Recursive Synthesis Framework Generates 37,484 Terminal-Agent Tasks at $0.05 Each

A recent paper published on arXiv (2608.05466) presents Recursive Synthetic Terminal Tasks (RST), a framework designed for the scalable creation of long-horizon terminal-agent tasks through recursive verified synthesis. This approach tackles the significant expense associated with generating high-quality long-horizon training data for terminal agents, which can range from hundreds to thousands of dollars per task due to the necessity for consistency among the instruction, environment, reference solution, and verifier. Human authorship is not scalable, and using large language models (LLMs) directly often disrupts these dependencies. RST begins with verified seed tasks, enhances the reference solution, adjusts the verifier and instruction to fit the new process, tests the outcome in a new sandbox, and utilizes accepted tasks as seeds for future iterations. Over fifteen recursive rounds, RST generated 37,484 synthesized terminal-agent tasks at approximately $0.05 each, with task complexity significantly increasing over the rounds and the median reference solution length expanding. The authors of this paper are researchers, and it was categorized as 'new' upon its announcement on arXiv. This research is pertinent to AI and machine learning, especially in synthetic data generation for training agents.