Tencent and Fudan Build 16,665 Workspaces for AI Agent Training

Tencent’s Hunyuan team and Fudan University developed WorkForge, a system that trains AI agents on file-based tasks and evaluates their completed work. The project created 16,665 workspaces spanning 40 professional fields and 60 file types, with an average of 42.3 files in each.

The workspaces draw on publicly available materials, including reports and spreadsheets. In one finance exercise, an agent compares stock-factor returns across regions, organizes the data and delivers its findings in a presentation, spreadsheet and CSV file.

WorkForge extracts facts that can be checked against the source documents, then uses them to create tasks and scoring criteria. Software checks numerical results and whether the required files were submitted. A separate AI system assesses the substance of the analysis.

The team generated demonstration workflows using GLM 5.3 alongside Claude Code and OpenClaw, then used those examples to fine-tune Qwen3.5 models. A 35B-A3B base model’s GDPval score increased from 45.5 to 73.6, while its APEX score rose from 5.0 to 21.3.

A separate 27B model scored 82.4 on GDPval, compared with 79.7 for Qwen3.5-397B, which has nearly 15 times as many parameters. The 27B model’s score remained below GPT-5.6 Sol’s 90.1.