RSI-Craft trains LLM agents with RSI-Dojo

Researchers introduce RSI-Dojo, a collection of executable model training environments created by an automatic pipeline called RSI-Craft to teach LLM agents how to conduct post-training. Using these environments, the authors collected high-quality training trajectories and fine-tuned Qwen models, resulting in significant performance gains on PostTrainBench that match frontier models such as GLM-5.2 and GPT 5.5. Open article →