acceptodds
Under review as a conference paper at ICLR 2027

VideoGAIA: A Benchmark for General AI Assistants on Agentic Video Understanding

Abstract

Video understanding is a fundamental task for evaluating the capabilities of multimodal large language models (MLLMs). However, existing leading models have already achieved approximately 90% accuracy on the Video-MME leaderboard, suggesting that conventional single-turn video understanding tasks are becoming increasingly saturated and insufficient for assessing the intelligence of advanced MLLMs. Towards this end, we introduce VideoGAIA, an agentic video understanding benchmark for general artificial intelligence (AI) assistants. Moving beyond one-shot video question answering, VideoGAIA formulates video understanding as a multi-turn, tool-augmented interaction process, where models must iteratively perceive videos, invoke external tools, gather complementary information, and integrate multimodal evidence across turns. VideoGAIA contains 271 model-human co-designed tasks covering diverse and complex real-world scenarios. Each video-question-answer instance is independently verified by three human experts to ensure both correctness and appropriate difficulty. All evaluated MLLMs, including frontier models such as GPT-5.5 and Kimi-K3, achieve less than 60% accuracy on VideoGAIA, highlighting its value as a high-quality and timely benchmark for evaluating next-generation MLLMs. We hope that VideoGAIA will facilitate the transition from conventional video understanding toward agentic video understanding.

Primary area:
datasets and benchmarks
Keywords:
Video Understanding, Video Agent

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_gCKXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VideoGAIA introduces 271 tasks across six domains to evaluate assistants that combine video evidence with external information retrieval. Tasks are generated from visual clues, filtered using models, and independently verified by human annotators. Twenty multimodal models are evaluated using a common agent loop with web search, webpage reading, and video reinspection. The highest reported accuracy is 58.30%. The paper also analyzes difficulty, tool use, and failure trajectories.

Critical strengths
  1. The tasks connect visual recognition with information seeking in a concrete way (§§3.1–3.2). Models need to identify relevant clues in the video and use them to retrieve an external fact. The modality and tool-necessity filters help support this task design.

  2. The human verification is a strong part of the construction process. Three annotators independently solve each candidate task, and tasks approved by two annotators receive further review (§3.2). The evaluation of 20 models with common prompts, tools, and interaction limits also provides a useful reference (§4.1).

  3. Figure 8 is a helpful example of what makes the tasks difficult. Recognizing Honda branding is insufficient to identify the engine displacement; the successful trajectory considers candidate bike models and checks a more specific visual cue. The example makes the interaction between visual recognition and external verification easy to understand.

Critical weaknesses
  1. I would like a clearer picture of the range of tasks produced by the construction process (§§3.1–3.3). The pipeline centers on identifying visual clues and retrieving related facts. This is a useful task family, but the six topical categories do not tell us how varied the actual problem-solving requirements are. A clearer account of the task mix would help readers understand what VideoGAIA primarily measures.

  2. Since every model already receives 20 uniformly sampled frames, it remains unclear how much performance depends on revisiting the video rather than identifying clues from the initial frames and searching the web (§4.1). Figure 8 illustrates the value of reinspection in one case, but all leaderboard results use the full agent loop. A comparison using only the initial frames and web tools would help clarify how important video reinspection is across the benchmark.

  3. A smaller reporting concern is that scores are averaged over three runs, but the variation is not reported, making small gaps in Table 1 difficult to interpret. The grading rules are provided in Appendix B.2; it would also help to know whether the judge’s decisions on model answers were checked against human judgments.

Questions for authors
  1. Could you describe the main evidence or reasoning requirements represented in the final task set, beyond the six topical categories?

  2. For one or two representative models, could you compare the full agent loop with a condition that retains the initial 20 frames and web tools but disables video reinspection?

  3. Could you report the variation across the three runs for the leading models and describe any manual checks of the automatic grader’s decisions?

5 comments
AuthorsOpenReview ↗

Response to Reviewer gCKX [Part 1]

We sincerely thank you for your time and constructive suggestions. In the following response, we try to address your concerns point by point.

W1 & Q1. Task Mix beyond the Six Topical Categories

W1. I would like a clearer picture of the range of tasks produced by the construction process (§§3.1–3.3). The pipeline centers on identifying visual clues and retrieving related facts. This is a useful task family, but the six topical categories do not tell us how varied the actual problem-solving requirements are. A clearer account of the task mix would help readers understand what VideoGAIA primarily measures.

Q1. Could you describe the main evidence or reasoning requirements represented in the final task set, beyond the six topical categories?

Thank you for your comment and question. We describe the evidence and reasoning requirements of all 271 tasks as follows.

Visual Evidence (a Task Can Involve Several)CountProportion
Objects, Products, Dishes, and Graphics12546.1%
On-Screen Text, Logos, Signs, and Emblems9334.3%
Animal and Plant Species or Natural Materials5721.0%
Places, Landmarks, and Scenes4416.2%
Actions and Events4115.1%
Specific People83.0%
Information to RetrieveCountProportion
Numeric Quantities and Specifications7628.0%
Proper Names of People, Organizations, Places, Products, or Works6022.1%
Common Terms, Categories, and Properties5219.2%
Years and Dates5018.5%
Scientific Names3312.2%

The topical categories reflect only part of this variation, since text, logo, and object evidence appears in all six categories. In addition, 72 questions (26.6%) refer to a specific moment or the order of events in the clip. We will add this breakdown to §3.3.

W2 & Q2. Contribution of Video Reinspection

W2. Since every model already receives 20 uniformly sampled frames, it remains unclear how much performance depends on revisiting the video rather than identifying clues from the initial frames and searching the web (§4.1). Figure 8 illustrates the value of reinspection in one case, but all leaderboard results use the full agent loop. A comparison using only the initial frames and web tools would help clarify how important video reinspection is across the benchmark.

Q2. For one or two representative models, could you compare the full agent loop with a condition that retains the initial 20 frames and web tools but disables video reinspection?

Thank you for your comment and question. We compare the full agent loop with a variant that keeps the initial 20 frames and web tools but disables video reinspection, as follows.

