ParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?

The researchers propose the ParetoRSI-Gym benchmark to evaluate the ability of AI agents to improve language models across domains through data-driven post-training, serving the evaluation of recursive self-improvement (RSI) systems. Agents cyclically perform data search, filtering, and training across five domains—mathematics, code, finance, medicine, and science—and read the results in each domain; apart from the data, the initial LLM, the five domains, the toolset, and the SFT pipeline are all fixed, and domain generalization benchmarks are used to distinguish capability improvement from overfitting to the optimized scores. Across three agent systems, three base models, five rounds of experiments, and multiple random seeds, the strongest Claude Code with Opus-5 improves the generalization score of the best checkpoint selected by the proxy metric by 9.2% relative to the starting model (an absolute improvement of 2.3 percentage points, equally averaged over the three base models and five domains), with a maximum single-domain relative improvement of 63% (medicine on Ministral-3-3B).