Tencent’s WorkForge Scales Verifiable Training for Work Agents
Tencent researchers introduce WorkForge, a framework that turns real-world files and workflows into verifiable environments for long-horizon agents.
From workflows to testable workspaces
Tencent and Fudan University researchers are presenting WorkForge as a way to address one of the least glamorous but most consequential bottlenecks in agent development: the shortage of realistic training environments where an agent’s work can be checked reliably.
The framework begins with professional workflows rather than isolated prompts. It identifies the resources, decisions and deliverables required for a job, then assembles workspaces from real-world files. WorkForge analyzes those workspaces to find factual anchors—observable details that can support both task construction and evaluation. It then creates instructions, solution plans and programmatic or semantic verifiers tied back to that evidence.
The team reports 16,700 verifiable environments spanning 40 professional domains and 60 file types. In post-training experiments, Qwen3.5-35B-A3B-Base improved its GDPVal score from 45.5 to 73.6 and its APEX score from 5.0 to 21.3. The paper also reports that a 27B model reached highly competitive results against stronger systems.
Why it matters
Most agent benchmarks still reward a final answer without fully exposing whether the system followed a dependable process. WorkForge instead treats the workspace—the files, intermediate decisions and produced artifacts—as part of the supervision signal. That is a meaningful shift for agents expected to handle finance, research, operations or office work over dozens of steps.
The unresolved issue is whether automatically synthesized verifiers remain trustworthy outside the paper’s selected domains. Still, the approach points toward a more scalable training economy for agents: less hand-built simulation, more evidence-linked work.
Uncle Cat take
WorkForge’s strongest contribution is not its Qwen score; it is tying evaluation to 60 file types, which gives enterprise agents a more credible path beyond toy tasks.