ModelCultureGeographyDaily LifeHistoryTechnologyEconomicsOverall
Qwen3.8-Max44.8342.8662.3050.0051.2850.0050.18
Qwen3.8-Max w/o Video Reinspection37.9333.3354.1045.4546.1550.0043.17
Gemini-3.1-Pro41.3849.2157.3854.5553.8550.0050.92
Gemini-3.1-Pro w/o Video Reinspection39.6642.8654.1045.4546.1516.6745.02

Disabling video reinspection lowers overall accuracy by 7.01 points for Qwen3.8-Max and 5.90 points for Gemini-3.1-Pro. The drop appears in five of the six categories for Qwen3.8-Max and in all six for Gemini-3.1-Pro. We will add this ablation to the revised paper.

AuthorsOpenReview ↗

Response to Reviewer gCKX [Part 2]

W3 & Q3. Run-to-Run Variance and Grader Validation

W3. A smaller reporting concern is that scores are averaged over three runs, but the variation is not reported, making small gaps in Table 1 difficult to interpret. The grading rules are provided in Appendix B.2; it would also help to know whether the judge’s decisions on model answers were checked against human judgments.

Q3. Could you report the variation across the three runs for the leading models and describe any manual checks of the automatic grader’s decisions?

Thank you for your comment and question. We report the overall accuracy (%) of three independent runs as follows.

ModelRun 1 (Table 1)Run 2Run 3Mean +/- Std
Seed2.0-Pro58.3057.9356.4657.56 +/- 0.98
Qwen3.7-Plus56.0957.5656.0956.58 +/- 0.85
Kimi-K354.6155.3555.3555.10 +/- 0.43
GPT-5.553.1456.0956.0955.10 +/- 1.70
Kimi-K2.652.7753.5152.4052.89 +/- 0.56
GPT-5.452.0352.4053.1452.52 +/- 0.56
Qwen3.5-Plus52.0349.4552.4051.29 +/- 1.61
GPT-5.250.9250.1850.9250.68 +/- 0.43
GLM-5V-Turbo50.1848.7152.4050.43 +/- 1.86
Qwen3.8-Max50.1849.8248.7149.57 +/- 0.77
Kimi-K2.549.8249.4549.4549.57 +/- 0.21
Gemini-3.1-Pro50.9249.0847.9749.32 +/- 1.49
MiMo-V2.548.7148.3450.5549.20 +/- 1.19
Qwen3.5-397B50.1848.7147.6048.83 +/- 1.30
Qwen3.6-Plus49.0847.2348.7148.34 +/- 0.98
Claude-Opus-4.647.9747.6048.7148.09 +/- 0.56
GLM-4.6V46.8644.2847.9746.37 +/- 1.89
Gemini-3.5-Flash43.9141.7044.6543.42 +/- 1.54
Claude-Opus-4.836.9038.0138.0137.64 +/- 0.64
Claude-Opus-4.734.6929.8933.9532.84 +/- 2.58

The standard deviation ranges from 0.21 to 2.58 points, and the Spearman correlation between the model rankings of any two runs is at least 0.88. Seed2.0-Pro ranks first in every run, and no model reaches 60% in any run. We will report the mean +/- std for all columns of Table 1 in the revision.

Regarding the grader, we did not conduct a separate manual check, since the judge receives the question, the prediction, and a short human-verified reference answer (3.73 tokens on average). We instead re-graded the Run 1 predictions of three leading models with two judges from other model families, as follows.

ModelGPT-5.5 JudgeDeepSeek-V4-Pro JudgeClaude-Opus-4.8 Judge
Seed2.0-Pro158/271156/271157/271
Qwen3.7-Plus152/271152/271151/271
Gemini-3.1-Pro138/271138/271138/271

The three judges differ by at most two of the 271 tasks and yield the same ranking, which supports the reliability of the LLM judge. This protocol also follows SimpleQA (Wei et al., 2024), Humanity's Last Exam (Phan et al., 2025), and BrowseComp (Wei et al., 2025), and Zheng et al. (2023) and Thakur et al. (2025) show that strong LLM judges agree well with human judgments. We will add a manual validation of the judge in future work.

References

  • Long Phan, Alice Gatti, Ziwen Han, et al. Humanity's Last Exam. arXiv preprint arXiv:2501.14249, 2025.
  • Aman Singh Thakur, Kartik Choudhary, Venkat Srinik Ramayapally, et al. Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges. In Proceedings of the Fourth Workshop on Generation, Evaluation and Metrics (GEM²), 2025.
  • Jason Wei, Nguyen Karina, Hyung Won Chung, et al. Measuring Short-Form Factuality in Large Language Models. arXiv preprint arXiv:2411.04368, 2024.
  • Jason Wei, Zhiqing Sun, Spencer Papay, et al. BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. arXiv preprint arXiv:2504.12516, 2025.
  • Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023.

Thank you again for your constructive feedback, which helped us improve the quality of this work. If our responses have addressed your concerns, we would be grateful if you could consider improving your rating. If you have any other questions or concerns that remain unresolved, we would be very happy to address them.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer gCKX,

Thank you again for your constructive feedback on our submission. In our response, we described the evidence and reasoning requirements of all 271 tasks, showed that disabling video reinspection lowers accuracy by 7.01 points for Qwen3.8-Max and 5.90 points for Gemini-3.1-Pro, and reported three-run results together with cross-judge validation.

As the discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns. If they have, we would appreciate it if you could consider improving your rating. If any questions remain, we would be very happy to address them before the deadline.

Best regards,

The Authors of Submission 1955, VideoGAIA

Reviewer_tc6kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VideoGAIA, a benchmark for evaluating agentic video understanding through multi-turn reasoning and external tool use. It comprises 271 model–human co-designed tasks, each independently verified by three human experts. All evaluated models achieve less than 60% accuracy, highlighting challenges in gathering and integrating multimodal evidence across turns.

Critical strengths
  • The paper is clearly organized. The construction pipeline and the worked example in Figure 8 make the intended evaluation setting understandable.

  • Testing 20 models with the same tools and interaction budget provides a useful comparative reference. The trajectory analyses extend the evaluation beyond aggregate accuracy by identifying failures in video perception, web retrieval, evidence reasoning, answer formulation, and agent execution.

