Beating AI News Flash: Tencent Hunyuan, in collaboration with Fudan University, proposes WorkForge, specifically designed to batch-build complex, verifiable work training environments for AI Agents.The team collected materials such as spreadsheets and reports from real public resources, covering 40 professional domains and 60 file types, building a total of 16,665 training environments. Each workspace contains an average of 42.3 files, far more complex than simulated environments that rely solely on AI-generated materials.For example, in financial tasks, the Agent needs to analyze stock factor returns across different regions, organize data, compare performance, and finally deliver PPT, Excel, and CSV files. WorkForge first extracts verifiable facts from the original files, then automatically generates tasks and scoring criteria. The program checks specific numbers and deliverable files, while a separate AI evaluates the analytical content, making training results as evidence-based as possible.The team used GLM 5.3, together with Claude Code and OpenClaw, to generate demonstration operation trajectories, then fine-tuned Qwen3.5 with this data. After training, the 35B-A3B base model's GDPval score rose from 45.5 to 73.6, and APEX from 5.0 to 21.3. Another 27B model reached 82.4 on GDPval, slightly higher than Qwen3.5-397B (79.7), which has nearly 15 times the parameter count, but still below GPT-5.6 Sol's 90.1.
