cs.AISep 30, 2026

WorkGenesis: Building the Worlds That Teach Agents to Work

Authors: Xinyu Zhu, Fenyi Liu, Yuzhu Cai, Shuo Tang, Rui Ye, Linfeng Zhang, Siheng Chen

Organizations: Shanghai Jiao Tong University · Endless Frontier · Beihang University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Figure 1: Performance comparison of different models on three authoritative occupational work benchmarks. Fx-Work-35B, trained on work scenarios synthesized by WorkGenesis, outperforms all comparable-scale models as well as DeepSeek-V4-Pro-Preview with 1.6T parameters.
Figure 2: Overview of the WorkGenesis framework. (a) Evidence-Based Work Construction : WorkGenesis repeatedly searches for suitable real-world anchor files and synthesizes additional reference files from them to construct the complete work scenario. It then assigns an independent agent to operate in this scenario and produce a reference deliverable. (b) Execution-Guided Consistency Verification : WorkGenesis scores the reference deliverable using a rubric and attributes unsatisfied rubric items, identifying defects in both the synthesized occupational task itself and the rubric. These defects are used as feedback signals to further revise the entire work scenario. This verification process repeats until no rubric item is attributed to the occupational scenario itself.
  GDPvalAA-v2APEX-Agents-AAJobBench 
ModelParamsRubricElopass@1MainEasyAVG
Frontier Models
Qwen3.8-Max2.4T-A95B91.53173938.952.5782.8151.14
Kimi-K32.8T-A104B91.64168735.543.8180.7246.22
GPT-5.5–90.50149129.930.1078.2236.52
GLM-5.2753B-A40B90.34151029.733.7975.4937.80
Table 1: Main results on GDPvalAA-v2, APEX-Agents-AA and JobBench. Fx-Work-35B significantly outperforms all ∼\sim 35B-sized models on all three benchmarks. It even surpasses the 1.6T DeepSeek-V4-Pro-Preview in GDPval Elo score, as well as on APEX-Agents and JobBench. Bold marks the best value within the ≤\leq 35B group. GDPval Elo scores are computed based on the August 4, 2026 snapshot and standard calculation methodology from Artificial Analysis. The average score for each model is computed by averaging the normalized Elo score on GDPvalAA-v2 and the normalized scores on APEX-Agents-AA and JobBench Main split.
Figure 3: Pairwise performance comparison of Fx-Work-35B and other baselines on GDPvalAA-v2 and JobBench. “Win” denotes the percentage of tasks on which Fx-Work-35B outperforms the compared model; “Tie” and “Lose” are defined analogously.
Figure 5
Figure 6: Number of turns per task before and after training. The horizontal axis is the number of turns required to complete a task and the vertical axis is the share of tasks. Fx-Work-35B exhibits stronger long-horizon task capabilities.
Figure 7
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

SectorOccupationSectorOccupation
FINCustomer Service RepresentativesFINFinancial Managers
FINFinancial and Investment AnalystsFINPersonal Financial Advisors
FINSecurities, Commodities, and Financial Services Sales AgentsGOVAdministrative Services Managers
GOVChild, Family, and School Social WorkersGOVCompliance Officers
GOVCourt, Municipal, and License ClerksGOVFirst-Line Supervisors of Police and Detectives
GOVPublic Safety TelecommunicatorsGOVRecreation Workers
Appendix
Table 2: The 65 occupation categories used in WorkGenesis. Sectors: FIN = Finance and Insurance, GOV = Government, HCS = Health Care and Social Assistance, INF = Information, MFG = Manufacturing, PST = Professional, Scientific, and Technical Services, RER = Real Estate and Rental and Leasing, RET = Retail Trade, WHL = Wholesale Trade, EDU = Educational Services, MIN = Mining, Quarrying, and Oil and Gas Extraction.
Model and initialization
Base modelQwen3.6-35B-A3B (sparse MoE)
Training data and mixture
Update steps1,537
Packing bin256,000 tokens (16k per GPU over CP)
Samplingshuffled and length-balanced per rollout
Optimization
Appendix
Table 3: Key training hyperparameters for training Fx-Work-35B.
Figure 9: Training loss curve of Fx-Work-35B.
Figure 10: Diversity of the constructed work. (a) Distribution of the 20K units of work over the 65 occupations.