Critical weaknesses
  • Section 3.2 requires multiple video anchors, observations from multiple frames, and at least two reasoning steps involving web search or webpage extraction. The construction filters and the example in Figure 8 support the intended design. However, the main evaluation does not establish how often successful solutions actually depend on adaptive interaction, rather than a fixed video-observation and retrieval pipeline. Multiple reasoning steps, multiple tool calls, and adaptive evidence acquisition are distinct properties; tool-use frequencies alone cannot establish the last of these.

  • Section 4.1 states that results are averaged over three independent runs, but the main results provide no run-to-run variability or confidence intervals. On a 271-task benchmark with several closely spaced scores, this makes it difficult to assess which ranking differences are robust. The evaluation description also does not report agreement between the LLM judge and human judgments of model responses, or a sensitivity check using an independent judge.

Questions for authors
  • The authors are recommended to clarify that how are the multiple-frame and retrieval-dependent reasoning requirements in Section 3.2 verified for the retained tasks. In particular, does human review verify that these dependencies are necessary, or primarily that the answer is unique and correct? Please distinguish a valid intended solution from evidence that simpler solutions are insufficient.

  • Can the authors provide controlled results comparing the adaptive agent with a fixed observation–search pipeline under comparable resource budgets? Results on models not used in tool-necessity filtering would be particularly informative. A substantial, consistent advantage would directly address my principal concern.

  • Some details need further explanation, e.g., What variation is observed across the three runs, and how stable are the reported ranking differences? Has the LLM judge been checked against human scoring of a representative sample of model responses or against an independent judge? Please report any existing validation and whether disagreements materially change the comparisons.

6 comments
AuthorsOpenReview ↗

Response to Reviewer tc6k [Part 1]

We sincerely thank you for your time and constructive suggestions. In the following response, we try to address your concerns point by point.

W1 & Q2. Adaptive Interaction versus a Fixed Observation-Search Pipeline

W1. Section 3.2 requires multiple video anchors, observations from multiple frames, and at least two reasoning steps involving web search or webpage extraction. The construction filters and the example in Figure 8 support the intended design. However, the main evaluation does not establish how often successful solutions actually depend on adaptive interaction, rather than a fixed video-observation and retrieval pipeline. Multiple reasoning steps, multiple tool calls, and adaptive evidence acquisition are distinct properties; tool-use frequencies alone cannot establish the last of these.

Q2. Can the authors provide controlled results comparing the adaptive agent with a fixed observation–search pipeline under comparable resource budgets? Results on models not used in tool-necessity filtering would be particularly informative. A substantial, consistent advantage would directly address my principal concern.

Thank you for your comment and question. We compare the adaptive agent with a fixed observation-search pipeline under the same 40-step budget, using two models that are not used in any construction or filtering stage, as follows.

ModelCultureGeographyDaily LifeHistoryTechnologyEconomicsOverall
Qwen3.8-Max (Adaptive)44.8342.8662.3050.0051.2850.0050.18
Qwen3.8-Max (Fixed)34.4838.1054.1047.7351.2850.0044.65
Gemini-3.5-Flash (Adaptive)37.9330.1647.5445.4564.1066.6743.91
Gemini-3.5-Flash (Fixed)36.2120.6339.3443.1856.4133.3337.27

The adaptive agent outperforms the fixed pipeline by 5.54 points on Qwen3.8-Max and 6.64 points on Gemini-3.5-Flash. It is never worse in any category and is better in four of the six categories for Qwen3.8-Max and in all six for Gemini-3.5-Flash. We will add this comparison to the revised paper.

Q1. Verification of the Multi-Frame and Retrieval Requirements

The authors are recommended to clarify that how are the multiple-frame and retrieval-dependent reasoning requirements in Section 3.2 verified for the retained tasks. In particular, does human review verify that these dependencies are necessary, or primarily that the answer is unique and correct? Please distinguish a valid intended solution from evidence that simpler solutions are insufficient.

Thank you for your question. The necessity of these dependencies is verified by the model-based filters in §3.2, which test whether simpler solutions fail, while human review verifies correctness and uniqueness, as follows.

StageVerified Property
Cross-Modal FilteringNeither the Video nor the Web Content Alone Suffices
Anchor AblationAt Least One Video Anchor Is Indispensable
Tool-Necessity AssessmentThree Models Fail to Answer without Tools
Human ReviewThe Answer Is Correct and Unique

The fixed-pipeline comparison in our response to W1 and Q2 further shows that simpler solutions underperform on the final benchmark. We do not ask annotators to judge necessity because, in our pilot tests, annotators, who are usually not AI experts, could not reliably judge whether a dependency is necessary and often misjudged it. We therefore restrict human review to correctness and uniqueness. We believe this does not weaken the contribution of this work.

AuthorsOpenReview ↗

Response to Reviewer tc6k [Part 2]

W2 & Q3. Run-to-Run Variance and Judge Validation

W2. Section 4.1 states that results are averaged over three independent runs, but the main results provide no run-to-run variability or confidence intervals. On a 271-task benchmark with several closely spaced scores, this makes it difficult to assess which ranking differences are robust. The evaluation description also does not report agreement between the LLM judge and human judgments of model responses, or a sensitivity check using an independent judge.

Q3. Some details need further explanation, e.g., What variation is observed across the three runs, and how stable are the reported ranking differences? Has the LLM judge been checked against human scoring of a representative sample of model responses or against an independent judge? Please report any existing validation and whether disagreements materially change the comparisons.

Thank you for your comment and question. We report the overall accuracy (%) of three independent runs as follows.

