Scaling RL to Long Videos

General Response

We sincerely appreciate your detailed and constructive feedback. Before addressing each concern individually, we first to highlight the major updates as follows.

1. Larger Dataset: We double the size of our dataset, LongVideo-Reason, from 52K to 104K reasoning data, with carefully filtering and data cleaning. Moreover, we include both multi-choice and open-ended QAs, making it more general.
2. VideoMME Update: Thanks to the updated dataset, our model achieved strong performance, e.g., 65.1% / 71.1% on VideoMME [1] without and with subtitles, outperforming the LongVILA [2] baseline, which achieved 60.1% / 65.1% on VideoMME [1].
3. Expanded Evaluation: We have included 7 additional video benchmarks for evaluation: ActivityNet-QA [3], PerceptionTest [4], VNBench [5], LongVideoBench [6], EgoSchema [7], NExT-QA [8], and MVBench [9]. Alongside VideoMME [1], we now compare LongVILA-R1 against LongVILA [2] baseline, and other methods [10, 11, 12, 13 ,14, 15] on a total of 8 benchmarks, demonstrating consistently strong performance. Results are show in the table below.
4. Extended Inputs: LongVILA-R1-7B is capable of processing 2,048 video frames per video, with customizable FPS settings.

Detailed Response for Each Question

We provide detailed responses to each of the concerns below.

Q1: “The performance is similar compared to LongVILA”

A: As shown in the table in General Response, LongVILA-R1-7B benefits from the updated large-scale dataset, achieving 65.1% / 71.1% on VideoMME [1] without and with subtitles, representing +5.0% and +6.0% improvements over LongVILA-7B, respectively.

Q2: “Missing Video-R1-7B in the 64-frames-setting”

A: Apologies for the oversight. The 64-frame result of Video-R1-7B [15] (61.4% on VideoMME without subtitles) was updated in its v2 version on arXiv on May 14, shortly before the NeurIPS submission deadline, and we missed it. LongVILA-R1-7B achieves 65.1% on VideoMME without subtitles, surpassing Video-R1-7B. We have included the comparison in the table in the General Response and will correct this value in the revision.

Q3: “The only strong performance on the own benchmark”

A: Thank you for the reminder. As shown in the table in General Response, we have evaluated LongVILA-R1 on seven additional benchmarks. Alongside VideoMME [1], it demonstrates strong performance across eight benchmarks in total, for example, 64.8% on ActivityNet-QA [3], 68.9% on PerceptionTest [4], and 75.5% on VNBench [5], showing significant improvements over existing models.

Q4: “Typos, inconsistencies, and overly touting”

A: Thank you for pointing this out. We will correct the typos and carefully refine the language in our revision to address inconsistencies and overly promotional phrasing.

Q5: “The results of LongVILA are missing from Table 2”

A: Thank you for the reminder. The results of the LongVILA [2] baseline are provided in the table below, and we will include them in the revision.

Q6: “Will the code for data generation and the MR-SP training strategy be made available?”

A: Yes, we are fully committed to release the training data, benchmark, data generation pipeline, and the whole training pipeline with the MR-SP training system.

Q7: “The leading open-source reasoning LLM”

A: It is the DeepSeek-R1-671B [16] model, and we will clarify this in our revision.

Q8: “What does ‘balanced’ here mean?”

A: By "balanced," we mean that the number of questions is balanced across the four categories: temporal, goal and purpose, plot and narrative, and spatial.

Q9: “Small comment on the soccer example“

A: Thank you for the suggestion. We will include additional examples in the revised version to better demonstrate the advantages of our model.

References

[1] Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis. CVPR 2025

[2] LongVILA: Scaling Long-Context Visual Language Models for Long Videos. ICLR 2025

[3] ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering. AAAI 2019

[4] Perception Test: A Diagnostic Benchmark for Multimodal Video Models. NeurIPS 2023

[5] Needle In A Video Haystack: A Scalable Synthetic Framework for Benchmarking Video MLLMs. ICLR 2025

[6] LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. NeurIPS 2024

[7] EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding. NeurIPS 2023

[8] NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions. CVPR 2021

[9] MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. CVPR 2024

[10] Video-LLaVA: Learning United Visual Representation by Alignment Before Projection. EMNLP 2024

[11] ShareGPT4Video: Improving Video Understanding and Generation with Better Captions. NeurIPS 2024

[12] VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. ArXiv 2024

[13] Kangaroo: A Powerful Video-Language Model Supporting Long-context Video Input. ArXiv 2024

[14] LLaVA-OneVision: Easy Visual Task Transfer. Trans. Mach. Learn. Res. 2025

[15] Video-R1: Reinforcing Video Reasoning in MLLMs. ArXiv 2025

[16] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. ACL 2025

Dear Reviewer XSh3,

Thank you once again for your constructive and insightful review, which has significantly helped improve the quality of our work.

