Dear Author, Thank you for your contribution. This is one of the most inspiring papers I've read recently, very concise and efficient. I also plan to try this approach, but I have some questions:
-
The paper seems to only mention SSL-R1-591K as the training set, while using datasets like MMBench as the test set. However, MMBench contains other subsets. I would like to ask if it is necessary to use subsets for training, as the task in SSL-R1-591K may differ from other test sets (e.g., MMBench). For example, I need to perform an image captioning generation task, and I want to know if directly using this proxy task will allow the model to generate image captions well (without needing to train the image captioning task), or if it only plays a supporting role (requiring training the image captioning task).
-
Assuming image captioning training is required, I would like to ask about the training order. For example, SFT first, then RL, and finally UPT (https://arxiv.org/abs/2505.22453). This method has self-boosting properties, but the synthesized data is essentially highly correlated with the test set. However, SSL-R1, being handled by a proxy task, might be irrelevant to the test set. I'd like to ask if the SSL-R1 method should be placed last? Or, for example, SFT -> SSL-R1 -> RL (OPD)?
Looking forward to your reply; any insights—whether from your experiments or theoretical thinking—would be greatly appreciated! 🙏
Dear Author, Thank you for your contribution. This is one of the most inspiring papers I've read recently, very concise and efficient. I also plan to try this approach, but I have some questions:
The paper seems to only mention SSL-R1-591K as the training set, while using datasets like MMBench as the test set. However, MMBench contains other subsets. I would like to ask if it is necessary to use subsets for training, as the task in SSL-R1-591K may differ from other test sets (e.g., MMBench). For example, I need to perform an image captioning generation task, and I want to know if directly using this proxy task will allow the model to generate image captions well (without needing to train the image captioning task), or if it only plays a supporting role (requiring training the image captioning task).
Assuming image captioning training is required, I would like to ask about the training order. For example, SFT first, then RL, and finally UPT (https://arxiv.org/abs/2505.22453). This method has self-boosting properties, but the synthesized data is essentially highly correlated with the test set. However, SSL-R1, being handled by a proxy task, might be irrelevant to the test set. I'd like to ask if the SSL-R1 method should be placed last? Or, for example, SFT -> SSL-R1 -> RL (OPD)?
Looking forward to your reply; any insights—whether from your experiments or theoretical thinking—would be greatly appreciated! 🙏