Li, Zhongzhi, Shi, Yucheng, Li, Zongxia, Wang, Ruhan, Li, Anhao, Li, Jiaxi, Yang, Junyao, Ke, Lei, Huang, Zixun, Wang, Yu, Huang, Jingyuan, Liu, Ninghao, Mi, Haitao, Liang, Leowei · Aug 5, 2026
Recursive Synthesis for Long-Horizon Terminal Tasks
Markdown · JSON · Original source
The paper’s authors present Recursive Synthetic Terminal Tasks (RST), a method for generating training tasks for terminal agents: software agents that carry out sequences of commands in a terminal. RST starts with verified tasks, extends their reference solutions, adjusts the instructions and verifiers, and checks each new task in a fresh sandbox before using accepted tasks as seeds for another round. The authors report that, across fifteen rounds, it generated 37,484 tasks at roughly $0.05 each, while task difficulty increased and DeepSeek-V4-Pro pass@4 fell from 90% at R1 to 2.5% at R15. They also report that supervised on trajectories from these tasks improved Qwen3.5 models by up to 10 points on three ; they report PPO scores for Qwen3.5-27B as well. The authors say synthesis yield and validation rates remained stable through the fifteenth round, despite rising difficulty, and argue that the process can continue beyond the scale reported. The results suggest a way to produce large amounts of increasingly difficult terminal-agent training data, but the supplied abstract does not describe the evaluation setup in enough detail to assess how these benchmark results would transfer to other tasks or settings.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Aug 5, 2026 (2026-08-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T23:35:43.195Z. First observed: 2026-11-09T05:04:01.181Z. Retrieved: 2026-11-12T23:34:11.957Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper’s abstract, which does not provide the evaluation setup needed to assess how the reported results transfer to other tasks or settings. The claim that recursion can continue beyond the reported scale is the authors’ interpretation of results through fifteen rounds; the abstract does not report results from later rounds.
Open question
Do synthesis yield and validation rates remain stable beyond the fifteen rounds reported by the authors?
Source-stated
The authors introduce RST, which recursively extends verified terminal-agent tasks and checks the resulting tasks before reusing accepted ones.
Starting from verified seed tasks, RST extends the reference solution, realigns the verifier and instruction to the new workflow, validates the result in a fresh sandbox, and reuses accepted tasks as seeds for subsequent rounds.
Source-stated
The authors report generating 37,484 tasks across fifteen rounds at roughly $0.05 per task, with increasing task difficulty and lower DeepSeek-V4-Pro pass@4.
Across fifteen recursive rounds, RST produces 37,484 synthesized terminal-agent tasks at roughly $0.05 per task. Task difficulty increases substantially over rounds: the median reference solution grows from 67 to 374 lines, the median number of executed commands grows from 40 to 244, and DeepSeek-V4
Source-stated
The authors report that training on generated-task trajectories improved Qwen3.5 benchmark results, and that agentic PPO raised Qwen3.5-27B scores on three benchmarks.
Fine-tuning on these trajectories improves 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, while agentic PPO lifts Qwen3.5-27B to 49.44%, 32.00%, and 22.07%
Source-stated
The authors report stable synthesis yield and validation rates through fifteen rounds, and say the process can continue beyond the scale described.
Moreover, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable as difficulty keeps climbing, indicating that the process can continue well beyond the scale reported here.
Source and contributor context
Li, Zhongzhi, Shi, Yucheng, Li, Zongxia, Wang, Ruhan, Li, Anhao, Li, Jiaxi, Yang, Junyao, Ke, Lei, Huang, Zixun, Wang, Yu, Huang, Jingyuan, Liu, Ninghao, Mi, Haitao, Liang, Leowei · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 359385cc-0159-47b8-8415-e1183758f2b8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T23:35:43.195Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 72199135193793b7f79e3c35b90efa4215935a0bc268d4c0e4cb8f23d5bac7b3. Source text hash: 2e0e89b2a767005fae08b3192985475118bc24c5c73f5c0f07f0bc7d102d3ae8.
Revision and source-check history
Version 1 · 2026-11-12T23:35:43.195Z
Recursive Synthesis for Long-Horizon Terminal Tasks
The paper’s authors present Recursive Synthetic Terminal Tasks (RST), a method for generating training tasks for terminal agents: software agents that carry out sequences of commands in a terminal. RST starts with verified tasks, extends their reference solutions, adjusts the instructions and verifiers, and checks each new task in a fresh sandbox before using accepted tasks as seeds for another round. The authors report that, across fifteen rounds, it generated 37,484 tasks at roughly $0.05 each, while task difficulty increased and DeepSeek-V4-Pro pass@4 fell from 90% at R1 to 2.5% at R15. They also report that supervised on trajectories from these tasks improved Qwen3.5 models by up to 10 points on three ; they report PPO scores for Qwen3.5-27B as well. The authors say synthesis yield and validation rates remained stable through the fifteenth round, despite rising difficulty, and argue that the process can continue beyond the scale reported. The results suggest a way to produce large amounts of increasingly difficult terminal-agent training data, but the supplied abstract does not describe the evaluation setup in enough detail to assess how these benchmark results would transfer to other tasks or settings.
