- The WorkFactory framework generates 112,686 structured, executable tasks for AI agent training across 612 occupations, eliminating the need for hand-labelled demonstrations at each step.
- Fine-tuning on WorkFactory-generated training trajectories pushed WorkPilot-35B-A3B’s Elo score from 879 to 1565 on the GDPval-AA v2.1 benchmark, though gains on out-of-distribution tasks remain untested.
- WorkFactory’s probe-and-repair loop catches data defects before they reach training, cutting the human curation overhead that has constrained professional-work agent datasets.
Getting AI agents to do real professional work reliably means training them on real professional tasks, at scale, with consistent structure. A paper posted to arXiv on October 27, 2026, by Ziyang Zeng and a team of 12 researchers proposes a framework called WorkFactory that converts high-level professional goals into 112,686 structured, executable tasks spanning 612 occupations, without requiring expert demonstrations at every step.
The Training Data Problem
Agents capable of legal analysis, financial auditing or similar knowledge work need supervised training data that is diverse, consistent and grounded in real professional tasks. Current methods cannot produce that at volume. WorkFactory’s approach is to treat professional work as a collection of “intents” and decompose each one systematically into actionable steps.
From Intent to Executable Plan
The process starts by translating a high-level professional intent into what the paper calls an “information support graph.” The graph maps dependencies between the pieces of information a task requires and where those pieces can be found. For a financial analyst working through a Q3 earnings review, that means tracing a chain from “locate the Q3 report” through “extract revenue figures” to “compare against the prior quarter,” before any execution begins. Each node specifies not just the action but the required data and its expected source.
Making dependencies explicit before task instantiation is what keeps generated tasks internally consistent. The framework mirrors the structured logic of actual professional work rather than producing loosely coupled steps an agent might not be able to follow end to end.
The Probe-and-Repair Loop
After planning, WorkFactory runs simulated “probe executions” on generated tasks to catch what the paper terms “support failures”: cases where required information is missing, inconsistent or unreachable in the defined environment. When a failure is detected, the system runs a repair pass that preserves the intent of the original task while fixing the defect.
The controlled component analyses in the paper show this directly reduces probe-diagnosed environment defects and improves downstream agent performance. It functions as an automated quality layer, cutting the human curation effort that would otherwise be needed to produce reliable training demonstrations at this scale. The approach embeds self-correction at the data generation stage rather than at inference, which is the more common point of intervention in agent research.
WorkPilot Results
The team used training trajectories from WorkFactory to fine-tune Qwen models at 4B, 27B and 35B-A3B scales, producing a family they call WorkPilot. Across four professional-work benchmarks, the WorkPilot models improved on their base counterparts. The sharpest result: WorkPilot-35B-A3B moved from an Elo score of 879 to 1565 on the GDPval-AA v2.1 benchmark.
That is a large jump. Elo scores on domain-specific benchmarks are notoriously sensitive to the composition of the comparison set, so the absolute number warrants some caution. What the result does establish is that automated task generation can produce training signal strong enough to move a model meaningfully on professional-work evaluations, without hand-labelled demonstrations at every step. Whether the gains hold on out-of-distribution tasks the framework was not designed around is left open by the paper.