Following your valuable suggestions, we have revised our paper and submitted a detailed rebuttal. In summary, we have:

1. Conducted comprehensive comparisons, including evaluations against LongVILA across 8 benchmarks with non-trivial improvements, and comparisons with Video-R1 (64-frame setting) on VideoMME (w/o sub) (see General Response, Q1, Q2, Q3, Q5).
2. Provided additional clarifications on typos, code release, and other details (see Q4, Q6, Q7, Q8, Q9).

We hope these updates have addressed your concerns. As the discussion period is drawing to a close, we would greatly appreciate it if you could let us know whether our response has resolved your questions.

Thank you again for your time and consideration.

Sincerely,

Authors of Paper 8328

General Response

We sincerely appreciate your detailed and constructive feedback. Before addressing each concern individually, we first to highlight the major updates as follows.

1. Larger Dataset: We double the size of our dataset, LongVideo-Reason, from 52K to 104K reasoning data, with carefully filtering and data cleaning. Moreover, we include both multi-choice and open-ended QAs, making it more general.
2. VideoMME Update: Thanks to the updated dataset, our model achieved strong performance, e.g., 65.1% / 71.1% on VideoMME [1] without and with subtitles, outperforming the LongVILA [2] baseline, which achieved 60.1% / 65.1% on VideoMME [1].
3. Expanded Evaluation: We have included 7 additional video benchmarks for evaluation: ActivityNet-QA [3], PerceptionTest [4], VNBench [5], LongVideoBench [6], EgoSchema [7], NExT-QA [8], and MVBench [9]. Alongside VideoMME [1], we now compare LongVILA-R1 against LongVILA [2] baseline, and other methods [10, 11, 12, 13 ,14, 15] on a total of 8 benchmarks, demonstrating consistently strong performance. Results are show in the table below.
4. Extended Inputs: LongVILA-R1-7B is capable of processing 2,048 video frames per video, with customizable FPS settings.

Detailed Response for Each Question

We provide detailed responses to each of the concerns below.

Q1: “The diversity of the LongVideo-Reason dataset”

A: Apologies for the confusion. The videos in LongVideo-Reason are not solely sourced from the Shot2Story [16] dataset. We also incorporated an additional 2,000 4K-resolution videos covering diverse scenarios, including autonomous driving, video games, household robotics, and wildlife. We used VILA-HD [17] to generate object bounding boxes within video frames and constructed spatial reasoning Q&A pairs based on these boxes and their corresponding captions. We will add detailed description of this in the revision.

The diversity of the LongVideo-Reason dataset also stems from its reasoning-annotated QA pairs, which are categorized into four reasoning types: Temporal, Goal and Purpose, Spatial, and Plot and Narrative. After thorough filtering and cleaning, the LongVideo-Reason dataset contains 104K QA pairs with explicit reasoning annotations. Representative visual examples are provided in Figures 10 and 11.

Additionally, the dataset contains a balanced mix of multiple-choice and open-ended questions, enhancing its versatility. As a result, as shown in the General Response, this dataset significantly contributes to the model’s strong performance across multiple benchmarks, which cover a diversity of video tasks.

Q2: “The novelty of the dataset generation and the training scheme”

A: We answer this from two perspectives:

(1) Dataset Generation

The novelty of our dataset generation lies in the pipeline design: we first use VLMs to generate text captions for video frames, and then leverage powerful reasoning-oriented LLMs to generate multi-step reasons. While advanced LLMs like DeepSeek-R1-671B [18] possess strong reasoning capabilities, they cannot process video inputs. Our pipeline bridges the gap between long-video VLMs and the reasoning-oriented LLMs.

Different from existing video datasets [19] which only provide QAs, our dataset includes detailed reasons for QA pairs. While some datasets do include reasoning annotations, they are designed for LLMs, such as mathematics [20], not videos.

Moreover, this dataset helps LongVILA-R1-7B to achieve strong performance (e.g., 65.1/71.1 on VideoMME) and serves as a solid baseline for future research, making it a meaningful contribution to the field.

