ParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?
The strongest agent system improved held-out scores by 9.2% relative to the starting models, or 2.3 absolute points.
ParetoRSI-Gym lets agents find and curate training data, train models, inspect per-domain results and repeat across math, code, finance, medicine and science. The initial models, tools and training pipeline stay fixed, while separate generalization benchmarks check whether gains extend beyond the scores agents optimize. Across five rounds, Claude Code with Opus-5 performed best; its reported gain averages the checkpoint selected by proxy scores equally across three base models and five domains. The authors also found that the agent harness mattered as much as the model driving it.
HF Daily Papers' note
ParetoRSI-Gym lets agents find and curate training data, train models, inspect per-domain results and repeat across math, code, finance, medicine and science. The initial models, tools and training pipeline stay fixed, while separate generalization benchmarks check whether gains extend beyond the scores agents optimize. Across five rounds, Claude Code with Opus-5 performed best; its reported gain averages the checkpoint selected by proxy scores equally across three base models and five domains. The authors also found that the agent harness mattered as much as the model driving it.
HF Daily Papers' note
score 5