GRPO with tool calling

[accupham]

Method description

Would it be possible to implement an RL environment that does multi-turn tool calling in the GRPO RL training loop? Right now it seems to be a one-shot inference before passing it to the custom reward function. I'd like to have a multi-turn interaction via tool calling step before passing the final result to the reward function.

Online tool calling would enable RL over a simulation with feedback from the environment. Seems vLLM supports all manner of tool calling. Could this be added as part of GRPO?

@qgallouedec

Open source status

• The method implementation is available
• The model weights are available
• The training datasets are available

Provide useful links for the implementation

No response