Questions on Multimodal and Agentic RL Support in Slime Roadmap

[Aurelius84]

Hi Slime Team (cc @zhuzilin ),

I have recently studied the Slime framework's source code and was impressed by its elegant and lightweight design. I am from the Xiaohongshu team, and we focus more on large-scale multimodal model and Agentic RL scenarios. I have several questions regarding the future roadmap of Slime and would appreciate your insights:

1. Are there any plans to support RL scenarios with large-scale multimodal foundation models?
2. Has there been any internal validation or experience with Agentic RL training in Slime?
3. Is there a roadmap for scaling to massive-size multimodal models and Agentic RL? Specifically, have you considered design strategies for rollouts addressing long-tail issues in such scenarios?
   I
   f possible, could you share some information relevant to these points? From an RL framework technology perspective, we are very interested in collaborating and advancing together with the Slime community.

Looking forward to your sharing and suggestions.

[yitianlian]

Hi! I’m Chengxing Xie, one of the main contributors, together with Zilin, on the Agentic RL component of the Slime framework. For Agentic RL scenarios, you can refer to the GLM 4.5 technical report, where we provide a detailed description of how we implement Agentic RL on large MoE models.

To address the long-tail issue in agentic settings, we adopt a fully asynchronous RL training approach, ensuring that rare or long-tail data does not block the overall training process. I’ve also written a blog post explaining our perspective on Agentic RL and the system design choices in Slime that enable it.

For multi-modality RL support, we recognize that the diverse architectures of VLMs make it challenging to align them within a unified RL training pipeline. We are actively working on solutions to better integrate these structures.

If you’re interested in collaborating, feel free to reach out!
cc @zhuzilin

[Aurelius84]

Thanks for reply~