Researchers introduce RSI-Dojo to train LLM agents for post-training
A new paper from arXiv introduces RSI-Dojo, a collection of 281 executable model training environments designed to teach large language model agents how to conduct post-training autonomously. The work addresses a gap in Recursive Self-Improvement research, where previous studies focused on workflow optimization rather than embedding the ability to train models directly into the agent itself.
Key points
- RSI-Dojo provides 281 executable environments for teaching LLM agents to perform post-training tasks.
- Fine-tuned Qwen models improved by 8.74 and 16.87 points on PostTrainBench using RSI-Dojo trajectories.
- RSI-Dojo-35B-A3B matched frontier models like GPT 5.5 on PostTrainBench-Lite with a score of 53.29.
To build this environment, the authors developed RSI-Craft, an automatic pipeline that converts human research artifacts into executable tasks. This pipeline collects papers from top venues, standardizes datasets and metrics, and filters tasks based on data size and reliability. The team then used frontier models to collect high-quality training trajectories, which were refined through rubric-based and outcome-based filtering.
The researchers fine-tuned Qwen3.5-9B and Qwen3.6-35B-A3B on these trajectories. The resulting models, RSI-Dojo-9B and RSI-Dojo-35B-A3B, showed improvements of 8.74 and 16.87 points on PostTrainBench, respectively. Notably, the 35B variant reached a score of 29.95 on PostTrainBench and 53.29 on PostTrainBench-Lite, matching the performance of frontier models like GLM-5.2 and GPT 5.5 in this specific task.
The story so far
3 episodes →- Researchers introduce RSI-Dojo to train LLM agents for post-trainingthis story
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Tencent releases WeVisOne-VL, a 2B/4B open vision model for detection, OCR and matting · 1 src
- Anthropic releases FlashProve open-source system for provable AI inference · 6 src
- Lama-dt outperforms baselines on meta-world and mujoco · 1 src
- SLATE benchmark tests long-horizon API selection in e-commerce automation · 1 src
- Researchers compare two fine-tuning methods for small medical AI models · 1 src
Comments
via GitHub Discussions