WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training

WorkFactory constructs executable professional tasks by planning information support and repairing environments through probe execution, then uses teacher trajectories to train the WorkPilot models.

Scaling supervised training for working agents requires diverse professional tasks that yield reliable expert demonstrations, but missing or inconsistent task information can obstruct teacher execution and weaken supervision. The information needed for an intended output may be distributed across artifacts and connected through intermediate calculations and judgments, making it hard to ensure all premises are available while leaving intermediate and final results for the solver. If this boundary is not preserved, teachers may introduce unsupported assumptions or leave essential work incomplete.

WorkFactory first synthesizes professional intents from occupational hierarchies, documented skills, and real deliverables, with LLM-based review. For each intent it plans backward using an information support graph that represents information units, carriers, support relations, and location bindings, while reserving derived results for the solver. Artifacts are materialized under a construction plan with programmatic and LLM-based checks and up to three automatic construction retries. A probe solver attempts the realized task from solver-visible inputs; blocker findings drive repairs that must preserve the intended work, followed by revalidation and up to three repair rounds. Tasks that pass are assigned to an independent teacher, whose trajectories are used for supervised fine-tuning.

• Fine-tuning on the WorkFactory corpus improves all three base models on four professional-work benchmarks; WorkPilot-35B-A3B gains on GDPval-AA, JobBench, Workspace-Bench, and OmegaUse-OfficeVal. §4.2
• WorkPilot-27B reaches a locally estimated 1612 Elo on GDPval-AA v2.1, 12 Elo above DeepSeek-V4.1-Flash and 34 points below GLM-5.3. §4.2
• In the 5K-task component analysis, information-level planning followed by execution-based repair reduces cross-file inconsistency incidence from 32.82% to 25.74% to 4.84% and raises GDPval-AA Elo from 1245 to 1293 to 1326. §4.4
• On a matched 30K-task subset, students trained on GLM-5.3 trajectories reach 1483 Elo and students trained on DeepSeek-V4-Flash-0731 trajectories reach 1248 Elo, both above the 879-Elo base. §4.5
• Execution-based repair on the 5K analysis set leaves 69.58% of tasks with zero repair rounds, 20.24% with one, 7.70% with two, and 2.48% reaching the three-round limit; 24 tasks remain blocked and are excluded. §C.3
• All three WorkPilot models lie on the non-dominated cost-performance frontier, with estimated costs of approximately $0.012, $0.031, and $0.396 per GDPval-AA task. §4.6

• The teacher model GLM-5.3 accepts only text input, so the models are trained and evaluated only through a text-only task-solving interface; the work does not examine learning from visual evidence directly.
• The training study focuses on supervised fine-tuning over teacher trajectories, not reinforcement learning.
• Non-blocking errors may remain in retained tasks; efficiency and open findings are advisory, and unclassified findings do not enter the eight-category table.
• Cost estimates are standardized estimates rather than provider invoices and exclude judges, verifiers, tool and web-search fees, sandbox compute, storage, and orchestration overhead.

• Intent synthesis uses occupational hierarchies, documented skills, and real professional deliverables via instruction backtranslation, followed by LLM-based review for coherence, feasibility, professional appropriateness, and plausibility. §3.2
• The information support graph contains information units, carriers, support relations, and location bindings, and is planned before materialization to make the premises-versus-work boundary explicit. §3.3
• The probe solver is told the task is synthetic, grounds diagnoses in sources and tool observations, and distinguishes environmental defects from solver mistakes and legitimate open choices; only blocker findings trigger repair. §C.1
• Repair preserves intended work by editing the request and existing reference files without updating the information support graph; the number, dependency relations, and types of reference files must remain unchanged. §C.2
• The corpus contains 112,686 tasks spanning 612 occupations, with 4.15 provided reference files per task on average, 317 reference-file extensions, and 53 deliverable-file extensions. §D.1
• Construction uses 11.656M model tokens per attempted task on average, with information-level planning the largest stage at 4.737M tokens, followed by execution and repair at 3.644M and material synthesis at 3.275M. §D.2
• The primary scaling sweep trains Qwen3.6-35B-A3B on nested subsets of 2K–110K tasks, with Qwen3.6-27B and Qwen3.5-4B additionally evaluated at 30K and 110K. §4.3