For workplace agents, a polished demonstration is only useful if the task itself contains enough consistent evidence to complete the work. This paper treats task-environment quality as a first-class training problem: plan what information an output needs, make that information accessible, then test and repair the task before collecting demonstrations.
Key Findings
- The framework builds an information support graph backward from each intended deliverable, distinguishing supplied facts from calculations and judgments the agent must perform. This helps prevent task construction from accidentally giving away the answer or omitting essential premises.
- Probe executions diagnose missing or contradictory support using only the solver-visible materials. Repairs address blockers while preserving the original objective, so the resulting teacher trajectories are grounded in executable tasks rather than hidden construction assumptions.
- The resulting corpus contains 112,686 tasks across 612 occupations. Fine-tuning Qwen models at 4B, 27B, and 35B-A3B improves results over their respective base models on four professional-work benchmarks; WorkPilot-35B-A3B rises from 879 to 1565 Elo on GDPval-AA v2.1.
Who Should Read It
Great fit for researchers building professional agents, task environments, or supervised training datasets who need reliable demonstrations at scale. Look elsewhere if your priority is multimodal workplace learning or reinforcement learning: the study uses text-only task solving and supervised fine-tuning, leaving both directions for future work.
