Tencent Hunyuan and Fudan propose WorkForge, building 16,665 Agent training environments

AT A GLANCE

The report states that Tencent Hunyuan and Fudan University proposed WorkForge, using 16,665 work environments to train and evaluate AI Agent。

Article

Tencent builds an 16700 work environment suite for Agents: let GLM demonstrate, while Qwen practices

Beating AI Flash reports that Tencent Hunyuan and Fudan University jointly proposed WorkForge, specifically for AI Agent to batch-build complex, verifiable work training environments.

The team collected spreadsheets, reports, and other materials from real public resources, covering 40 professional fields and 60 file types, and built a total of 16,665 training environments. Each workspace contains an average of 42.3 files, making it far more complex than simulated environments that rely solely on AI to generate materials from scratch.

For example, in financial tasks, an Agent must analyze stock factor returns across different regions, organize data, compare performance, and ultimately deliver PPT, Excel, and CSV。WorkForge first extracts verifiable facts from the source files, then automatically generates tasks and scoring criteria. A program checks specific figures and deliverable files, while another AI evaluates the analytical content, making training results as evidence-based and traceable as possible.

The team used GLM 5.3, together with Claude Code and OpenClaw, to generate demonstration action trajectories, then used these data to fine-tune Qwen3.5. After training, the 35B-A3B base model's score on GDPval increased from 45.5 to 73.6，APEX increased from 5.0 to 21.3. Another 27B model achieved GDPval on 82.4, slightly higher than 15 with nearly Qwen3.5-397B（79.7 times as many parameters), but still below GPT-5.6 Sol's 90.1.

Original link https://m.theblockbeats.info/flash/373207

Key points

Beating AI reports that Tencent Hunyuan and Fudan University jointly proposed WorkForge, for batch-building complex, verifiable AI Agent work training environments.
The team collected spreadsheets, reports, and other materials from real public resources, covering 40 professional fields and 60 file types, and built a total of 16,665 training environments.
Each workspace contains an average of 42.3 files.
WorkForge It first extracts verifiable facts from the source files, then automatically generates tasks and scoring criteria.
The team used GLM 5.3, together with Claude Code and OpenClaw, to generate demonstration action trajectories, then used these data to fine-tune Qwen3.5。
The report shows that the 35B-A3B base model's GDPval score increased from 45.5 to 73.6，APEX score increased from 5.0 to 21.3。
Another 27B model's GDPval score was 82.4；Qwen3.5-397B was 79.7，GPT-5.6 Sol was 90.1。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

WorkForge Like a simulated office for AI assistants: it organizes public materials into workspaces containing multiple files, and generates tasks, scoring rules, and demonstration steps, enabling models to practice organizing data, analyzing content, and delivering PPT, Excel, and CSV deliverables.

Why it matters to readers

The method described in the report seeks to train Agents in more complex file environments with checkable results; the listed evaluation scores show that the trained models improved.

Beginners can understand it as preparing a complete set of exercises, reference operations, and an automated scoring system for AI, rather than merely having AI answer individual questions.

Risks and unknowns

The supplied materials do not include a paper, code repository, dataset, or official announcements from Tencent Hunyuan or Fudan University; the related project and figures therefore cannot be independently verified on this basis.
The report does not specify the full evaluation setup, so it cannot be confirmed whether scores for different models are directly comparable.
The materials do not state whether WorkForge has entered product or commercial use.
Whether the training results can be reproduced by third parties remains unclear.

Tencent Hunyuan and Fudan propose WorkForge, building 16,665 Agent training environments

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center