cs.AIOct 10, 2026

WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions

Authors: Junyi Zhou, Yufei Wang, Qiyuan Zhang, Chuhan Wu, Zhengxi Li, Wentao Zhao, Jie Zhou, Le Tian, 

Organizations: City University of Hong Kong · WeChat AI, Tencent · MBZUAI

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Developing terminal-based agents for long-horizon execution demands synthetic tasks that tightly couple complex instructions with authentic, interactive environments. However, current automated synthesis suffers from a fundamental asymmetry: instruction-oriented methods fabricate low-fidelity, synthetic environments, whereas resource-oriented methods treat real artifacts merely as static inputs, degrading task instructions into shallow, trivial operations. To address this limitation, we introduce WorkBorne, a synthesis framework that uses artifact transitions as the central primitive for modeling the full life-cycle of agentic tasks. In WorkBorne, each transition specifies who performs the task, which artifacts anchor the initial state and intended outcome, and the corresponding relationship between them. Specifically, WorkBorne first constructs coherent and diverse artifact corpora from iterative web retrieval, then discovers non-trivial transitions by identifying meaningful input-output artifact assemblies and the relations that connect them. Each discovered transition is finally instantiated into grounded agentic tasks. We construct 1,061 artifact corpora and 7,882 task environments, yielding 15,485 high-quality trajectories, with 53.6% exceeding 30 assistant turns. Experiments on Qwen3.5-27B and Qwen3.5-35B-A3B show that WorkBorne substantially improves over base models (i.e., improving Terminal-Bench 2.1 from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2) and achieves the strongest overall performance among other synthesis baselines (e.g., TMax (Ivison et al., 2026), RST (Li et al., 2026a), Terminal-Universe (Wu et al., 2026a)). Further analysis shows that transition is strongly associated with execution depth, while corpus breadth and transition depth provide complementary scaling gains.

Figures & tables

Figure 1: Overview of WorkBorne. WorkBorne models agentic tasks as artifact transitions T=(p,Ain,Aout,r)\mathcal{T}=(p,\mathcal{A}^{\mathrm{in}},\mathcal{A}^{\mathrm{out}},r) and synthesizes them in three stages: (1) Artifact Corpus Construction: building real-world artifact pools; (2) Transition Discovery: discovering grounded transitions through assembly, planning, and validation; and (3) Task Instantiation: compiling transitions into aligned instructions, executable environments, and evaluation rubrics.
Figure 2
Table 2: Main results on three agentic benchmarks: Terminal-Bench Lite, Terminal-Bench 2.1, and ClawEval. We report mean raw reward for TB Lite, pass@1 for TB 2.1, and pass@3, pass^3, and average score for ClawEval. Our Terminal-Bench scores are averaged over three runs. The best result among synthesis methods is shown in bold , and the second-best result is underlined . Green values in parentheses indicate absolute improvements over the corresponding Qwen3.5 base model. ∗\ast marks scores we measure on the released models under the configurations in Appendix D ; other scores are taken from the corresponding reports.
 EnvironmentTrajectoryPerformance
MethodFile countFile typesRubricsAssistant turnsTool callsLong-horizon rate (%)TB LiteTB 2.1
Instruction-first10.85.59.125.935.827.468.347.2
Input-only283.215.57.627.836.633.069.048.3
WorkBorne347.318.310.644.557.566.671.151.7
Table 3: Environment, trajectory, and performance comparison across construction pipelines. Each pipeline contributes 1,000 environments and 1,000 training trajectories. Counts are means over the sampled environments and retained trajectories; long-horizon rate is the percentage of retained trajectories with >30>30 assistant turns.
Factorη\eta
Input artifact count0.049
Total artifact count0.015
Input artifact form0.183
Output anchor form0.306
Artifact assembly0.314
Transition relation type0.140
Table 4: Association with execution depth. Larger η\eta indicates stronger association.
Figure 3: Transition structure corresponds to different execution horizons. Left: selected working input and output anchor assemblies. Right: an expansion of the highlighted repository-to-repository assembly, comparing four transition relations while holding the assembly fixed. Bars show the percentage of tasks whose median execution exceeds 30 assistant turns; the adjacent column reports the median across tasks, rounded to the nearest turn.
Figure 4: Both scaling paths improve performance. Transition scaling holds corpus coverage fixed, while corpus scaling holds transitions per corpus fixed. The horizontal axis shows the number of training trajectories. The vertical axis shows mean raw reward for TB Lite and pass@1 for TB 2.1, both expressed as percentages and averaged over three runs. Point labels denote C×TC\times T , where CC is the number of corpora and TT is the number of transitions per corpus; the two paths share the 1,000×81{,}000\times 8 endpoint. * After failure filtering, this configuration contains slightly fewer than 8,000 tasks ( <2%<2\% shortfall); we fill the remaining trajectory budget with additional rollouts of retained tasks under different harnesses.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