Original source
cs.AI · Article
Source details
Original source date: Aug 5, 2026 (2026-08-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T23:35:43.195Z. First observed: 2026-11-09T05:04:01.181Z. Retrieved: 2026-11-12T23:34:11.957Z.
Limits
The supplied evidence is the paper’s abstract, which does not provide the evaluation setup needed to assess how the reported results transfer to other tasks or settings. The claim that recursion can continue beyond the reported scale is the authors’ interpretation of results through fifteen rounds; the abstract does not report results from later rounds.
Open question
Do synthesis yield and validation rates remain stable beyond the fifteen rounds reported by the authors?
Source-stated
The authors introduce RST, which recursively extends verified terminal-agent tasks and checks the resulting tasks before reusing accepted ones.
Starting from verified seed tasks, RST extends the reference solution, realigns the verifier and instruction to the new workflow, validates the result in a fresh sandbox, and reuses accepted tasks as seeds for subsequent rounds.
Source-stated
The authors report generating 37,484 tasks across fifteen rounds at roughly $0.05 per task, with increasing task difficulty and lower DeepSeek-V4-Pro pass@4.
Across fifteen recursive rounds, RST produces 37,484 synthesized terminal-agent tasks at roughly $0.05 per task. Task difficulty increases substantially over rounds: the median reference solution grows from 67 to 374 lines, the median number of executed commands grows from 40 to 244, and DeepSeek-V4
Source-stated
The authors report that training on generated-task trajectories improved Qwen3.5 benchmark results, and that agentic PPO raised Qwen3.5-27B scores on three benchmarks.
Fine-tuning on these trajectories improves 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, while agentic PPO lifts Qwen3.5-27B to 49.44%, 32.00%, and 22.07%
Source-stated
The authors report stable synthesis yield and validation rates through fifteen rounds, and say the process can continue beyond the scale described.
Moreover, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable as difficulty keeps climbing, indicating that the process can continue well beyond the scale reported here.
Source and contributor context
Li, Zhongzhi, Shi, Yucheng, Li, Zongxia, Wang, Ruhan, Li, Anhao, Li, Jiaxi, Yang, Junyao, Ke, Lei, Huang, Zixun, Wang, Yu, Huang, Jingyuan, Liu, Ninghao, Mi, Haitao, Liang, Leowei · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 359385cc-0159-47b8-8415-e1183758f2b8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T23:35:43.195Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 72199135193793b7f79e3c35b90efa4215935a0bc268d4c0e4cb8f23d5bac7b3. Source text hash: 2e0e89b2a767005fae08b3192985475118bc24c5c73f5c0f07f0bc7d102d3ae8.
uncertain · 2026-11-12T23:35:43.195Z · anthropic/claude-opus-5.5
Publication: 359385cc-0159-47b8-8415-e1183758f2b8
attribution — pass: The draft consistently attributes RST, its results and its scalability claim to the paper's authors. It also keeps the no-ceiling statement framed as the authors' argument, not as established fact.
We present Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for constructing long-horizon terminal-agent tasks at scale.
summary — uncertain: The draft explains terminal agents well, but central jargon is unexplained for a general reader. 'pass@4' and 'R1'/'R15' carry the main difficulty claim. Repair: 'share of tasks solved within four attempts, from round 1 to round 15.' 'They report agentic PPO scores as well' omits the actual result: 49.44%, 32.00% and 22.07%, or relative gains of 20.0%, 41.2% and 21.9%. It should state at least the relative gains. The authors' force is also softened: 'shows no ceiling' and 'well beyond' become 'beyond the scale reported.'
agentic PPO lifts Qwen3.5-27B to 49.44%, 32.00%, and 22.07% on the three benchmarks, corresponding to relative gains of 20.0%, 41.2%, and 21.9% over the base model
date — pass: The displayed original publication date of 2026-08-05 matches the v1 submission date. It does not use the v4 revision date of 5 Nov 2026 or the November feed date.
[Submitted on 5 Aug 2026 (v1), last revised 5 Nov 2026 (this version, v4)]
support — pass: Each claim matches the abstract. The closing sentence uses 'suggest' and acknowledges that the abstract cannot establish transfer to other settings. The question, which asks whether stability holds beyond 15 rounds, follows directly from the authors' extrapolation without presuming an answer. The limitations correctly note that later rounds were not reported.
Moreover, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable as difficulty keeps climbing
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Further training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference