Auto Research Loop with Specialist Agents Develops Training Recipes

A research paper published on arXiv (2605.05724) presents an automated research framework that operates as a closed empirical loop influenced by external measurements. In each experiment, a hypothesis is proposed along with an editable code, a result owned by the evaluator, and feedback for future proposals. The result is a traceable record of proposals, code modifications, experiments, scores, and failure classifications. Specialized agents divide recipe surfaces and disseminate measured lineage throughout the trials. This lineage feedback allows agents to convert evaluator results—such as crashes, budget excesses, size failures, and accuracy misses—into modifications at the program level. In total, there were 1,197 headline-run trials and 600 Parameter Golf control trials, with no human intervention in proposals, edits, scores, or failures post-setup.