HarnessTrajectoriesShare (%)
Mini-SWE1,61410.4
Stirrup1,4379.3
OpenHands1,4219.2
Terminus21,2087.8
Hermes1,1907.7
OpenClaw1,1317.3
Appendix
Table 5: Trajectory statistics by harness in the final training set. Shares use all 15,485 training trajectories.
  Task overlaps 
BenchmarkInstructionsTraining instructionsTask metadataTrajectoriesMax. Jaccard
TB Lite1000000.0067
TB 2.1890000.0074
ClawEval general1610000.0122
Appendix
Table 6: Contamination checks against evaluation instructions. Task overlaps are assessed through 13-gram matching and review. Jaccard similarity uses normalized 5-gram sets of task instructions. Text is normalized with Unicode NFKC and case folding, using alphanumeric word units and individual Han characters. Instructions shorter than 13 units are checked by full normalized instruction containment.
SettingValue
Hardware128 H20 GPUs
Tensor / pipeline / context / data parallelism8 / 1 / 1 / 16
Sequence parallelismEnabled
Micro / global batch size1 / 32
Gradient accumulation steps2
Training epochs3
Appendix
Table 7: Shared optimization, parallelism, and execution settings for both base models.
SettingValue
Expert parallelism8
Expert tensor parallelism1
Router auxiliary loss coefficient10−610^{-6}
Token dispatcheralltoall
Grouped GEMMEnabled
Permutation fusionEnabled
Appendix
Table 8: Expert parallelism and MoE settings for Qwen3.5-35B-A3B.
SettingTerminal-BenchClawEval (General)
Concurrency2540
Context window (tokens)131,072262,144
Thinking budget (tokens)32,76832,768
Temperature1.00.6
Top- pp0.950.95
Appendix
Table 9: Evaluation settings for Terminal-Bench and ClawEval.
MetricInstruction-firstInput-onlyWorkBorne
Task environments
File count10.8283.2347.3
File types5.515.518.3
Directories4.866.985.3
Maximum path depth1.44.85.6
File categories3.84.75.2
Appendix
Table 10: Full construction comparison. Values are arithmetic means except the long-horizon rate, which gives the percentage of retained trajectories exceeding 30 assistant turns.
MethodTask Complexity ↑\uparrowWorkspace Realism & Coherence ↑\uparrowInstruction–Environment Alignment ↑\uparrow
Instruction-first2.612.844.08
Input-only2.944.173.71
WorkBorne4.214.464.35
Appendix
Table 11: Human evaluation of synthesized tasks across construction pipelines. Each method is evaluated on 50 tasks matched by the same underlying skills. Higher is better.
Rubric Quality DimensionMean Score ↑\uparrowRated ≥4\geq 4 (%) ↑\uparrow
Requirement Coverage4.3688.0
Faithfulness & Correctness4.4892.0
Verifiability4.2284.0
Average4.3588.0
Appendix
Table 12: Human audit of evaluation rubrics for the 50 sampled WorkBorne tasks. Higher is better.
Figure 5: Reusable corpora support varied task transformations and workspaces. (a) Distributions of task counts and distinct transition types per corpus. (b) The five most frequent transition types cover 77.2% of tasks. (c) File composition counts occurrences across task environments. (d) Task coverage counts environments containing each file category, so categories overlap. Source code contributes the largest file share, while documents occur in nearly every workspace.
Figure 6: Artifact forms in working inputs and output anchors. Bars show the percentage of tasks containing each artifact form based on source metadata; categories overlap.
GroupTasksMedian turns>> 30>> 50
Input / output anchor assemblies
Document / Document5312946.0%14.1%
Document / Repository7733767.5%28.1%
Repository / Repository5284378.8%38.4%
Document + repository / Repository2404475.4%38.8%
Transition relations within repository / repository
Appendix
Table 13: Execution depth by transition structure. Threshold columns give percentages of tasks exceeding each median turn count. Assemblies list inputs / output anchors; transition groups use five selected raw labels within the repository-to-repository assembly. Median turns are rounded to the nearest integer.
More in this topic

