Questions and Discussion Regarding Your Paper

[Lee-xeo]

First and foremost, congratulations on your excellent work and significant contributions to the field of large multimodal models. After a careful reading, we have a few questions we would be grateful to discuss with you.

On the Comparison with the Qwen-3-VL-Think Variant:
We noted that alongside Qwen-3-VL, a "Think" variant was also released. As we understand it, both variants share the same model architecture but differ in their training data and strategies. Given that the Qwen-3-VL-Think variant is also optimized via reinforcement learning and is capable of processing both image and video tasks, its technical setup appears highly relevant to your work. Could you please elaborate on the reason for not including a comparative analysis with the Qwen-3-VL-Think variant in your paper? We believe such a comparison could further highlight the unique advantages of your proposed method.

On Variable Control in the Data Modality Ablation Study:
In your ablation study on data modalities, you demonstrated the impact of each modality by removing its corresponding data, which provides valuable insights. However, we have a question regarding the experimental design: could this approach potentially introduce a confounding variable? Specifically, the observed performance degradation might be attributed not only to the "absence of a modality's capability" but also to the "reduction in the total volume of training data." We were wondering if you had considered alternative experimental setups to more precisely isolate the effects of these two factors—for instance, by replacing the removed data with an equal amount of data from other modalities. This could potentially strengthen the conclusions drawn from the ablation study.

On Comparison with Other State-of-the-Art Reinforcement Learning Methods:
Your paper effectively demonstrates the efficacy of EMA-GRPO through comparisons with methods like GRPO. Concurrently, we have seen the emergence of several new and promising RL algorithms in the community, such as GAPO, GSPO and SPAO (used by qwen3-vl-think). We are curious if you have considered benchmarking your method against these more recent algorithms. We believe that a broader comparative analysis would further solidify the significance and robustness of EMA-GRPO, and we would be very interested to see how it performs.

Thank you again for your time and for your outstanding contribution to the field. We look forward to your response.

[tulerfeng]

Thank you very much for your careful reading of our work.

1. We agree that the Qwen-3-VL-Think model is a very strong and relevant baseline, and we highly appreciate the impressive work from the Qwen team in developing such a powerful thinking-oriented multimodal model. We will add comparison in the future.

2. This is a very insightful point. In practice, during the RL stage, we observe that the performance improvement typically converges within roughly 200～300 global steps (with global batch size 128 and mini-batch size 32). Therefore, the final performance is not sensitive to the total amount of training data.

3. Our proposed EMA-GRPO is primarily designed to address the heterogeneous reward characteristics across different visual tasks, which is a distinct challenge from those tackled by other GRPO variants. Technically, EMA-GRPO mainly modifies the estimation of the reward standard deviation, and is therefore orthogonal to methods like GSPO rather than conflicting with them. In fact, these techniques can potentially be combined, and we are particularly interested in GSPO for its effectiveness in MoE training, which we plan to explore in future work.

[crystalroundy19162545-ship-it]

I share a similar concern regarding the comparative analysis. In your paper, the comparisons are primarily conducted against methods based on Qwen2.5. However, given that Qwen3VL is inherently powerful enough to outperform previous approaches, it might be more appropriate to also compare against other methods based on Qwen3VL (and clarify this) for both image and video tasks. Otherwise, it is easy to attribute the performance gains primarily to the strength of Qwen3VL itself.

[tulerfeng]

@crystalroundy19162545-ship-it

Hi, Thanks for pointing out.

As Qwen3VL-8B-Instruct. as just released about 1 month ago, there're few models based on Qwen3VL-8B-Instruct. Therefore, we mainly choose the strong Qwen3VL-8B-Instruct as the baseline.

[GuoYumin-hku]

@tulerfeng I want to share another question, in the section of introducing EMA-GRPO, it says:

Specifically, examples with very small or very large variance receive stronger updates, while medium-difficulty samples—whose rollouts usually have large variance—are under-optimized. As a result, the reinforcement learning within a task becomes biased.

So what is the exact opinion about large variance exmaples here? Or do you mean the "variance" at first stands for difficulty?

[tulerfeng]

@GuoYumin-hku Hi, thanks for your question.

Yes, here variance refers to the standard deviation of rollout rewards for a given question. For example, a reward sequence like
[1,1,1,1,0,0,0,0] has larger variance, while [0,0,0,0,0,0,0,1] has smaller variance, which can lead to imbalanced gradient updates. The rollout variance of a question is influenced by both its intrinsic difficulty and the model’s current capabilities. This issue was first discussed in “Understanding R1-Zero-Like Training: A Critical Perspective.”