WorkFactory lifts WorkPilot-35B Elo from 879 to 1565 on GDPval
TL;DR
- Fine-tuning on WorkFactory data lifts WorkPilot-35B-A3B from 879 to 1565 Elo on GDPval-AA v2.1, a roughly 1.78x gain.
- WorkFactory generates 112,686 executable professional tasks spanning 612 occupations using an 'information support graph' of output dependencies.
- All three WorkPilot models (4B, 27B, 35B-A3B) beat their Qwen base models across four professional-work benchmarks, per the abstract.
Fine-tuning on 112,686 synthetic professional tasks lifts a 35B-A3B Qwen agent from 879 to 1565 Elo on the GDPval-AA v2.1 benchmark, according to a paper posted to arXiv on October 27 by researchers led by Ziyang Zeng.
The pipeline, called WorkFactory, generates tasks spanning 612 occupations. The authors describe the method this way: "information-level planning builds an information support graph to guide task instantiation, explicitly representing the dependencies and locations of information required for the intended outputs." Probe executions catch tasks the teacher can't actually complete; those get repaired without changing the intended work.
Three WorkPilot models are released, fine-tuned at 4B, 27B, and 35B-A3B scales. "All three models improve over their respective base models across four professional-work benchmarks," the abstract reports. Only the 35B-A3B's Elo figure is cited.
The paper names neither the teacher model producing the demonstrations nor the other three benchmarks. 612 occupations is a count, not a list.
Originally reported by paper
Read the original article →Original headline: WorkFactory Builds 112K Professional Tasks Across 612 Jobs; WorkPilot-35B Elo Nearly Doubles