ModelRun 1 (Table 1)Run 2Run 3Mean +/- Std
Seed2.0-Pro58.3057.9356.4657.56 +/- 0.98
Qwen3.7-Plus56.0957.5656.0956.58 +/- 0.85
Kimi-K354.6155.3555.3555.10 +/- 0.43
GPT-5.553.1456.0956.0955.10 +/- 1.70
Kimi-K2.652.7753.5152.4052.89 +/- 0.56
GPT-5.452.0352.4053.1452.52 +/- 0.56
Qwen3.5-Plus52.0349.4552.4051.29 +/- 1.61
GPT-5.250.9250.1850.9250.68 +/- 0.43
GLM-5V-Turbo50.1848.7152.4050.43 +/- 1.86
Qwen3.8-Max50.1849.8248.7149.57 +/- 0.77
Kimi-K2.549.8249.4549.4549.57 +/- 0.21
Gemini-3.1-Pro50.9249.0847.9749.32 +/- 1.49
MiMo-V2.548.7148.3450.5549.20 +/- 1.19
Qwen3.5-397B50.1848.7147.6048.83 +/- 1.30
Qwen3.6-Plus49.0847.2348.7148.34 +/- 0.98
Claude-Opus-4.647.9747.6048.7148.09 +/- 0.56
GLM-4.6V46.8644.2847.9746.37 +/- 1.89
Gemini-3.5-Flash43.9141.7044.6543.42 +/- 1.54
Claude-Opus-4.836.9038.0138.0137.64 +/- 0.64
Claude-Opus-4.734.6929.8933.9532.84 +/- 2.58

The standard deviation ranges from 0.21 to 2.58 points, and the Spearman correlation between the model rankings of any two runs is at least 0.88. Seed2.0-Pro ranks first in every run, and no model reaches 60% in any run. We will report the mean +/- std for all columns of Table 1 in the revision.

Regarding the judge, we re-graded the Run 1 predictions of three leading models with two independent judges from other model families, as follows.

ModelGPT-5.5 JudgeDeepSeek-V4-Pro JudgeClaude-Opus-4.8 Judge
Seed2.0-Pro158/271156/271157/271
Qwen3.7-Plus152/271152/271151/271
Gemini-3.1-Pro138/271138/271138/271

The three judges differ by at most two of the 271 tasks and yield the same ranking, so judge disagreements do not materially change the comparisons. We did not conduct a separate manual check, since each judge receives the question, the prediction, and a short human-verified reference answer (3.73 tokens on average). This protocol also follows SimpleQA (Wei et al., 2024), Humanity's Last Exam (Phan et al., 2025), and BrowseComp (Wei et al., 2025), and Zheng et al. (2023) and Thakur et al. (2025) show that strong LLM judges agree well with human judgments. We will add a manual validation of the judge in future work.

References

  • Long Phan, Alice Gatti, Ziwen Han, et al. Humanity's Last Exam. arXiv preprint arXiv:2501.14249, 2025.
  • Aman Singh Thakur, Kartik Choudhary, Venkat Srinik Ramayapally, et al. Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges. In Proceedings of the Fourth Workshop on Generation, Evaluation and Metrics (GEM²), 2025.
  • Jason Wei, Nguyen Karina, Hyung Won Chung, et al. Measuring Short-Form Factuality in Large Language Models. arXiv preprint arXiv:2411.04368, 2024.
  • Jason Wei, Zhiqing Sun, Spencer Papay, et al. BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. arXiv preprint arXiv:2504.12516, 2025.
  • Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023.

Thank you again for your thoughtful feedback, which helped us improve the quality of this work. If our responses have addressed your concerns, we would be grateful if you could consider improving your rating. If you have any other questions or concerns that remain unresolved, we would be very happy to address them.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer tc6k,

Thank you again for your thoughtful feedback on our submission. In our response, we showed that under the same 40-step budget, the adaptive agent outperforms a fixed observation-search pipeline by 5.54 points on Qwen3.8-Max and 6.64 points on Gemini-3.5-Flash, two models not used in any construction or filtering stage, and is never worse in any category. We also clarified that the model-based filters verify the necessity of the multi-frame and retrieval dependencies, and reported three-run results and cross-judge validation.

As the discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns. If they have, we would appreciate it if you could consider improving your rating. If any questions remain, we would be very happy to address them before the deadline.

Best regards,

The Authors of Submission 1955, VideoGAIA

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

The paper introduces VideoGAIA, a benchmark of 271 tasks designed to evaluate agentic video understanding beyond conventional single-turn video question answering. Models are initially provided with 20 uniformly sampled video frames and must answer open-ended questions through iterative reasoning and tool use, including web search, webpage retrieval, and targeted video reinspection. The benchmark is constructed through a multi-stage model-assisted pipeline involving video captioning, visual anchor extraction, question and answer generation, evidence retrieval, and quality filtering. Candidate tasks undergo cross-modal and tool-necessity checks, followed by independent verification by three human annotators, yielding 271 tasks across six domains and two difficulty levels. The authors evaluate 20 frontier multimodal models under a unified ReAct-style agent framework with a maximum of 40 steps, using an LLM-based answer judge and averaging results over three runs. The strongest model, Seed2.0-Pro, achieves 58.3% accuracy, with all evaluated models remaining below 60%. Additional analyses examine difficulty-dependent performance, tool-use patterns, error categories, computational cost, and representative agent trajectories.

Critical strengths

Strengths

  1. Timely and well-motivated framing. The paper convincingly argues that conventional single-turn video question answering is approaching saturation and motivates a shift toward evaluating video understanding in tool-augmented, agentic settings. Figure 1 effectively illustrates this distinction.

  2. Rigorous multi-stage benchmark construction. The construction pipeline incorporates several meaningful quality-control mechanisms, including video-anchor ablation, unimodal-solvability filtering, tool-necessity checks, and independent verification by three human annotators with additional adjudication when needed. Requiring multiple video anchors and external reasoning steps provides a reasonable operational definition of agentic video understanding.

  3. Broad and standardized model evaluation. Evaluating 20 frontier multimodal models under a unified framework, with consistent prompts, tools, and interaction budgets, provides a valuable comparative reference for the community. The inclusion of the agent and evaluation prompts in Appendix B further supports reproducibility.

  4. Informative analyses beyond leaderboard performance. The paper examines difficulty-dependent performance, tool-use patterns, computational costs, and dominant failure modes. In particular, the contrast between tool-call volume and accuracy, the prevalence of video perception errors, and the substantial L1/L2 performance gaps provide useful insights into current model limitations. Figure 8 offers a clear illustration of hypothesis maintenance and evidence verification versus premature commitment.

  5. Open-ended evaluation format. Using open-ended answers with semantic evaluation, rather than multiple-choice questions, reduces reliance on option-elimination strategies and better reflects practical information-seeking tasks.

