Scaling Verifiable Environments for Long-horizon Work Agents

Jiazheng Zhang; Long Ma; Yunxian Yang; Zhiheng Xi; Zhikai Lei; Yajie Yang; Chenyang Liao; Enyu Zhou

cs.CLOct 4, 2026

Scaling Verifiable Environments for Long-horizon Work Agents

Authors: Jiazheng Zhang, Long Ma, Yunxian Yang, Zhiheng Xi, Zhikai Lei, Yajie Yang, Chenyang Liao, Enyu Zhou, +18 more

Abstract

Work agents operate over digital artifacts to execute professional knowledge-intensive work, requiring training environments that support long-horizon interaction and trustworthy verification. However, hand-crafted environments incur prohibitive engineering overhead that prevents environment scaling, whereas synthesis methods sacrifice workspace complexity, realism, or grounded verifiability. To bridge this gap, we introduce WorkForge, a scalable synthesis framework for constructing verifiable work-agent environments from real-world resources. Starting from expert workflows, WorkForge first identifies the resources, decisions, and deliverables required by each workflow. It then retrieves relevant real-world files and organizes them into a workspace. WorkForge inspects the workspace to extract concrete, checkable facts about its content. These factual anchors fix which task types the workspace can support and how their outcomes can be verified. Therefore, WorkForge derives each task's instructions, solution plan, and complementary programmatic and semantic verifiers directly from these factual anchors, keeping verification traceable to observable workspace evidence. Furthermore, we construct 16.7K verifiable environments across 40 professional domains, with workspaces collectively covering 60 file types. Post-training Qwen3.5-35B-A3B-Base improves GDPVal from 45.5 to 73.6 and APEX Score from 5.0 to 21.3, while enabling Qwen3.5-27B to achieve highly competitive performance and outperform strong competitors. Our analyses confirm the efficacy of the proposed method and reveal consistent scaling behaviors across both data volume and interaction horizons.

Figures & tables

Figure 1: Comparison of synthetic and realistic workspaces. Realistic workspaces are substantially more complex: they contain more files, span a broader range of file types, and distribute task-relevant evidence throughout the directory hierarchy. In contrast, synthetic workspaces are highly curated, with most task-relevant information concentrated in a small set of identifiable files.

Figure 2: Overview of WorkForge pipeline. Conditioned on professional scene blueprints, WorkForge acquires and instantiates real-world resources into authentic workspaces, derives tasks and complementary verifiers, and collects scored trajectories through interaction.

Frontier Proprietary Models
Model	WildClaw		ClawEval		GDPval	APEX		ALE
	Score	Pass@1	Score	Passˆ3	Score	Score	Pass@1	Score	Pass@1
Claude Opus 5.0	64.5	32.4	82.3	74.2	91.0	58.5	42.7	48.3	26.7
Claude Opus 4.8	64.0	30.5	81.5	72.1	89.1	46.2	34.0	46.2	24.8
GPT-5.6 Sol	53.5	29.5	77.3	68.9	90.1	51.8	37.5	48.3	26.7
Gemini 3.7 Flash	52.6	18.1	69.2	60.0	86.1	51.1	36.9	46.9	25.7

Table 2: Performance (%) of agent capability on WildClaw, ClawEval, GDPval, APEX, and Agent’s Last Exam. The best results are in bold and the second-best results are underlined within each group. Additional evaluation details are provided in Appendix C.4 .

Figure 4: Scaling analysis of WorkForge . Performance on ALE and WildClaw with increasing (a) Training data volume and (b) Interaction length.

Figure 5: Effect of trajectory harness. We compare two training groups, each using 8k trajectories generated by Claude Code or OpenClaw, respectively.

	ALE		APEX
	Score	Pass@1	Score
LLM only	15.62	4.8	11.18
Programmatic	14.89	4.8	10.92
Hybrid	22.11	11.4	15.39

Table 3: Ablation of verification-based trajectory selection. Each variant uses trajectories selected by the semantic score, the programmatic score, or their combination.

Figure 6: Execution behavior and rubric-level performance on GDPval. (a) The x-axis presents the mean number of distinct external web domains visited per task; the y-axis presents evidence retrieval within the workspace as a percentage of tool calls. Arrows connect each base model to its fine-tuned counterpart; color records reasoning length after tool errors. (b) Pass rate across five rubric categories. Details about rubric categories are described in Table 11 .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Domain taxonomy and task distribution of the WorkForge SFT corpus. The left panel shows all 40 domains, and the right panel reports the ten most represented domains.

#	Domain	Count	%	#	Domain	Count	%
1	Office Productivity	1,394	8.36	21	Commerce & eCommerce	275	1.65
2	Data & Database	1,368	8.21	22	Devices & IoT	271	1.63
3	Business & Enterprise	1,324	7.94	23	Medical	243	1.46
4	Finance & Payments	1,237	7.42	24	Visual Recognition & Image Processing	237	1.42
5	Advertising & Marketing	983	5.90	25	Storage & File Management	212	1.27
6	AI & Machine Learning	701	4.21	26	Events	211	1.27

Appendix

Table 4: Complete domain distribution. Counts are computed over 16,665 SFT tasks.

Configuration	Value
Optimizer	AdamW
Adam coefficients	$\beta_{1}=0.9$ , $\beta_{2}=0.95$
Peak learning rate	$1\times 10^{-5}$
Weight decay	0.1
Gradient clipping	1.0
Learning-rate schedule	Cosine decay to $1\times 10^{-7}$

Appendix

Table 5: Supervised fine-tuning configuration used for the main models and training-data ablations.

Parameter	Value
Temperature	0.7
Top- $p$	0.9
Top- $k$	20
Maximum generated tokens per response	32,768
Maximum context length	262,144

Appendix

Table 6: Decoding configuration used for model evaluation.

Parameter	Value
Maximum position embeddings	262,144
RoPE base ( $\theta$ )	1,000,000
RoPE scaling type	YaRN
Scaling factor	6.4
Original maximum position embeddings	40,960

Appendix

Table 7: Shared YaRN configuration for the Qwen3-based baselines.

Benchmark	Evaluation set	Reported metrics	Rollouts/task
WildClaw	Full set	Score, Pass@1	1
Claw-Eval	Full set	Score, Passˆ3	3
GDPval	201 text-only tasks	Rubric score	1
APEX	Full set	Score, Pass@1	1
ALE	Full set	Score, Pass@1	1

Appendix

Table 8: Benchmark-specific metrics and sampling protocols.

Benchmark	Harness	Wall-clock limit	Interaction limit
WildClaw	OpenClaw	24,000 s	Not set
Claw-Eval	OpenClaw	3,600 s	50 turns
GDPval	Claude Code	14,400 s	300 turns
APEX	ReAct Toolbelt	14,400 s	250 steps
ALE	Claude Code	28,800 s	Unlimited

Appendix

Table 9: Evaluation harnesses and per-task agent execution budgets.

Row	Task ID	Occupation
12	38889c3b-e3d4-49c8-816a-3cc8e5313aba	Audio and Video Technicians
13	ff85ee58-bc9f-4aa2-806d-87edeabb1b81	Audio and Video Technicians
14	4b894ae3-1f23-4560-b13d-07ed1132074e	Audio and Video Technicians
56	e222075d-5d62-4757-ae3c-e34b0846583b	Film and Video Editors
57	c94452e4-39cd-4846-b73a-ab75933d1ad7	Film and Video Editors
58	75401f7c-396d-406d-b08e-938874ad1045	Film and Video Editors

Appendix

Table 10: Multimodal tasks excluded from the GDPval evaluation set. Row numbers refer to the official 220-task set.

Category	Criteria	Share	Base	SFT	$\Delta$
Visual / layout quality	371	4.2%	74.1%	87.1%	+13 pp
Numerical claims	523	5.9%	64.6%	72.8%	+8.2 pp
Cross-source reconciliation	103	1.2%	86.4%	92.2%	+5.8 pp
Reasoning / argumentation	200	2.2%	82.0%	86.5%	+4.5 pp
Content coverage	7,718	86.6%	77.4%	83.4%	+6.0 pp

Appendix

Table 11: GDPval rubric pass rates across five deliverable-quality categories.

Group	Turns per trajectory	Trajectories
G1	0–20	1,000
G2	21–40	1,000
G3	41–60	1,000

Appendix

Table 12: Training subsets for trajectory-length scaling.

Group	Productive calls	Skill discovery	Tool calls	Distinct actions	Tool types used
G1	11.7	21.0%	17.7	17.0	3.3
G2	23.1	21.9%	38.3	33.0	3.6
G3	24.0	33.3%	42.0	38.3	4.0

Appendix

Table 13: Tool-use behavior under trajectory-length scaling on WildClaw. Longer-turn supervision increases exploration intensity, breadth, and productive execution.

Variant	Selection rule	Trajectories
LLM only	$S^{\mathrm{sem}}\geq 0.7$	1,000
Programmatic only	$S^{\mathrm{prog}}\geq 0.7$	1,000
Hybrid	$0.5S^{\mathrm{sem}}+0.5S^{\mathrm{prog}}\geq 0.7$	1,000

Appendix

Table 14: Trajectory selection rules for the verification-filter ablation.

