The effectiveness of stepGRPO

[CCCTong]

Thanks for your excellent work! But I am still curious about the improvement brought by stepGRPO. Have you compared the results of GRPO training after fine-tuning with Mulberry260K data in the warm up phase?

[CCCTong]

The GRPO fine-tuning mentioned here is similar to other RL method such as VLM-R1 and Visual-RFT, which only uses format + accuracy reward

[RohollahHS]

The GRPO fine-tuning mentioned here is similar to other RL method such as VLM-R1 and Visual-RFT, which only uses format + accuracy reward

I think that in Table 4 of the paper, the outcome-level reward actually refers to GRPO-based RL tuning, which only uses the final answer to compute the reward.