WorkForge scales verifiable real-workspace environments for work agents

WorkForge, from researchers at Tencent's LLM Department and Fudan University, is a method for building training environments for "work agents": agents that turn a folder of real files into professional deliverables such as a review report, a spreadsheet or a committee briefing. The authors argue that hand-built environments cost too much to scale. Automatically synthesized ones, in turn, tend to be small and over-curated, with the relevant evidence sitting in a few obvious files, and graders that are not tied to what is actually in the workspace can invent requirements.

The pipeline starts from expert-written workflow descriptions. An acquisition agent then searches for and downloads real repositories, datasets and web files and assembles them into a workspace, keeping their original folder structure. Next, WorkForge inspects that workspace to extract concrete, checkable facts. Each task's instructions, a hidden solution plan and two kinds of grader are derived from those facts. Programmatic checks test exact values and file states, and an LLM judge scores open-ended quality against rubrics. In the paper's worked example, a factor-investing task sits on 297 CSV files. One extracted fact corrects an assumption: the emerging-market five-factor series starts in July 1989, not July 1990, and a grader checks for it.

The authors built 16.7K such environments across 40 professional domains and 60 file types. They had GLM 5.3 solve the tasks inside the Claude Code and OpenClaw harnesses, then used the scored attempts as supervised fine-tuning data. Trained on this data, Qwen3.5-35B-A3B-Base rose from 45.5 to 73.6 on GDPval and from 5.0 to 21.3 on APEX. Qwen3.5-27B, already post-trained, went from 79.4 to 82.4 on GDPval, against 89.1 for Claude Opus 4.8 and 90.1 for GPT-5.6 Sol in the same table. Selecting training attempts by both grader scores together worked better than using either one alone, and performance rose with more data and with longer attempts.

All benchmark numbers come from the authors' own evaluation runs. GDPval, for example, was scored on 201 of its 220 tasks, leaving out the multimodal ones. The gains come from fine-tuning on another model's graded work, and the paper found some harness-specific transfer: models did better when trained and tested in the same agent harness. The authors also list limits. WorkForge depends on expert workflows and publicly accessible files, so proprietary domains are hard to cover. Its automatically built graders are only as good as the workspace profiling, have had limited expert calibration, and judge only final deliverables, not the steps an agent took.