Hugging Face researchers have introduced Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework designed to generate long-horizon terminal-agent training data at scale. The approach addresses the high cost of manually creating such data, which typically ranges from hundreds to thousands of dollars per task due to the need for consistency across instructions, environments, reference solutions, and verifiers.
RST starts from verified seed tasks and iteratively extends the reference solution, realigns the verifier and instruction, validates the result in a fresh sandbox, and reuses accepted tasks as seeds for subsequent rounds. Over fifteen recursive rounds, the framework produced 37,484 synthesized terminal-agent tasks at an estimated cost of $0.05 per task.
Task difficulty increased substantially with each round: the median reference solution grew from 67 to 374 lines, the median number of executed commands rose from 40 to 244, and the pass@4 rate of DeepSeek-V4-Pro dropped from 90% at round 1 to 2.5% at round 15, indicating progressively harder tasks.
To demonstrate training utility, the team collected rejection-sampled trajectories from Qwen3.5 models on the synthesized tasks and used them for supervised fine-tuning. This improved Qwen3.5-27B and Qwen3.5-122B-A10B by up to 10 points on Terminal-Bench 2, Terminal-Bench Hard, and Long-Horizon Terminal Bench. Furthermore, agentic PPO training lifted Qwen3.5-27B to 49.44%, 32.00%, and 22.07% on the three benchmarks, representing relative gains of 20.0%, 41.2%, and 21.9% over the base model.
Notably, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable while difficulty continues to climb, suggesting the process can be extended well beyond the reported scale. The dataset is available via a Hugging Face collection.