Questions about FlowGRPO optimization

[Darius-H]

GenRL/README.md

Line 94 in 82a867e

### 📝 Video OCR Training Comparison

In the Readme, it's mentioned that your project can make the OCR score rise faster during training than flow grpo and the growth curve smoother. Could you introduce what specific optimizations you have implemented? Thanks

[Harahan]

Actually, I align the algorithm and hyper-parameter of this repo with flow grpo. The reasons could be attributed to implementation. For example, we fix some bugs (e.g., yifan123/flow_grpo#107, which actually is caused by the pre-sampling mechanism of num_worker and still not solved in that pr).