendix F.

W2: Insufficient distinction from VideoDR and VideoSearcher, with missing per-skill oracle evaluations.

Thank you for this suggestion. Unlike VideoDR, which primarily introduces a benchmark and evaluates existing models, VideoRover additionally provides task-specific training data and SFT/RL training for an 8B model. VideoRover-Bench also provides balanced evaluation across six video-length–difficulty strata and nine knowledge domains.

We acknowledge that VideoSearcher already supports temporal refinement, spatial inspection, and multi-tool reasoning. Unlike VideoSearcher, which constructs training examples by adapting existing image–text QA pairs to retrieved videos, VideoRover starts from encyclopedia knowledge, grounds candidate entities in videos, and generates questions linking verified visual evidence to external facts. VideoRover also provides an explicit visit tool, allowing the model to independently select retrieved webpages for further inspection. We have added a brief discussion of VideoDR and VideoSearcher to the Related Work section.

Providing an oracle response for every tool invocation is challenging because trajectories contain multiple rounds and tool requests depend on the evolving research state. Following your suggestion in W1, we instead construct stage-level oracle experiments using the annotations retained during VideoRover-Bench construction. Oracle visual supplies the relevant visual entity identities before inference, oracle retrieval supplies answer-relevant webpages, and oracle all supplies both. In every setting, the model retains access to all four original tools.

MethodS-ES-MS-HL-EL-ML-HAvg.
VideoRover-RL80.0064.0040.0070.0052.0042.0058.00
+ Oracle visual84.0074.0062.0080.0070.0070.0073.33
+ Oracle retrieval84.0072.0046.0078.0062.0058.0066.67
+ Oracle all96.0084.0086.0094.0086.0094.0090.00

Both individual oracles improve performance, indicating that visual entity identification and external retrieval are important bottlenecks. The larger gain from oracle visual suggests that identifying the relevant entities among video frames is particularly challenging under these settings. Providing both yields the largest improvement, demonstrating their complementary value. The remaining errors suggest that obtaining the correct entities and sources is not sufficient by itself, with evidence extraction and subsequent reasoning remaining potential limitations. We have added these experiments and analyses to Section 6.4.

Dear Reviewer iKax,

With less than 48 hours remaining in the discussion period, we would like to kindly follow up on our rebuttal. We have made every effort to respond thoroughly to your comments and would greatly appreciate any further feedback. If any concerns remain, please let us know so that we can provide additional clarification before the deadline.

Thank you again for your time and consideration. Even a brief update would be very helpful.

Sincerely,

The Authors

Dear Reviewer iKax,

As the discussion deadline approaches, we would sincerely appreciate your feedback on our responses and additional experiments. Could you please let us know whether they affect your assessment or whether any concerns remain unresolved? We would be happy to provide further clarification.

Thank you for your time and consideration. Even a brief update would be very helpful.

Sincerely,

The Authors

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

This paper introduces VideoRover, a framework for open-world video question answering that combines video cropping, image search, text search, and webpage browsing in an iterative reasoning process. The authors construct 26K supervised fine-tuning trajectories, 3K challenging RL instances, and VideoRover-Bench, a 300-question benchmark stratified by video duration and difficulty. Based on Qwen3-VL-8B, the model is trained through SFT followed by GSPO. Experiments show substantial improvements over the same-backbone tool-use baseline and gains across several general video benchmarks.

