Question: Training time for DemyAgent and RLAnything? #12
Description
Hi, thanks for the great work on both papers!
I've read through both the papers (arXiv:2510.11701 and arXiv:2602.02488) and the repository configs carefully, but could not find any explicit training time for either work. I wanted to ask for these numbers to better understand the compute budget needed for reproduction.
What the papers/configs already tell us
DemyAgent (Appendix A.1):
- Hardware: 8× Tesla-A100-80G (single node)
- SFT: 5 epochs, batch size 32
- RL: 3 epochs, batch size 64
- → Missing: how many hours/days did SFT and RL each take?
RLAnything:
- GUI Agent (OSWorld): 12 nodes × 8 GPUs = 96× A100, 240 RL steps
- LLM Agent (AlfWorld): 8 nodes × 8 GPUs = 64× A100, 200 RL steps
- Coding LLM: 4 nodes × 8 GPUs = 32× A100, 300 RL steps
- → Missing: training time per setting?
This information would be very helpful for the community to plan reproduction efforts and estimate costs. Thanks in advance!