LLM Agent Trajectory Synthesis

LLM Agent Trajectory Synthesis creates and validates interaction sequences to train and improve language-model agents.

See all 69 papers

Explore similar work

CardsList
Jul 30, 2026

Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training

Training terminal agents at scale requires diverse, verifiable terminal tasks and high-quality interaction trajectories, yet acquiring such data remains a significant challenge. Existing synthesis methods face two key limitations: (1) weak reliability caused by the disconnect between task generation and real execution, and (2) limited diversity and scalability due to dependence on existing repositories. We propose Meta-Task, a framework that redefines terminal task synthesis as a Terminal-Bench-format task itself: an agent operates within a real container environment to iteratively generate, execute, and verify tasks, so that synthesized components are checked for internal consistency and executability within the generation loop itself. Building upon this, we decouple the target task requirements along multiple dimensions, introduce a multi-phase mechanism that dynamically designs novel task specifications before producing the actual tasks, and incorporate optional external material support to enhance diversity and realism. We additionally apply LLM-as-Judge filtering to ensure the quality of the final training data. Experiments on Terminal-Bench 2.0 show that fine-tuning on only 3,221 Meta-Task synthesized trajectories achieves 22.5% and 31.8% Avg Pass@1 for Qwen3-14B and Qwen3-32B respectively, outperforming concurrent approaches with significantly less training data.

Zhihong Pan, Jiyuan He, Kai Zhang +5

State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China · Meituan

Synthetic Data GenerationTerminal AgentsLLM Agent Trajectory SynthesisLLM Agent Training
AskPDF
Sep 30, 2026

WorkGenesis: Building the Worlds That Teach Agents to Work

The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them; and (2) Execution-Guided Consistency Verification, which renders a reference deliverable inside the constructed work, attributes every unsatisfied rubric item to the agent, the task, or the rubric, and uses task and rubric defects as feedback to iteratively repair the work until it passes the audit. Experimental results demonstrate that Fx-Work-35B, trained with simple supervised fine-tuning (SFT) on only 20K units of work synthesized by WorkGenesis, achieves the highest scores among all comparable-scale baselines on the five reported metrics across GDPvalAA-v2, APEX-Agents-AA, and JobBench (31.00 versus 24.79 average score), and even surpasses frontier models such as the 1.6T DeepSeek-V4-Pro-Preview. These results show that WorkGenesis provides scalable training data for working agents.

Xinyu Zhu, Fenyi Liu, Yuzhu Cai +4

Shanghai Jiao Tong University · Endless Frontier · Beihang University

LLM Agent TrainingSynthetic Data Generation for Language Models
AskPDF
Aug 5, 2026

Recursive Synthesis for Long-Horizon Terminal Tasks

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 R1 to 2.5% at R15. 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.

Zhongzhi Li, Yucheng Shi, Zongxia Li +8

Tencent HY LLM Frontier · University of Georgia · University of Maryland, College Park +5

Terminal AgentsSynthetic Benchmark GenerationLLM Agent TrainingLong-Horizon Agent Tasks
AskPDF