Critical weaknesses
  • Potential selection bias from model-assisted filtering. The benchmark construction removes tasks that Gemini-3.1-Pro, GPT-5.5, or Claude-Opus-4.7 can solve without tools, while additional cross-modal filtering relies on Gemini-3.1-Pro and GPT-5.5 (§3.2). This may systematically disadvantage these models and related model families in the final evaluation. In particular, the poor performance of Claude-Opus-4.7/4.8 cannot be confidently attributed to non-monotonic scaling without accounting for this selection process. The authors should quantify rejection rates by filtering model and discuss the implications for leaderboard comparisons.

  • Insufficient baselines to isolate agentic capabilities. The paper does not report controlled comparisons against no-tool, text-only with web search, or fixed-frame baselines. These are important for establishing whether performance genuinely depends on adaptive video inspection and multimodal evidence integration rather than visually grounded web retrieval. The cross-modal filtering procedure partially addresses this concern, but does not replace systematic ablations on the final benchmark. An oracle visual-anchor baseline would also help separate perception failures from retrieval and reasoning failures.

  • Limited statistical reliability of model comparisons. The benchmark contains only 271 tasks, with particularly small subsets for L1 (62) and Economics (6). Although results are averaged over three runs, no variance estimates, confidence intervals, or significance tests are reported. Consequently, the reliability of close model rankings and family-level comparisons remains unclear. Reporting per-run variability and uncertainty estimates would strengthen the conclusions.

  • Insufficient validation of annotation and evaluation quality. GPT-5.5 serves as both a benchmark contestant and the primary answer judge, without a reported human–judge agreement study. Similarly, although each candidate receives three human evaluations, the manuscript provides limited information about annotator qualifications, agreement, and verification procedures. The reported annotation effort (over 100 hours for 847 candidates, each reviewed by three annotators) also warrants clarification. Finally, the criteria used to calibrate L1/L2 difficulty and assign dominant error categories are insufficiently specified.

  • Reproducibility concerns with web-dependent ground truth and evaluation protocols. Tasks depend on external webpages and search results that may change over time, yet the manuscript does not describe a frozen evidence corpus or retrieval timestamps. This creates potential ambiguity in reference answers and limits reproducibility across evaluations.

  • Limited differentiation from related benchmarks and incomplete data documentation. Although VideoDR, Video-BrowseComp, and LongVidSearch are discussed, the paper provides little systematic comparison of their task requirements, evaluation protocols, and coverage. A concise benchmark comparison would help establish VideoGAIA's distinct contribution. Additionally, the licensing, redistribution, and maintenance policies for videos collected from public platforms remain unclear, while the discussion of potential privacy and societal impacts is limited.

Questions for authors

See Critical Weaknesses

  • For each of the three tool-necessity models and the two cross-modal models, how many candidate tasks did that model alone eliminate? Can you report accuracy of all 20 models on the 847-task pre-filter pool (or a subset) to show how much of the Claude/Gemini deficit is induced by filtering?

  • Will you release a frozen snapshot of the retrieved evidence pages and search results so that results remain reproducible as the web changes? How will you handle answers that become outdated (e.g., ownership, prices, latest models)?

6 comments
AuthorsOpenReview ↗

Response to Reviewer XYxc [Part 1]

We sincerely thank you for your time and constructive suggestions. In the following response, we try to address your concerns point by point.

W1 & Q1. Selection Bias from Model-Assisted Filtering

W1. Potential selection bias from model-assisted filtering. The benchmark construction removes tasks that Gemini-3.1-Pro, GPT-5.5, or Claude-Opus-4.7 can solve without tools, while additional cross-modal filtering relies on Gemini-3.1-Pro and GPT-5.5 (§3.2). This may systematically disadvantage these models and related model families in the final evaluation. In particular, the poor performance of Claude-Opus-4.7/4.8 cannot be confidently attributed to non-monotonic scaling without accounting for this selection process. The authors should quantify rejection rates by filtering model and discuss the implications for leaderboard comparisons.

Q1. For each of the three tool-necessity models and the two cross-modal models, how many candidate tasks did that model alone eliminate? Can you report accuracy of all 20 models on the 847-task pre-filter pool (or a subset) to show how much of the Claude/Gemini deficit is induced by filtering?

Thank you for your comment and question. The 847 candidates are the tasks remaining after model-based filtering and before human verification (§3.2). The numbers of candidates removed by each filtering model alone are as follows.

Filtering StageModelCandidates Removed by This Model Alone
Cross-Modal FilteringGemini-3.1-Pro (Video Only)2.5K
Cross-Modal FilteringGPT-5.5 (Web Content Only)1.5K
Tool-Necessity AssessmentGemini-3.1-Pro1.1K
Tool-Necessity AssessmentGPT-5.51.3K
Tool-Necessity AssessmentClaude-Opus-4.70.7K

We also evaluate all 20 models on the 847 candidates, and the accuracy (%) is as follows. These candidates have not passed human verification, and the 576 tasks later removed by annotators mostly have incorrect or non-unique answers, so this accuracy should be interpreted with caution.

Model847 CandidatesFinal 271 Tasks (Mean)
Seed2.0-Pro78.2857.56
Qwen3.8-Max72.2649.57
Kimi-K371.7855.10
Qwen3.7-Plus70.9656.58
GPT-5.570.7255.10
Kimi-K2.669.7852.89
Gemini-3.1-Pro69.4249.32
GPT-5.468.8352.52
Kimi-K2.568.2449.57
Qwen3.6-Plus67.5348.34
MiMo-V2.566.9449.20
Gemini-3.5-Flash66.8243.42
GPT-5.266.7150.68
GLM-5V-Turbo66.3550.43
Qwen3.5-Plus66.1251.29
Qwen3.5-397B65.2948.83
GLM-4.6V64.2346.37
Claude-Opus-4.661.9848.09
Claude-Opus-4.761.5132.84
Claude-Opus-4.859.2737.64

Claude-Opus-4.7 removes the fewest candidates alone, yet the three Claude-Opus models are the three lowest on the 847 candidates, so their deficit is unlikely to come from the filtering. GPT-5.5 removes 2.8K candidates alone and still ranks fifth on the 847 candidates and ties for third on the final benchmark, which shows no consistent penalty for filtering models.