(2) Training Scheme

The novelty of our training scheme lies in the co-design of algorithm and infrastructure, even though the SFT+RL algorithm itself is not new. A central component of this scheme is the MR-SP system, which is carefully tailored for efficient long-video RL. It features several key designs:

1. We first implemented video encoding with balanced distribution across multiple GPUs. (Traditional RL frameworks often run into GPU OOM issues when processing more than 512 frames)
2. We introduce a chunked distributed gathering strategy for multimodal inputs and embeddings. (Traditional RL frameworks often suffer from CPU OOM issues due to the combination of long video frames and a large number of rollouts)
3. We further cache and reuse video embeddings during both the rollout and pre-filling phases, effectively reducing redundant computation. (Traditional RL frameworks often suffer from high latency due to repetitive processing)

Without this co-design, the training algorithm alone is insufficient to handle long-video training effectively. In contrast, our training framework is capable of training on 3,600 video frames using a single node with 8 A100 GPUs.

We are fully committed to releasing the dataset, the complete code of dataset generation pipeline, and the training framework, including the MR-SP system, to support future research in the community.

Q3: “Evaluation on Video-MME”

A: Thanks for your reminder, we include more VLMs comparison on VideoMME, including NVILA-8B-Video [21] (CVPR 2025), Apollo-7B [22] (CVPR 2025), and VideoLLaMA3-7B [23] (ArXiv 2025).

LongVILA-R1 achieves performance comparable to VideoLLaMA3. Although VideoLLaMA3 does not incorporate RL training, it uses 5.72 million samples during video fine-tuning, Table 4 in VideoLLaMA3 [23]. In contrast, we use only 0.21 million samples in total across both SFT and RL, resulting in higher data efficiency and much lower training cost than VideoLLaMA3.

Q4: “The comparison with Gemini-1.5-Pro and GPT-4o”

A: To address concerns of zero-shot evaluation, we further compare LongVILA-R1-7B with Gemini-1.5-Pro, GPT-4o, and GPT-4o-mini on ActivityNet-QA, VNBench, and VideoMME. LongVILA-R1-7B outperforms Gemini-1.5-Pro and GPT-4o on ActivityNet-QA and VNBench, and inferior to them on VideoMME, where LongVILA-R1-7B surpasses GPT-4o-mini with only 7B parameters.

Q5: “Long video frames”

A: As mentioned in the General Response, LongVILA-R1-7B is capable of processing up to 2,048 video frames per video. We progressively increase the input frames when evaluating on VNBench, and observe that performance improves as the frames increases.

Q6: “Variable or high frame rates”

A: LongVILA-R1-7B supports both variable and high FPS settings. We evaluate it on VideoMME using FPS values of 8, 16, and 32, all of which result in strong performance.

References

[1] Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis. CVPR 2025

[2] LongVILA: Scaling Long-Context Visual Language Models for Long Videos. ICLR 2025

[3] ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering. AAAI 2019

[4] Perception Test: A Diagnostic Benchmark for Multimodal Video Models. NeurIPS 2023

[5] Needle In A Video Haystack: A Scalable Synthetic Framework for Benchmarking Video MLLMs. ICLR 2025

[6] LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. NeurIPS 2024

[7] EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding. NeurIPS 2023

[8] NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions. CVPR 2021

[9] MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. CVPR 2024

[10] Video-LLaVA: Learning United Visual Representation by Alignment Before Projection. EMNLP 2024

[11] ShareGPT4Video: Improving Video Understanding and Generation with Better Captions. NeurIPS 2024

[12] VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. ArXiv 2024

[13] Kangaroo: A Powerful Video-Language Model Supporting Long-context Video Input. ArXiv 2024

[14] LLaVA-OneVision: Easy Visual Task Transfer. Trans. Mach. Learn. Res. 2025

[15] Video-R1: Reinforcing Video Reasoning in MLLMs. ArXiv 2025

[16] Shot2Story: A New Benchmark for Comprehensive Understanding of Multi-shot Videos. ICLR 2025

[17] Scaling Vision Pre-Training to 4K Resolution. CVPR 2025

[18] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. ACL 2025

[19] Video Instruction Tuning With Synthetic Data. ArXiv 2024

[20] Training Verifiers to Solve Math Word Problems. ArXiv 2021

[21] NVILA: Efficient Frontier Visual Language Models. CVPR 2025

[22] Apollo: An Exploration of Video Understanding in Large Multimodal Models. CVPR 2025

[23] VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding. ArXiv 2025

Dear Reviewer oP4r,

Thank you once again for your constructive and insightful review, which has significantly helped improve the quality of our work.

