rLLM is Agentica’s open-source framework for post-training language agents. In DeepSWE: Training a Fully Open-sourced, State-of-the-Art Coding Agent by Scaling RL it is the system on which the DeepSWE-Preview coding agent was trained, as were the team’s earlier releases DeepCoder-14B-Preview and DeepScaleR-1.5B-Preview. That post describes DeepSWE-Preview as powered by rLLM and states the team’s mission as democratizing RL for LLMs.
The post’s contribution notes describe work on rLLM as part of the DeepSWE project: implementing agent and environment abstractions for it, optimizing its performance, and designing and implementing the initial rLLM system used to train DeepSWE.