ParetoRSI-Gym evaluates multi-domain recursive self-improvement
Original title: ParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?
Claude Code with Opus-5 raised equal-weighted generalization scores 9.2% relative, or 2.3 absolute points, over the starting models.
ParetoRSI-Gym evaluates recursive self-improvement when an agent searches for data, curates it, trains, reads per-domain results and repeats. The initial LLM, five domains (math, code, finance, medicine and science), toolset and SFT pipeline stay fixed, while generalization benchmarks separate capability gains from overfitting. Across three agent systems, three base models, five rounds and multiple seeds, Claude Code with Opus-5 is the strongest system. Further analysis says the harness matters as much as the model and that each agent keeps a distinct data-selection strategy.
Unconfirmed: The 9.2% relative and 63% single-domain figures are the authors' reported results, not an independent evaluation.
- The reported gain is averaged equally over three base models and five domains.
- Medicine on Ministral-3-3B shows a single-domain relative gain of up to 63%.
- The score is for the system's proxy-best checkpoint versus the starting model.
- Only the training data changes inside each fixed evaluation loop.
- Products
- Claude Code
- AI models
- Opus-5Ministral-3-3B
From the original article
1 min read at huggingface.co
Recursive self-improvement (RSI) requires an agentic system to design, train, and evaluate its own successor. Data-driven post-training is a central part of that loop, yet existing evaluations measure it on a single benchmark, where an agent selects from a static pool or distills a stronger model to push one score up in one run.
By Haojin Wang, Yangzhen Wu, Zeyi Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.