Following your valuable suggestions, we have revised our paper and submitted a detailed rebuttal. In summary, we have provided:

1. Detailed discussions on the diversity of the dataset as well as the novelty of the dataset generation process and training scheme (see Q1 & Q2).
2. Comprehensive comparisons across 8 benchmarks, including VideoMME, along with new evaluations against Gemini-1.5-Pro and GPT-4o (see General Response, Q3 & Q4).
3. Additional ablation studies on long video frames and variable / high frame rates.

We hope these updates have addressed your concerns. As the discussion period is drawing to a close, we would greatly appreciate it if you could let us know whether our response has resolved your questions. If there are any remaining concerns, we will make every effort to address them.

Thank you again for your time and consideration.

Sincerely,

Authors of Paper 8328

General Response

We sincerely appreciate your detailed and constructive feedback. Before addressing each concern individually, we first to highlight the major updates as follows.

1. Larger Dataset: We double the size of our dataset, LongVideo-Reason, from 52K to 104K reasoning data, with carefully filtering and data cleaning. Moreover, we include both multi-choice and open-ended QAs, making it more general.
2. VideoMME Update: Thanks to the updated dataset, our model achieved strong performance, e.g., 65.1% / 71.1% on VideoMME [1] without and with subtitles, outperforming the LongVILA [2] baseline, which achieved 60.1% / 65.1% on VideoMME [1].
3. Expanded Evaluation: We have included 7 additional video benchmarks for evaluation: ActivityNet-QA [3], PerceptionTest [4], VNBench [5], LongVideoBench [6], EgoSchema [7], NExT-QA [8], and MVBench [9]. Alongside VideoMME [1], we now compare LongVILA-R1 against LongVILA [2] baseline, and other methods [10, 11, 12, 13 ,14, 15] on a total of 8 benchmarks, demonstrating consistently strong performance. Results are show in the table below.
4. Extended Inputs: LongVILA-R1-7B is capable of processing 2,048 video frames per video, with customizable FPS settings.

Detailed Response for Each Question

We provide detailed responses to each of the concerns below.

Q1: “Evaluations on more benchmarks”

A: Thank you for your suggestion. We fully agree on the importance of evaluating the performance of our reasoning model on a more comprehensive set of benchmarks. As mentioned in General Response, we evaluated LongVILA-R1 on eight benchmarks, including VideoMME (w/o sub), VideoMME (w/ sub) [1], ActivityNet-QA [3], PerceptionTest [4], VNBench [5], LongVideoBench [6], EgoSchema [7], NExT-QA [8], and MVBench [9]. Under these benchmarks, LongVILA-R1 consistently outperforms LongVILA [2] and demonstrates more comprehensive imporvements compared to other video VLMs with similar parameter scales.

Q2: “The primary technical contributions are in the Sequence Parallel training, while the architecture and training methods (GRPO) are directly applied.”

A: We really appreciate the reviewer's acknowledgement of one of our core contributions, MR-SP system, which enable training on long videos with up to 3,600 frames using a single node with 8 A100 GPUs.

In addition to the technical contributions of MR-SP, our work also provides a dataset, a benchmark, and, more importantly, a comprehensive pipeline for generating large-scale datasets tailored for long-video RL training. We believe this contribution has a big value in this area.

The novelty of our training scheme lies in the co-design of the algorithm and infrastructure, even though the SFT+RL algorithm itself is not new. The infrastructure features several key design elements:

1. We first implemented video encoding with balanced distribution across multiple GPUs. (Traditional RL frameworks often run into GPU OOM issues when processing more than 512 frames.)
2. We introduce a chunked distributed gathering strategy for multimodal inputs and embeddings. (Traditional RL frameworks often suffer from CPU OOM issues due to the combination of long video frames and a large number of rollouts.)
3. We further cache and reuse video embeddings during both the rollout and pre-filling phases, effectively reducing redundant computation. (Traditional RL frameworks often suffer from high latency due to repetitive processing.)

Without this co-design, the training algorithm alone is insufficient to handle long-video training effectively. In contrast, our training framework is capable of training on 3,600 video frames using a single node with 8 A100 GPUs.

We are fully committed to releasing the dataset, the complete code of the dataset generation pipeline, and the training framework, including the MR-SP system, to support future research in the community.