Explore similar work

CardsList
May 25, 2026cs.AI

JobBench: Aligning Agent Work With Human Will

Current benchmarks for occupational AI agents are scoped primarily by economic values, telling a replacement story. We introduce JobBench, which evaluates AI agents on the workflows that experts identify as high-priority for delegation, empowering humans based on their needs instead of replacing them with GDP value. JobBench covers 130 agentic tasks across 35 occupations. Each task is packaged as a workspace of heterogeneous reference files, requiring the agent to reason through the cluttered information streams of real professional work. Outputs are graded by a fact-anchored chain of rubrics, averaging 35.6 binary criteria per task. We evaluate 36 models; the strongest, Claude Opus~4.7 under Claude Code, reaches only 45.9 %. We hope JobBench shifts the community's target labour-market effect from replacement to enhancement: building agents that do what humans actually want delegated, not only what is most economically valuable.

Yuetai Li, Yichen Feng, Zhangchen Xu +21

University of Washington · 10Bake AI · University of California, Santa Barbara +10

Economies
AskPDF
May 9, 2026cs.LG

SkillGen: Verified Inference-Time Agent Skill Synthesis

Skills are a promising way to improve LLM agent capabilities without retraining, while keeping the added procedure reusable and controllable. However, high-quality skills are still largely written by hand. We introduce SkillGen, a multi-agent framework that synthesizes a single auditable skill from trajectories generated by a base agent. The output is a human-readable artifact that can be inspected before use. Rather than merely summarizing trajectories, SkillGen leverages contrastive induction over both successful and failed trajectories to identify reusable success patterns, recurring failure modes, and behaviors that appear in nearby successes but are missing from failures. SkillGen then generates candidate skills and iteratively refines the skill. A key novelty in SkillGen is that we model agent skills as interventions to empirically verify the net effect of skills on the overall performance. Specifically, we compare outcomes on the same instances with and without the skill, so that we account for both repairs (cases where the skill fixes a baseline failure) and regressions (cases where the skill breaks a baseline success). Across a broad range of agents and datasets, SkillGen consistently improves held-out performance, outperforms existing skill-generation baselines, and produces skills that transfer across models.

Yuchen Ma, Yue Huang, Han Bao +5

Munich Center for Machine Learning, LMU Munich · University of Notre Dame · Microsoft Research

Reusable Agent SkillsSkillsSynthesis
AskPDF
Jul 15, 2026cs.SE

NexForge: Scaling Executable Agent Tasks via Requirement-First Synthesis

Scaling executable agent training data is bottlenecked by substrate-first methods that tie task generation to predefined tools, repositories, or skill graphs: expanding coverage requires manual expansion of the substrate, each new domain demands a bespoke pipeline, and the resulting task distributions often reflect substrate convenience rather than real-world demand. We introduce NexForge, a requirement-first framework that compiles free-form capability requirements into executable agent training data. NexForge first performs research-based demand discovery to identify representative task forms, realistic scenarios, and their relative prevalence. It then applies distribution-aware task compilation and automatically retrieves or constructs the files, repositories, dependencies, and runtime configurations required to materialize each task, followed by teacher rollout collection and trajectory distillation. The same pipeline, without any domain-specific infrastructure, produces 3,600 terminal tasks and 2,000 office tasks, improving Qwen3.5-35B-A3B Base from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval; scaling to 43.2K terminal tasks reaches 58.4%, surpassing Claude Opus 4.6. Scaled further, NexForge-synthesized data contributes to the training of Nex-N2, a family of publicly available agent models that lift Qwen3.5-35B-A3B to 75.3% on Terminal-Bench 2.1 and to 1585 Elo on GDPval -- achieving state-of-the-art open-source performance and surpassing several frontier proprietary systems. Nex-N2 models are available at https://nex.sii.edu.cn/

Jiarong Zhao, Zhikai Lei, Zhiheng Xi +5

1East China Normal University · 2Shanghai Qiji Zhifeng Co., Ltd

Synthetic TaskForgeScaling
AskPDF