Recursive Synthesis for Long-Horizon Terminal Tasks

Hugging Face Daily Papers Papers
recursive-synthesis long-horizon terminal-agents synthetic-data fine-tuning reinforcement-learning

Summary

This paper introduces Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for generating long-horizon terminal-agent training data at scale, producing 37,484 tasks and improving Qwen3.5 models on terminal benchmarks.

High-quality long-horizon training data for terminal agents is expensive to produce, often costing hundreds to thousands of dollars per task, because each task must keep the instruction, environment, reference solution, and verifier mutually consistent. Human authoring does not scale, and direct generation with large language models (LLMs) often breaks these dependencies. We present Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for constructing long-horizon terminal-agent tasks at scale. 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. 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-Pro pass@4 drops from 90% at R_1 to 2.5% at R_{15}$. To demonstrate training utility, we collect rejection-sampled Qwen3.5 trajectories on the synthesized tasks and use them for supervised fine-tuning. 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\% on the three benchmarks, corresponding to relative gains of 20.0\%, 41.2\%, and 21.9\% over the base model. 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.
Original Article
View Cached Full Text

Cached at: 08/07/26, 01:54 AM

Paper page - Recursive Synthesis for Long-Horizon Terminal Tasks

Source: https://huggingface.co/papers/2608.05466 Published on Aug 5

·

Submitted byhttps://huggingface.co/zli12321

LZXon Aug 6

Authors:

,

,

,

,

,

,

,

,

,

Abstract

High-qualitylong-horizontrainingdataforterminalagentsisexpensivetoproduce,oftencostinghundredstothousandsofdollarspertask,becauseeachtaskmustkeeptheinstruction,environment,referencesolution,andverifiermutuallyconsistent.Humanauthoringdoesnotscale,anddirectgenerationwithlargelanguagemodels(LLMs)oftenbreaksthesedependencies.WepresentRecursiveSyntheticTerminalTasks(RST),arecursiveverifiedsynthesisframeworkforconstructinglong-horizonterminal-agenttasksatscale.Startingfromverifiedseedtasks,RSTextendsthereferencesolution,realignstheverifierandinstructiontothenewworkflow,validatestheresultinafreshsandbox,andreusesacceptedtasksasseedsforsubsequentrounds.Acrossfifteenrecursiverounds,RSTproduces37,484synthesizedterminal-agenttasksatroughly\0.05pertask.Taskdifficultyincreasessubstantiallyoverrounds:themedianreferencesolutiongrowsfrom67to374lines,themediannumberofexecutedcommandsgrowsfrom40to244,andDeepSeek-V4-Propass@4dropsfrom90%atR_1to2.5%atR_{15}$.Todemonstratetrainingutility,wecollectrejection-sampledQwen3.5trajectoriesonthesynthesizedtasksandusethemforsupervisedfine-tuning.Fine-tuningonthesetrajectoriesimprovesQwen3.5-27BandQwen3.5-122B-A10Bbyupto10pointsonTerminal-Bench~2,Terminal-BenchHard,andLong-HorizonTerminalBench,whileagenticPPOliftsQwen3.5-27Bto49.44\%,32.00\%,and22.07\%onthethreebenchmarks,correspondingtorelativegainsof20.0\%,41.2\%,and21.9\%overthebasemodel.Moreover,after15rounds,therecursionshowsnoceiling:synthesisyieldandvalidationratesremainstableasdifficultykeepsclimbing,indicatingthattheprocesscancontinuewellbeyondthescalereportedhere.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2608\.05466

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.05466 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.05466 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.05466 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

@DeFiMinty: Can AI systems keep generating harder training tasks for AI agents? New research from Tencent suggests they can. Recurs…

X AI KOLs Following

New research from Tencent introduces Recursive Synthetic Terminal Tasks (RST), a method that progressively generates harder, verifiable training tasks for AI agents. Starting from 639 tasks, it produced 37,484 verified tasks across 15 rounds, and reinforcement learning with these tasks improved Qwen3.5-27B from 22.7% to 32.0% on Terminal-Bench Hard.

CLI-Universe: Towards Verifiable Task Synthesis Engine for Terminal Agents

Hugging Face Daily Papers

CLI-Universe is a synthesis engine that generates verifiable terminal-agent tasks via multi-dimensional capability taxonomy and evidence-guided research, producing a distilled dataset of 6,000 trajectories. Fine-tuning Qwen3-32B on this dataset achieves 33.4% on Terminal-Bench 2.0, setting a new state-of-the-art for open-source models at or below 32B parameters.

Terminal-World: Scaling Terminal-Agent Environments via Agent Skills

arXiv cs.CL

Terminal-World introduces a fully automated pipeline that uses agent skills to synthesize high-quality training data for terminal agents, enabling models to outperform baselines with only 1.2% of the training data. The method co-derives task instructions, environments, and teacher trajectories from skill primitives.