Critical strengths
  1. Well-motivated task. Connecting temporal video grounding with external knowledge retrieval addresses a practical limitation of video-only reasoning.
  2. Comprehensive data pipeline. The construction process provides multi-step tool-use supervision, supported by cross-model verification and human review.
  3. Strong empirical improvements. SFT and RL raise the aggregate accuracy from 31.86% for the same-backbone agentic baseline to 57.71%. Tool ablations support the complementary value of video inspection and external retrieval.
  4. Broad evaluation. Results across an existing benchmark, the proposed benchmark, and general video benchmarks provide evidence of transfer beyond the training setting.
Critical weaknesses
  1. Coordination benefits are not isolated: Tool ablations do not establish the advantage of bidirectional interaction over fixed pipelines under matched budgets.
  2. Evaluation validity needs clarification: Training–test separation is insufficiently documented, while the small benchmark and lack of uncertainty estimates limit confidence in modest gains.
  3. Evidence reliability remains uncertain: Qualitative examples contain unresolved inconsistencies, suggesting that webpage browsing does not necessarily ensure faithful verification.
Questions for authors

See Critical Weaknesses

6 comments
AuthorsOpenReview ↗

Response to Reviewer awmf

We sincerely thank the reviewer for the constructive feedback. We hope our responses below address your concerns.

W1: Unclear benefits of bidirectional interaction over fixed pipelines under matched budgets.

Thank you for this suggestion. To more directly assess the benefit of flexible tool coordination, we compare VideoRover-RL with a fixed-order variant using the same model, tools, and maximum number of interaction turns. This variant enforces crop_video → image_search → text_search → visit within each round. After each round, the model can either answer or start another round, while remaining free to select tool arguments.

MethodVideoDRS-ES-MS-HL-EL-ML-HAvg.
VideoRover-RL56.0080.0064.0040.0070.0052.0042.0057.71
+ Fixed order48.0074.0058.0034.0064.0044.0034.0050.86

Enforcing this fixed order reduces overall accuracy, with decreases on VideoDR and all six VideoRover-Bench strata. These results support the practical benefit of flexible scheduling over this fixed iterative pipeline, allowing the model to revisit video evidence, refine searches, or continue webpage verification as needed rather than complete a prescribed tool cycle. We have added this experiment to Appendix F.

W2: Unclear train–test separation, small benchmark size, and missing uncertainty estimates.

Thank you for raising this concern. We apologize that the data split and verification protocol were insufficiently documented. Training and benchmark data are strictly separated, with no benchmark video or associated QA instance included in either SFT or RL training.

Each benchmark example undergoes automated grounding and secondary verification, followed by manual checks of question clarity, answer correctness, localized video evidence, external evidence, and their joint dependency. This process is particularly labor-intensive because annotators must locate relevant video segments and validate the complete video-to-web evidence chain. Existing benchmarks are similarly limited in scale, including VideoDR (originally 100 examples, later expanded to 500), VSQA (278), and VideoDR-Bench (200). The VSQA and VideoDR-Bench datasets have not been publicly released, so we have not evaluated on them. VideoRover-Bench contains 300 manually verified examples, balanced across two video-length groups and three difficulty levels, with 50 examples per subset and coverage of nine knowledge domains. Difficulty is manually assessed, with trajectory interaction counts used as an auxiliary signal.

To quantify evaluation variability, we conducted two additional independent runs under the same settings. Across the three runs, the overall score is 57.76% ± 1.08 percentage points (mean ± sample standard deviation). These results indicate that VideoRover-RL’s overall performance is reasonably consistent across repeated evaluations rather than dependent on a single run.

W3: Inconsistent qualitative evidence raises concerns about verification faithfulness.

Thank you for raising this concern. We acknowledge that some qualitative traces contain inaccurate or conflicting intermediate evidence. The benchmark answers and supporting evidence were manually verified independently of model-generated trajectories, and we also saved the corresponding evidence pages. We can therefore be reasonably confident in the correctness of the final answers.

However, the model's intermediate reasoning may not be fully reliable. Web content can be noisy or inconsistent, and related work on open-web deep research also discusses challenges involving noisy search results and evidence attribution [1,2]. The visit tool provides access to source content, but does not guarantee its reliability. In addition, limitations in the model's reasoning abilities can lead to errors in extracting, attributing, and reconciling evidence. These factors can produce inaccurate intermediate claims even when the final answer is correct.

The Limitations section already acknowledges our model's dependence on the availability and quality of external search results, highlighting evidence reliability as an ongoing challenge.

Reference

[1] Huang, Wenxuan, et al. "Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models." ICML 2026.

[2] Du, Mingxuan, et al. "Deepresearch bench: A comprehensive benchmark for deep research agents." ICLR 2026.

Dear Reviewer awmf,

With less than 48 hours remaining in the discussion period, we would like to kindly follow up on our rebuttal. We have made every effort to respond thoroughly to your comments and would greatly appreciate any further feedback. If any concerns remain, please let us know so that we can provide additional clarification before the deadline.

Thank you again for your time and consideration. Even a brief update would be very helpful.

Sincerely,

The Authors

Dear Reviewer awmf,

As the discussion deadline approaches, we would sincerely appreciate your feedback on our responses and additional experiments. Could you please let us know whether they affect your assessment or whether any concerns remain unresolved? We would be happy to provide further clarification.

Thank you for your time and consideration. Even a brief update would be very helpful.

Sincerely,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Follow-up Comment to Reviewers

Dear Reviewers,

Thank you again for your time and constructive feedback. In response to your comments, we have provided the following experiments and clarifications:

  • Reviewer awmf: Fixed-order tool-use comparison; repeated evaluations with uncertainty estimates; clarification of train–test separation and benchmark verification; discussion of evidence faithfulness.
  • Reviewer iKax: Two-stage inference and oracle-entity evaluations; visual and retrieval bottleneck analyses; clarification of differences from VideoDR and VideoSearcher.
  • Reviewer rUvC: Candidate-pool sensitivity analysis; manual quality-audit statistics; smaller evidence-extractor ablation; resolution-sensitivity experiments; clarification of screening procedures and collection biases.
  • Reviewer Ncqi: Latency and inference-cost analyses; question-only web-search and direct-answer evaluations; comparisons with Video-DeepResearch, VideoZoomer, and Video-o3; clarification of training effects, evaluation settings, and evidence reliability.
  • Reviewer aWSx: Clarification of contributions in data construction and training; comparison of fixed-order and flexible tool use; trajectory analysis across VideoRover, VideoSearcher, and Video-DeepResearch.

As the discussion period is now more than halfway through, we would greatly appreciate your feedback on whether these responses address your concerns and affect your overall assessment. If any points remain unresolved, please let us know so that we can address them during the remaining discussion period.

Thank you for your continued engagement. Even a brief update would be very helpful.

Sincerely,

The Authors

AuthorsOpenReview ↗

Summary of the Rebuttal and Discussion

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for their time and all reviewers for their constructive feedback. We promptly conducted additional experiments and prepared detailed responses, making every effort to engage actively in the discussion. We are encouraged by the positive follow-up assessments from Reviewers awmf, Ncqi, and aWSx.

As of this comment, we have not received follow-up feedback from Reviewers iKax and rUvC. For completeness, we summarize our responses to their concerns below:

  • Reviewer iKax: We conducted two-stage and oracle-entity evaluations, with integrated VideoRover outperforming two-stage inference (57.71% vs. 43.43%). We added visual and retrieval oracle analyses to identify bottlenecks, and clarified differences from VideoDR and VideoSearcher in benchmark design, data construction, and training.
  • Reviewer rUvC: We provided candidate-pool sensitivity analysis, a manual audit of 614 candidate examples, and an 8B evidence-extractor ablation showing comparable performance. We also qualified the retrieval-dependence claim, acknowledged collection biases, and evaluated resolution sensitivity for OCR and fine-grained perception.

We believe these responses provide substantive evidence addressing the concerns raised, and key results and clarifications have been incorporated into the revised manuscript. We respectfully ask that the additional evidence and full discussion record be considered alongside the original reviews in the final assessment.

Thank you all for your time and consideration.

Sincerely,

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.