Request for Agentic RL training details (thank you for open-sourcing!)

[zkangning]

I’m trying to follow/reproduce the Agentic RL part, but some training details are not fully clear from the current docs.

Would you consider open-sourcing the training code/details for the Agentic RL part to make it easier for others to reproduce and build on your results?

Thanks again for the great open-source release!

[Raibows]

Hi @zkangning

Thank you for your interests! Since the synthesized environments is exposed via MCP, it should be easy to adapt to tools-learning agentic RL framework. If you have any implementation questions, please feel free to email me.

The RL codebase may take additional internal check for release. Please stay tuned!

[zkangning]

@Raibows Thank you for your response. I am likewise confused about the BFCL v3 results reported for multiple models in the paper. According to the official results as well as my own experiments, the Qwen3-8B Base Model achieves an Overall metric of approximately 65% on BFCL v3, whereas the paper reports 53.83%.

I also downloaded AWM-8B and evaluated it using the official Qwen3Handler, obtaining an Overall Score of 65.61%.

Therefore, I would like to ask how the BFCL v3 metrics reported in the paper are measured. If the official evaluation code is used, the results for the Base Models Qwen3-4B, Qwen3-8B, and Qwen3-14B should be close to 63%, 66%, 68%, respectively.

Sincerely look forward to your reply. If possible, could you make the evaluation code publicly available to help us align the metrics reported in the paper.

[Raibows]

Hi @zkangning,

We use a MCP specialized system prompt in awm/core/agent.py when we evaluate all the benchmarks. Thus, we slightly modified the BFCL code to bridge the instructed agent behavior and evaluation. Could you please test how the model perform under the used system prompt?

I will try to release the evaluation codes. Thank you!