AuthorsOpenReview ↗

Response to Reviewer XYxc [Part 2]

W2. Baselines That Isolate Agentic Capabilities

Insufficient baselines to isolate agentic capabilities. The paper does not report controlled comparisons against no-tool, text-only with web search, or fixed-frame baselines. These are important for establishing whether performance genuinely depends on adaptive video inspection and multimodal evidence integration rather than visually grounded web retrieval. The cross-modal filtering procedure partially addresses this concern, but does not replace systematic ablations on the final benchmark. An oracle visual-anchor baseline would also help separate perception failures from retrieval and reasoning failures.

Thank you for your comment. We report the accuracy (%) on the final benchmark for the full agent loop, a fixed-frame baseline without video reinspection, a variant without web search and page visit, and a no-tool baseline, as follows.

ModelCultureGeographyDaily LifeHistoryTechnologyEconomicsOverall
Qwen3.8-Max44.8342.8662.3050.0051.2850.0050.18
Qwen3.8-Max w/o Video Reinspection37.9333.3354.1045.4546.1550.0043.17
Qwen3.8-Max w/o Web Search and Page Visit36.2134.9255.7445.4543.5933.3342.80
Qwen3.8-Max w/o All Tools34.4834.9254.1031.8217.9516.6735.79
Gemini-3.1-Pro41.3849.2157.3854.5553.8550.0050.92
Gemini-3.1-Pro w/o Video Reinspection39.6642.8654.1045.4546.1516.6745.02
Gemini-3.1-Pro w/o Web Search and Page Visit32.7626.9845.9043.1838.4633.3336.90
Gemini-3.1-Pro w/o All Tools25.8620.6339.3427.2728.2116.6728.04

Removing any tool group lowers accuracy for both models, and removing all tools costs 14.39 points for Qwen3.8-Max and 22.88 points for Gemini-3.1-Pro. The fixed-frame baseline remains 7.01 points below the full agent loop for Qwen3.8-Max and 5.90 points for Gemini-3.1-Pro, so performance depends on adaptive video inspection beyond visually grounded web retrieval. We will add these baselines to the revised paper.

W3. Statistical Reliability of Model Comparisons

Limited statistical reliability of model comparisons. The benchmark contains only 271 tasks, with particularly small subsets for L1 (62) and Economics (6). Although results are averaged over three runs, no variance estimates, confidence intervals, or significance tests are reported. Consequently, the reliability of close model rankings and family-level comparisons remains unclear. Reporting per-run variability and uncertainty estimates would strengthen the conclusions.

Thank you for your comment. We report the overall accuracy (%) of three independent runs as follows.

ModelRun 1 (Table 1)Run 2Run 3Mean +/- Std
Seed2.0-Pro58.3057.9356.4657.56 +/- 0.98
Qwen3.7-Plus56.0957.5656.0956.58 +/- 0.85
Kimi-K354.6155.3555.3555.10 +/- 0.43
GPT-5.553.1456.0956.0955.10 +/- 1.70
Kimi-K2.652.7753.5152.4052.89 +/- 0.56
GPT-5.452.0352.4053.1452.52 +/- 0.56
Qwen3.5-Plus52.0349.4552.4051.29 +/- 1.61
GPT-5.250.9250.1850.9250.68 +/- 0.43
GLM-5V-Turbo50.1848.7152.4050.43 +/- 1.86
Qwen3.8-Max50.1849.8248.7149.57 +/- 0.77
Kimi-K2.549.8249.4549.4549.57 +/- 0.21
Gemini-3.1-Pro50.9249.0847.9749.32 +/- 1.49
MiMo-V2.548.7148.3450.5549.20 +/- 1.19
Qwen3.5-397B50.1848.7147.6048.83 +/- 1.30
Qwen3.6-Plus49.0847.2348.7148.34 +/- 0.98
Claude-Opus-4.647.9747.6048.7148.09 +/- 0.56
GLM-4.6V46.8644.2847.9746.37 +/- 1.89
Gemini-3.5-Flash43.9141.7044.6543.42 +/- 1.54
Claude-Opus-4.836.9038.0138.0137.64 +/- 0.64
Claude-Opus-4.734.6929.8933.9532.84 +/- 2.58

The standard deviation ranges from 0.21 to 2.58 points, and the Spearman correlation between the model rankings of any two runs is at least 0.88. The family-level trends in §4.2 also hold under the three-run means, as accuracy increases from Kimi-K2.5 to Kimi-K2.6 to Kimi-K3 and from GLM-4.6V to GLM-5V-Turbo, while Claude-Opus-4.7 and Claude-Opus-4.8 remain below Claude-Opus-4.6. We will report the mean +/- std for all columns of Table 1 in the revision.

AuthorsOpenReview ↗

Response to Reviewer XYxc [Part 3]

W4. Validation of Annotation and Evaluation Quality

Insufficient validation of annotation and evaluation quality. GPT-5.5 serves as both a benchmark contestant and the primary answer judge, without a reported human–judge agreement study. Similarly, although each candidate receives three human evaluations, the manuscript provides limited information about annotator qualifications, agreement, and verification procedures. The reported annotation effort (over 100 hours for 847 candidates, each reviewed by three annotators) also warrants clarification. Finally, the criteria used to calibrate L1/L2 difficulty and assign dominant error categories are insufficiently specified.

Thank you for your comment. We address each point as follows.

LLM Judge. We re-graded the Run 1 predictions of three leading models with two judges from other model families, and the number of correct answers is as follows.

ModelGPT-5.5 JudgeDeepSeek-V4-Pro JudgeClaude-Opus-4.8 Judge
Seed2.0-Pro158/271156/271157/271
Qwen3.7-Plus152/271152/271151/271
Gemini-3.1-Pro138/271138/271138/271

The three judges differ by at most two of the 271 tasks and yield the same ranking. We will add a manual validation of the judge in future work.

Annotators. The annotators come from an external data annotation company and hold bachelor's degrees, but they are not domain experts. We trained them on a pilot set of 10 sample tasks, and the review checks whether each answer is correct and unique.

