Researchers scale synthetic tasks for terminal agents
RST and CalibForge propose verified task-generation methods for training terminal agents.
Why it matters
The work targets a central bottleneck for AI agents: producing long-horizon, reliable training data without expensive human task authoring. Solver-calibrated and recursively verified task synthesis could make terminal-agent training more scalable and measurable.
The key points
- 1.RST generated 37,484 verified terminal-agent tasks across 15 rounds.
- 2.CalibForge used solver behavior to calibrate 5,431 learnable tasks.
- 3.Both papers focus on scalable supervision for long-horizon terminal agents.
Two new papers describe systems for generating executable, verifiable training tasks for terminal agents. Recursive Synthetic Terminal Tasks starts from verified seed tasks and recursively extends solutions, instructions and verifiers, producing 37,484 synthesized tasks at roughly $0.05 per task across 15 rounds. CalibForge generates 5,431 calibrated tasks by using solver behavior to revise candidates toward task difficulty suited for learning, with reported gains on Terminal-Bench 2.0, SWE-bench Pro and Doc2Repo.
⚡ Try this today
Read the papers before building terminal-agent training pipelines that rely on synthetic executable tasks.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
- arXiv cs.AIRecursive Synthesis for Long-Horizon Terminal TasksAug 7, 12:00 PM↗
- arXiv cs.LGRecursive Synthesis for Long-Horizon Terminal TasksAug 7, 12:00 PM↗
- arXiv cs.LGCalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal TasksAug 7, 12:00 PM↗
- arXiv cs.CLCalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal TasksAug 7, 12:00 PM↗
- HF Daily PapersCalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal TasksAug 6, 4:00 AM↗
Enjoyed this brief? Get the next one in your inbox.
More in Research
Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency
arXiv:2608.
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
arXiv:2608.
[Paper] Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iter