Difference with the Slime project

[dangerzone]

It seems like there's a lot of similarities between AgentRL and Slime when it comes to exposing a single API for rollout generation, deep SGLang integration, and async RL training, with the key difference being the addition of cross-policy sampling and task advantage normalization.

Is AgentRL the future of RL post-training at Z.ai and does this project replace Slime? If starting a new multi-turn agentic RL training project, should one build on top of AgentRL or slime?

[zhc7]

Hi, these two frameworks were developed independently and concurrently by two separate teams, and are not meant to replace each other. The general principles of the training component are basically the same, and—as far as I know—the speed is roughly comparable when working with models under 32B. There are some differences in focus: this project offers deeper and more native support for the development and deployment of an agentic environment. Feel free to use whichever framework you prefer. Hope this helps!