Annotation Effort. The reported 100+ hours refer to the elapsed working time of the annotation project, counted in 8-hour workdays, during which the 15 annotators worked in parallel.

Difficulty and Error Labels. Both the L1/L2 difficulty level and the dominant error category are assigned by GPT-5.5 through its API.

W5 & Q2. Reproducibility of Web-Dependent Answers

W5. Reproducibility concerns with web-dependent ground truth and evaluation protocols. Tasks depend on external webpages and search results that may change over time, yet the manuscript does not describe a frozen evidence corpus or retrieval timestamps. This creates potential ambiguity in reference answers and limits reproducibility across evaluations.

Q2. Will you release a frozen snapshot of the retrieved evidence pages and search results so that results remain reproducible as the web changes? How will you handle answers that become outdated (e.g., ownership, prices, latest models)?

Thank you for your comment and question. After the anonymity period, we will release the model trajectories, including the retrieved webpage results, so that the evidence behind each evaluation can be inspected and reused. Each reference answer is also supported by the evidence retrieved during construction (§3.2). Answers that may change over time are not a weakness of the benchmark but an inherent property of deep research tasks on the open web.

Reviewer_qJL8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces VideoGAIA, a benchmark of 271 human-verified questions for evaluating how video agents use tools such as thinking with videos, web search, and webpage reading to understand videos. The dataset construction involves frame captioning, anchor extraction, and query generation and checking with closed-source models. Candidate questions are first filtered based on the number of reasoning steps, whether tools are needed, and whether answering requires combining different information sources. Human reviewers then carry out the final screening. All 20 evaluated models score below 60% under the same ReAct harness. The error analysis identifies misinterpreting visual clues and failing to retrieve useful information as common causes of failure.

Critical strengths
  • The benchmark offers a different perspective on video understanding and provides a representative setting for evaluating agentic video understanding.
  • Combining model-based filtering with human review makes the benchmark challenging and helps ensure data quality.
  • The main experiments are thorough, the results are convincing, and the additional analyses are useful.
Critical weaknesses
  • The benchmark contains only 271 questions, with just 6 in Economics. The paper does not provide enough information about how much the scores vary across runs. Reporting the results of all three independent runs, along with their standard deviations, would help.
  • The authors do use single-modality and no-tool tests to filter candidate questions during dataset construction, which shows that they have considered whether tools are necessary. However, they do not systematically report results on the final test set with external tools disabled, or with web search, webpage reading, or video reinspection individually removed. This makes it difficult to measure how much each tool-use contributes to the final results, or to tell whether differences between models mainly come from their initial understanding of the video or from their later search and evidence integration.
Questions for authors

The tool-call format includes a memory field for brief notes. How are these notes used in later steps, and what role do they play in the agentic loop?

6 comments
AuthorsOpenReview ↗

Response to Reviewer qJL8 [Part 1]

We sincerely thank you for your time and constructive suggestions. In the following response, we try to address your concerns point by point.

W1. Benchmark Size and Run-to-Run Variance

The benchmark contains only 271 questions, with just 6 in Economics. The paper does not provide enough information about how much the scores vary across runs. Reporting the results of all three independent runs, along with their standard deviations, would help.

Thank you for your comment. We report the overall accuracy (%) of three independent runs as follows.

ModelRun 1 (Table 1)Run 2Run 3Mean +/- Std
Seed2.0-Pro58.3057.9356.4657.56 +/- 0.98
Qwen3.7-Plus56.0957.5656.0956.58 +/- 0.85
Kimi-K354.6155.3555.3555.10 +/- 0.43
GPT-5.553.1456.0956.0955.10 +/- 1.70
Kimi-K2.652.7753.5152.4052.89 +/- 0.56
GPT-5.452.0352.4053.1452.52 +/- 0.56
Qwen3.5-Plus52.0349.4552.4051.29 +/- 1.61
GPT-5.250.9250.1850.9250.68 +/- 0.43
GLM-5V-Turbo50.1848.7152.4050.43 +/- 1.86
Qwen3.8-Max50.1849.8248.7149.57 +/- 0.77
Kimi-K2.549.8249.4549.4549.57 +/- 0.21
Gemini-3.1-Pro50.9249.0847.9749.32 +/- 1.49
MiMo-V2.548.7148.3450.5549.20 +/- 1.19
Qwen3.5-397B50.1848.7147.6048.83 +/- 1.30
Qwen3.6-Plus49.0847.2348.7148.34 +/- 0.98
Claude-Opus-4.647.9747.6048.7148.09 +/- 0.56
GLM-4.6V46.8644.2847.9746.37 +/- 1.89
Gemini-3.5-Flash43.9141.7044.6543.42 +/- 1.54
Claude-Opus-4.836.9038.0138.0137.64 +/- 0.64
Claude-Opus-4.734.6929.8933.9532.84 +/- 2.58

The standard deviation ranges from 0.21 to 2.58 points, and the Spearman correlation between the model rankings of any two runs is at least 0.88. Seed2.0-Pro ranks first in every run, and no model reaches 60% in any run. We will report the mean +/- std for all columns of Table 1 in the revision.

Regarding the benchmark size, the 271 tasks are those that survive our filtering of 100K seed videos into 847 candidates and the verification of each candidate by three annotators (§3.2). This scale is comparable to the 300-question test set of GAIA (Mialon et al., 2024).

W2. Per-Tool Ablations on the Final Benchmark

The authors do use single-modality and no-tool tests to filter candidate questions during dataset construction, which shows that they have considered whether tools are necessary. However, they do not systematically report results on the final test set with external tools disabled, or with web search, webpage reading, or video reinspection individually removed. This makes it difficult to measure how much each tool-use contributes to the final results, or to tell whether differences between models mainly come from their initial understanding of the video or from their later search and evidence integration.

Thank you for your comment. We report the accuracy (%) on the final benchmark after removing video reinspection, web search and page visit, or all tools, as follows.

