RSI-Dojo: Training LLM Agents to Automate Post-Training, Matching Frontier Models
dair_ai · x · 2026-11-04
A paper from CUHK, ByteDance, Tsinghua and SJTU introduces RSI-Dojo, teaching LLM agents to run model post-training themselves — a step toward recursive self-improvement.
- RSI-Craft pipeline converts research papers into executable training environments (datasets, metrics, eval code); 281 post-training tasks survive reliability/difficulty filtering.
- Frontier-model trajectories are collected with an improved harness and used to fine-tune Qwen models.
- RSI-Dojo-9B gains +8.74 on PostTrainBench; RSI-Dojo-35B-A3B gains +16.87 to 29.95 and hits 53.29 on PostTrainBench-Lite, matching GLM-5.2 and GPT 5.5.
Unlike prior harness/workflow-level RSI work, this internalizes post-training capability into the model itself.
More from coding & agent
- Veteran agent builder distrusts auto-compaction, rolls his own optimized pipeline — omarsar0 · 2026-11-04
- quiet.md open-sources a Rust+GPUI local notes app with Feishu DocxXML compatibility — vista8 · 2026-11-04
- YC-backed Verno launches AI that edits and error-checks industrial P&ID engineering documents — ycombinator · 2026-11-04
- Personal agents' real value is as a context buffer, not prompt engineering — andreisavu · 2026-11-04
- ClickUp rolls out Brain² browser use for sites without official integrations — testingcatalog · 2026-11-04
- Cerebral Valley teams with Meta and Google DeepMind for SF hackathon with $11,000 in prizes — jocarrasqueira · 2026-11-04