RSI-Craft paper trains models to run their own post-training
Turning papers into executable post-training environments is a concrete step toward models that improve themselves, and the benchmark parity with GLM-5.2 and GPT 5.5 shows the approach already competes with frontier results. Practitioners fine-tuning smaller open models can study the 281-task pipeline and trajectory filtering method to replicate the gains. The released Qwen-based checkpoints make the results directly testable.
A new paper on recursive self-improvement introduces RSI-Craft, a pipeline that converts research papers into executable training environments with datasets, metrics and evaluation code. After filtering for reliability and difficulty, 281 post-training tasks remained. Frontier models solved these tasks under an improved harness, and the filtered trajectories were used to fine-tune Qwen models. The 9B version gains 8.74 points, while RSI-Dojo-35B-A3B gains 16.87 points on PostTrainBench, reaching 29.95 — comparable to GLM-5.2 and GPT 5.5.
Great new paper on recursive self-improvement.
Self-improving AI is one of the most discussed ideas right now. This paper explores training a model to run LLM post-training itself.
(bookmark it)
The team built RSI-Craft, a pipeline that turns research papers into executable training environments, including datasets, metrics, and evaluation code. After filtering for reliability and difficulty, 281 post-training tasks remain. Frontier models solved them under an improved harness, and the filtered trajectories were used to fine-tune Qwen models. The 9B version gains 8.74 points.
RSI-Dojo-35B-A3B gains 16.87 points on PostTrainBench and reaches 29.95, similar to GLM-5.2 and GPT 5.5.
https://t.co/wOTXJwHPzm
Chat with Paper: https://t.co/oKjJdi8r40