Q3: “Result on standard video benchmarks like MVBench, ApolloBench”

A: Thank you for the reminder. As mentioned above, we have reported results on seven additional benchmarks, including MVBench [9] as you suggested. As for ApolloBench [16], the benchmark is not publicly available at this time. We will evaluate our model on it when it becomes open-sourced.

Q4: “More details and visualizations of the dataset”

A: Thank you for the reminder. We will include additional details and visualizations of the dataset in the revised version.

We illustrate the data distribution in Figure 3 and the dataset build-up pipeline in Figure 4. The videos span a wide range of categories, including travel, sports, education, animals, blogs, news, music, technology, comedy, entertainment, film, and gaming. The QA pairs are categorized into four reasoning types: Temporal, Goal and Purpose, Spatial, and Plot and Narrative. After thorough filtering and cleaning, the LongVideo-Reason dataset contains 104K QA pairs with explicit reasoning annotations. Representative visual examples are provided in Figures 10 and 11.

In addition, we further collected 2,000 additional 4K-resolution videos spanning scenarios such as autonomous driving, video games, household robotics, and wildlife. We employed VILA-HD [17] to generate object bounding boxes in video frames and constructed spatial reasoning Q&A pairs based on these boxes and corresponding captions.

Additionally, the dataset contains a balanced mix of multiple-choice and open-ended questions, enhancing its versatility. To facilitate future research in video reasoning, we will release the dataset and the complete dataset construction pipeline to the community. We will also add more visualizations of the dataset in the future version of the paper.

References

[1] Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis. CVPR 2025

[2] LongVILA: Scaling Long-Context Visual Language Models for Long Videos. ICLR 2025

[3] ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering. AAAI 2019

[4] Perception Test: A Diagnostic Benchmark for Multimodal Video Models. NeurIPS 2023

[5] Needle In A Video Haystack: A Scalable Synthetic Framework for Benchmarking Video MLLMs. ICLR 2025

[6] LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. NeurIPS 2024

[7] EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding. NeurIPS 2023

[8] NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions. CVPR 2021

[9] MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. CVPR 2024

[10] Video-LLaVA: Learning United Visual Representation by Alignment Before Projection. EMNLP 2024

[11] ShareGPT4Video: Improving Video Understanding and Generation with Better Captions. NeurIPS 2024

[12] VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. ArXiv 2024

[13] Kangaroo: A Powerful Video-Language Model Supporting Long-context Video Input. ArXiv 2024

[14] LLaVA-OneVision: Easy Visual Task Transfer. Trans. Mach. Learn. Res. 2025

[15] Video-R1: Reinforcing Video Reasoning in MLLMs. ArXiv 2025

[16] Apollo: An Exploration of Video Understanding in Large Multimodal Models. CVPR 2025

[17] Scaling Vision Pre-Training to 4K Resolution. CVPR 2025

General Response

We sincerely appreciate your detailed and constructive feedback. Before addressing each concern individually, we first to highlight the major updates as follows.

1. Larger Dataset: We double the size of our dataset, LongVideo-Reason, from 52K to 104K reasoning data, with carefully filtering and data cleaning. Moreover, we include both multi-choice and open-ended QAs, making it more general.
2. VideoMME Update: Thanks to the updated dataset, our model achieved strong performance, e.g., 65.1% / 71.1% on VideoMME [1] without and with subtitles, outperforming the LongVILA [2] baseline, which achieved 60.1% / 65.1% on VideoMME [1].
3. Expanded Evaluation: We have included 7 additional video benchmarks for evaluation: ActivityNet-QA [3], PerceptionTest [4], VNBench [5], LongVideoBench [6], EgoSchema [7], NExT-QA [8], and MVBench [9]. Alongside VideoMME [1], we now compare LongVILA-R1 against LongVILA [2] baseline, and other methods [10, 11, 12, 13 ,14, 15] on a total of 8 benchmarks, demonstrating consistently strong performance. Results are show in the table below.
4. Extended Inputs: LongVILA-R1-7B is capable of processing 2,048 video frames per video, with customizable FPS settings.

Detailed Response for Each Question

We provide detailed responses to each of the concerns below.

