Automated Synthesis of Terminal Agent Tasks
Listen to the summary
Uses a voice available on your device
Audio options
On this page
Key Takeaways
- RST generated 37,484 verified tasks over fifteen recursive rounds without manual human authoring.
- The method maintains task consistency by defining transformation contracts and using sandboxed static checks plus oracle execution.
- Fine-tuning Qwen3.5-27B and Qwen3.5-122B-A10B on synthesized data yielded significant performance gains on benchmarks like Terminal-Bench Hard.
- Qwen3.5-27B achieved a performance increase of up to 41.2 percent using PPO reinforcement learning.
Summary & Methodology Analysis
The researchers developed Recursive Synthesis, or RST, to solve the problem of scaling high-quality, long-horizon training data for agents. Instead of manual data production, RST uses a seed task from a verified pool and applies a taxonomy of rewrite operators to define new, extended task requirements. Each transformation requires a contract that defines the new evidence and requirements, ensuring the reference solution, verifier, and public instructions remain synchronized throughout the expansion. This process effectively automates the creation of complex workflows while maintaining mutual consistency between the environment and the task instruction.
Once a candidate task is generated, it is validated in a fresh sandbox environment through static checks and oracle execution. If validated, the task enters a seed pool for further recursive rounds or is used for supervised fine-tuning. For reinforcement learning, the team uses a dedicated pool of accepted tasks for verifier-based learning. This approach allows the model to learn from successful rollouts of its own synthetic tasks, which improved performance for models like Qwen3.5-27B and Qwen3.5-122B-A10B on benchmarks including Terminal-Bench 2 and Long-Horizon Terminal Bench. Fine-tuning with PPO, or Proximal Policy Optimization, a reinforcement learning algorithm that stabilizes policy updates, drove the observed 41.2 percent relative gain.
While effective, the RST pipeline has notable constraints. Because the system requires an executable proof of solvability for every generated task, the method is limited to scenarios that can be fully verified within a sandboxed terminal environment. Additionally, while the system preserves domain diversity, the authors observed a high-similarity tail in later recursive rounds, meaning task variety may eventually decrease as the synthesis proceeds. The paper does not specify the exact total compute time required for these fifteen rounds beyond the per-task cost of 0.05 dollars.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What problem does this paper address?
It addresses the difficulty and high cost of producing high-quality, long-horizon training data for terminal-based agents at scale.
Q2. How much does it cost to generate a task?
The generation cost is approximately 0.05 dollars per task.
Q3. Does this method require human input for every task?
No, the method is designed to avoid human authoring by recursively synthesizing tasks from a seed pool.
Q4. What is the role of the transformation contract?
It defines new requirements and evidence for a task, ensuring the reference solution, verifier, and public instruction remain aligned.
Q5. How are tasks validated after synthesis?
Each candidate task is validated in a fresh sandbox using static checks and oracle execution.
Q6. Which models were fine-tuned using the synthesized data?
The authors fine-tuned Qwen3.5-27B and Qwen3.5-122B-A10B.
Q7. What are the limitations regarding task types?
Synthesized tasks must have an executable proof of solvability, limiting the method to tasks verifiable in a sandboxed terminal environment.
Q8. What happens to the diversity of tasks in later rounds?
The authors acknowledge a high-similarity tail in later rounds, which they note requires future attention.
Q9. How many tasks were generated by the system?
Over fifteen recursive rounds, RST generated 37,484 verified tasks.