Explore similar work

Oct 5, 2026cs.AI

VERA: Scaling Verifiable Environments for Agentic co-Evolution

Competent agents need precise and verifiable environments, such as sandboxes that are resumable at any stage and evolve from observable evidence. However, most long-horizon work exposes how rare these are: for example, an agent in medical research must ground a finding, classify it, and write a report over dozens of dependent steps, yet recent environments score only the outcome. To address the challenges in stable training, we present VERA, which builds such environments at scale and lets agents evolve on them. VERA builds these environments from initial trajectories: an agent writes rubrics, executable checks, a judge verifies each sandbox, and only those that pass enter the training bank. On these environments, VERA alternates between two updates: train the model with rubric rewards, or edit the harness skills. We also create a verifier which gates model checkpoints and harness edits using explicit development-set acceptance criteria. This attribution distinguishes VERA's co-evolution from single-axis baselines: its updates target not only the cause but the outcome. With an open-source corpus of 9,000+ long-horizon verifiable environments, a 9B model paired with its co-evolved agent beats the strongest baseline by 10.3 and 13.0 points in the two domains. At 27B, it surpasses the baseline on AutoCoWorkBench (71.6) and AutoMedBench (80.7), transfers to unseen workflows, and retains general capabilities.

Junqi Liu, Yongyang Pan, Zhuosong Jiang +14

University of California, Santa Cruz · University of Illinois Urbana-Champaign · National University of Singapore +4

Reinforcement Learning with Verifiable Rewards Long-Horizon LLM Agents Self-Evolving Agents

Ask PDF

Aug 22, 2026cs.AI

Training Needs Trustworthy Worlds: Verified Synthetic Web Environments for Agent Learning

Web agents promise to automate complex digital workflows, but their training remains limited by synthetic environments that look plausible while hiding broken links, inconsistent states, or infeasible tasks. We address the gap between scalable environment generation and trustworthy agent learning by constructing synthetic web environments that are executable, auditable, and grounded in backend state. Our framework represents each generated website as a structured scaffold of pages, navigation links, database records, state-change markers, and task constraints, then verifies and repairs structural, semantic, consistency, and feasibility defects before policy training. During interaction, ordinary UI transitions are executed deterministically, while persistent backend updates are invoked only through validated state-change markers, enabling dense rewards compiled from verified task-progress predicates. Across 500 synthetic environments spanning six domains, our method reduces task-blocking defects and improves feasible-task rate from 48.6% to 94.8%, while producing stronger PPO policies and improving transfer to WebArena, WebShop, and MiniWoB++ without LLM calls at evaluation time. These results show that verified synthetic environments can serve as a scalable and reliable training substrate for compact web agents, shifting synthetic webagent learning from surface-level plausibility toward executable, state-grounded supervision.

Chenghao Zhang, Yuxi Cheng, Saisai Hu +3

University of Pennsylvania · Pace University · Shenzhen Campus of Sun Yat-sen University

Reinforcement Learning with Verifiable Rewards Web Agents

Ask PDF

Sep 30, 2026cs.CL

GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

Working agents need to read diverse files, coordinate tools, and produce deliverables. Training such agents requires tasks built on many real files with verifiable results, but few pipelines exist to synthesize this kind of data. Existing pipelines either generate files with models, which lack realism and diversity, or build tasks on real files without task-specific verifiers, leaving result quality unchecked. We introduce GraphForge, an evidence-graph based framework that grounds both the task and its verification in real files. Starting from occupation-grounded seeds for controlled diversity, GraphForge assembles a workspace of real files for each seed and builds an evidence graph over their relations. Since the task statement and rubrics are both derived from this graph, task requirements are backed by the workspace files and each criterion is anchored to the files needed to verify it. An initial rollout further tests executability, and a revision agent repairs the task and rubrics against the original files before trajectories are collected. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories brings GDPVal to 1445.7 (+65.7) under OpenHands, and Workspace-Bench-Lite and SpreadsheetBench II to 63.7 (+7.7) and 24.0 (+13.7) under Claude Code. Rejection fine-tuning on the SFT model's own rollouts, with candidates selected by the evidence-anchored rubrics, yields further improvements on all three benchmarks, suggesting that the rubrics provide a useful selection signal.

Qisheng Su, Hanchen Wang, Guanru Zhu +10

University of Science and Technology of China · Shanghai Innovation Institute · Fudan University +1

AI Agent Benchmarks LLM Agent Training LLM Agent Planning

Ask PDF