Q1: “Missing Key Baseline, Limited Performance Gains, Limited Evaluation Scope”

A: As mentioned in the General Response, we have evaluated LongVILA-R1-7B on eight benchmarks, including the three you suggested: ActivityNet-QA[3], LongVideoBench [6], and EgoSchema[7], and compared to the LongVILA-7B baseline. Across all benchmarks, LongVILA-R1-7B consistently outperforms LongVILA-7B [9], achieving 65.1% and 71.1% on VideoMME [1], with +5.0% and +6.0% improvements over LongVILA-7B [2].

Q2: “Cost-Benefit Analysis”

A: Thank you for your constructive suggestion. We justify the performance gains from LongVILA-7B to LongVILA-R1-7B on VideoMME [1], along with the associated training overhead on H100. Both CoT SFT and RL fine-tuning contribute to meaningful improvements, from 65.1 → 66.9 → 71.1 (with subtitles). While CoT SFT and RL require 82 GPU-days and 224 GPU-days, the performance gains are non-trivial.

In addition, as shown in the table in General Response, LongVILA-R1-7B achieves notable improvements over LongVILA-7B on several other benchmarks—for example, from 59.5% to 64.8% on ActivityNet-QA [3], 58.1% to 68.9% on PerceptionTest [4], and 63.0% to 75.5% on VNBench [5].

References

[1] Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis. CVPR 2025

[2] LongVILA: Scaling Long-Context Visual Language Models for Long Videos. ICLR 2025

[3] ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering. AAAI 2019

[4] Perception Test: A Diagnostic Benchmark for Multimodal Video Models. NeurIPS 2023

[5] Needle In A Video Haystack: A Scalable Synthetic Framework for Benchmarking Video MLLMs. ICLR 2025

[6] LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. NeurIPS 2024

[7] EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding. NeurIPS 2023

[8] NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions. CVPR 2021

[9] MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. CVPR 2024

[10] Video-LLaVA: Learning United Visual Representation by Alignment Before Projection. EMNLP 2024

[11] ShareGPT4Video: Improving Video Understanding and Generation with Better Captions. NeurIPS 2024

[12] VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. ArXiv 2024

[13] Kangaroo: A Powerful Video-Language Model Supporting Long-context Video Input. ArXiv 2024

[14] LLaVA-OneVision: Easy Visual Task Transfer. Trans. Mach. Learn. Res. 2025

[15] Video-R1: Reinforcing Video Reasoning in MLLMs. ArXiv 2025

Dear Reviewer HSHu,

Thank you once again for your constructive and insightful review, which has significantly helped improve the quality of our work.

Following your valuable suggestions, we have revised our paper and submitted a detailed rebuttal. In summary:

1. We conducted comprehensive comparisons across 8 benchmarks, including the key baseline LongVILA, and achieved non-trivial performance improvements (see General Response & Q1).
2. We provided a thorough cost-benefit analysis, showing that the training cost is reasonable and relative to the performance gains (see Q2).

We hope these updates have addressed your concerns. As the discussion period is drawing to a close, we would greatly appreciate it if you could let us know whether our response has resolved your questions.

Thank you again for your time and consideration.

Sincerely,

Authors of Paper 8328

Dear Reviewer HSHu,

We sincerely appreciate your thoughtful and constructive review, which has greatly helped improve the quality of our work.

Following your valuable feedback, we have carefully revised the paper and submitted a detailed rebuttal. As the discussion phase is coming to an end, we would greatly appreciate it if you could kindly let us know whether our response has addressed your concerns.

Thank you again for your time and consideration.

Sincerely,

Authors of Paper 8328

Dear Reviewer HSHu,

We hope this message finds you well.

Following your valuable review of our paper, we have carefully revised the paper and submitted a detailed rebuttal addressing your concerns. As the discussion period will end in just over one day, we would greatly appreciate it if you could let us know whether our responses have resolved your concerns.

Thank you again for your time and contribution to the review process.

Sincerely,

Authors of Paper 8328

Dear Reviewer,

The authors have responded to your reviews. Does it change your opinion? Please kindly engage in discussion and provide any feedback or questions if you may have. Thanks!

AC

Dear Reviewer HSHu,

I hope this message finds you well.

Following your valuable review of our paper, we have carefully addressed all your concerns in our rebuttal. Your review has been very helpful in improving our paper, and we have revised the paper based on your comments.

As the discussion period will conclude in about only 10 hours, we would be very grateful if you could share your thoughts on our responses before the deadline. Your feedback would be invaluable in ensuring a complete and balanced discussion.

Thank you again for your time and effort in reviewing our work.

Sincerely,

Authors of Paper 8328