ModelCultureGeographyDaily LifeHistoryTechnologyEconomicsOverall
Qwen3.8-Max44.8342.8662.3050.0051.2850.0050.18
Qwen3.8-Max w/o Video Reinspection37.9333.3354.1045.4546.1550.0043.17
Qwen3.8-Max w/o Web Search and Page Visit36.2134.9255.7445.4543.5933.3342.80
Qwen3.8-Max w/o All Tools34.4834.9254.1031.8217.9516.6735.79
Gemini-3.1-Pro41.3849.2157.3854.5553.8550.0050.92
Gemini-3.1-Pro w/o Video Reinspection39.6642.8654.1045.4546.1516.6745.02
Gemini-3.1-Pro w/o Web Search and Page Visit32.7626.9845.9043.1838.4633.3336.90
Gemini-3.1-Pro w/o All Tools25.8620.6339.3427.2728.2116.6728.04

Removing any tool group lowers accuracy for both models, and removing all tools costs 14.39 points for Qwen3.8-Max and 22.88 points for Gemini-3.1-Pro. Qwen3.8-Max leads by 7.75 points without tools but trails by 0.74 points with all tools, so differences between models depend substantially on search and evidence integration. We will add this ablation to the revised paper.

AuthorsOpenReview ↗

Response to Reviewer qJL8 [Part 2]

Q1. Role of the Memory Field

The tool-call format includes a memory field for brief notes. How are these notes used in later steps, and what role do they play in the agentic loop?

Thank you for your question. The memory field passes brief context from earlier steps to later tool calls, with the goal of improving the success rate of tool use. For example, a video reinspection call can fail when it requests too many frames or too high a resolution for the context window. The error message is then kept in memory, so the retry can use fewer frames or a lower resolution. We observe that adding memory raises the tool-call success rate from about 80% to above 95%. We will clarify this in Appendix B.1.

References

  • Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, et al. GAIA: a benchmark for General AI Assistants. In International Conference on Learning Representations (ICLR), 2024.

Thank you again for your valuable feedback, which helped us improve the quality of this work. If our responses have addressed your concerns, we would be grateful if you could consider improving your rating. If you have any other questions or concerns that remain unresolved, we would be very happy to address them.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer qJL8,

Thank you again for your valuable feedback on our submission. In our response, we reported all three runs with standard deviations between 0.21 and 2.58 points, tool ablations on the final benchmark showing that removing all tools costs 14.39 points for Qwen3.8-Max and 22.88 points for Gemini-3.1-Pro, and how the memory field raises the tool-call success rate from about 80% to above 95%.

As the discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns. If they have, we would appreciate it if you could consider improving your rating. If any questions remain, we would be very happy to address them before the deadline.

Best regards,

The Authors of Submission 1955, VideoGAIA

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank the Area Chair and all reviewers for their careful assessment and constructive feedback. We have responded to each reviewer with three-run results, tool ablations on the final benchmark, a comparison with a fixed observation-search pipeline, cross-judge validation, filtering statistics, and clarifications of the benchmark construction and data policy. The main results are summarized below.

Contribution

VideoGAIA evaluates whether video agents can ground evidence in the input video, use external tools, and integrate information across turns to solve real-world tasks. It contains 271 human-verified tasks built through a human-model co-design pipeline, and all 20 evaluated frontier MLLMs remain below 60% accuracy under a unified agent loop.

Main Responses

Run-to-Run Variance and Judge Reliability (qJL8, XYxc, tc6k, gCKX). We report all three runs for the 20 models. The standard deviation of overall accuracy ranges from 0.21 to 2.58 points, the Spearman correlation between the model rankings of any two runs is at least 0.88, and Seed2.0-Pro ranks first in every run. Re-grading three leading models with DeepSeek-V4-Pro and Claude-Opus-4.8 changes their scores by at most two of the 271 tasks and preserves their ranking.

Tool Ablations and Adaptive Interaction (qJL8, XYxc, tc6k, gCKX). On the final benchmark, removing any tool group lowers accuracy for both Qwen3.8-Max and Gemini-3.1-Pro, and removing all tools costs 14.39 and 22.88 points, respectively. Disabling video reinspection while keeping the initial 20 frames and web tools lowers accuracy by 7.01 and 5.90 points. Under the same 40-step budget, the adaptive agent outperforms a fixed observation-search pipeline by 5.54 points on Qwen3.8-Max and 6.64 points on Gemini-3.5-Flash, two models not used in any construction or filtering stage, and it is never worse in any category.

Selection Bias from Filtering (XYxc). We report the number of candidates removed by each filtering model alone and the accuracy of all 20 models on the 847 candidates. Claude-Opus-4.7 removes the fewest candidates, and GPT-5.5 removes 2.8K candidates yet ties for third on the final benchmark, which shows no consistent penalty for the filtering models.

Benchmark Construction and Annotation (qJL8, XYxc, tc6k, gCKX). The 271 tasks are filtered from 100K seed videos and verified by three annotators, a scale comparable to the 300-question test set of GAIA. The model-based filters verify that the multi-frame and retrieval dependencies are necessary, and human review verifies that each answer is correct and unique. We also characterize all tasks by the visual evidence to recognize and the information to retrieve, and clarify the annotator background, the annotation effort, and how difficulty levels and error categories are assigned.

Related Benchmarks and Data Policy (XYxc). Compared with VideoDR, the most closely related benchmark, VideoGAIA requires more web retrieval hops, removes tasks that frontier models solve without tools, and lets agents revisit the video. The annotations are released under the CC BY-NC 4.0 license together with the video files for research use only, and the model trajectories with the retrieved webpage results will be released after the anonymity period.

Memory Field (qJL8). The memory field passes brief context, such as the error message of a failed tool call, to later tool calls, which raises the tool-call success rate from about 80% to above 95%.

We believe these results and clarifications address the main concerns raised by the reviewers. As the discussion period ends on November 18 (AoE), we would be grateful if the reviewers could let us know whether our responses have resolved their concerns, so that we still have time to address any remaining questions. We sincerely appreciate the time and effort devoted to our submission.

Best regards,

The Authors of Submission 1955, VideoGAIA

Dear Reviewers,

We are approaching the deadline of the discussion period. Please go through the authors' rebuttal at your earliest convenience, engage in the discussion, indicate any points the rebuttal missed as follow-up comments, and update your rating based on the rebuttal. As the paper received mixed reviews, your additional input is really important for this paper.

Thank you for your time and participation

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.