# Researchers release RSI-Dojo environments for training agents to run post-training

> RSI-Dojo is a set of executable environments built from published research, and fine-tuning Qwen models on its trajectories gains 8.74 and 16.87 PostTrainBench points.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Research
- Published: 2026-11-08T16:21:29+00:00
- Updated: 2026-11-08T17:09:43.801719+00:00
- Web page: https://ai-news.ikangai.com/story/researchers-release-rsi-dojo-environments-for-training-agents-to-run-post-traini/
- Seen in: DAIR.AI (Elvis Saravia)
- Change: capability (New capability)
- Applies to: Qwen3.5-9B, Qwen3.6-35B-A3B, RSI-Dojo-35B-A3B

## What changed

Researchers release RSI-Dojo, executable environments built from published research, and fine-tuning Qwen models on their trajectories produces RSI-Dojo-35B-A3B and tuned Qwen3.5-9B.

Researchers from CUHK, ByteDance, Tsinghua University and Shanghai Jiao Tong University have released RSI-Dojo, a collection of executable environments. They are meant to teach LLM agents to run post-training on models themselves. The authors report that fine-tuning Qwen3.5-9B and Qwen3.6-35B-A3B on trajectories from these environments improves them by 8.74 and 16.87 PostTrainBench points over their base models. The larger tuned model, RSI-Dojo-35B-A3B, reaches 29.95 on PostTrainBench and 53.29 on PostTrainBench-Lite. The authors say this matches frontier models such as GLM-5.2 and GPT 5.5.

The authors' premise is that most automated model-training work has targeted workflows or harnesses around the model. It has not targeted the model's own ability to conduct post-training. To teach that ability, they argue, one needs executable training environments. Those environments produce trajectories showing how to select data, train models, diagnose failures and iterate. RSI-Dojo is meant to supply those environments. The source of the tasks is RSI-Craft, a pipeline. It collects research papers from top venues, puts their datasets and metrics behind a unified interface and reproduces their evaluation dependencies. Tasks are then filtered on data size, execution reliability and discriminability, which is the ability to separate strong agents from weak ones. That leaves 281 seed tasks across math, writing, code, medical, knowledge and reasoning.

The training data comes from frontier models that attempt these tasks under an improved harness. The authors keep only trajectories that pass rubric-based and outcome-based filters.

The comparison with GLM-5.2 and GPT 5.5 needs careful reading. The authors say RSI-Dojo-35B-A3B matches those models, and the paper's own abstract uses the word matching. The DAIR.AI summary of the paper says similar to. Both describe the same numbers, so the gap between them is one of wording rather than data. The DAIR.AI summary also notes the scope: its tasks follow the PostTrainBench style, and longer-horizon research settings are left to future work.

The authors frame the work as a step toward recursive self-improvement, the idea that AI systems can build and improve AI systems. Their argument is that such a system needs agents able to train models. They say this skill can be placed in the weights of a mid-sized open model rather than only in the surrounding scaffolding. The method also offers a repeatable source of new training tasks, since published research can be converted into environments.

## Why it matters

The work moves the ability to train models into the weights of a mid-sized open model. Converting research into environments gives a repeatable source of training tasks.

## Original source

- [(2611.03698) Training LLM Agents to Automate LLM Post-Training with RSI-Dojo](https://arxiv.org/abs/2611.03698)
- [Training LLM Agents to Automate LLM Post-Training with RSI-Dojo | DAIR.AI Academy | DAIR.AI Academy](https://academy.dair.ai/papers/training-llm-agents-to-automate-llm-post-training-with-rsi-dojo-2611.03698)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Researchers release RSI-Dojo environments for training agents to run post-training", 2026-11-08, https://ai-news.ikangai.com/story/researchers-release-rsi-dojo-environments-for-training-agents-to-run-post-traini/
