Tencent and Fudan's WorkForge synthesizes 16.7K verifiable environments, lifting GDPVal from 45.5 to 73.6
Tencent-Hunyuan · hf · 2026-10-15
Tencent's LLM Department and Fudan University introduce WorkForge, a scalable framework for synthesizing verifiable work-agent environments from real-world resources (paper on arXiv).
The problem: work agents need training environments that support long-horizon interaction and trustworthy verification, but hand-crafted environments are too expensive to scale, while synthesis methods sacrifice workspace complexity, realism, or grounded verifiability.
How it works:
- Starts from expert workflows, identifying the resources, decisions, and deliverables each requires;
- Retrieves real-world files into a workspace and extracts concrete, checkable factual anchors;
- Task instructions, solution plans, and both programmatic and semantic verifiers are derived directly from these anchors, keeping verification traceable to observable workspace evidence.
Results: 16.7K verifiable environments across 40 professional domains covering 60 file types. Post-training Qwen3.5-35B-A3B-Base improves GDPVal from 45.5 to 73.6 and APEX Score from 5.0 to 21.3, while Qwen3.5-27B becomes highly competitive. Analyses show consistent scaling behaviors in both data volume and interaction horizon.
More from coding & agent
- Sentry CEO says AI agent is now the company's No.1 code contributor — zeeg · 2026-10-15
- Dev: I accept about half of AI coding suggestions, a third surprise me positively — nickbaumann_ · 2026-10-15
- Armin Ronacher on why he prefers React after building with Solid.js, plus what's new with Pi — ianand · 2026-10-15
- Agents write most new code now — has your review process scaled? — ronbodkin · 2026-10-15
- 8k context beats 480k: OpenAI Decisions API cuts costs 84% at 98.25% F1 — OpenAIDevs · 2026-10-15
- Sentry CEO: AI agents still write messy code but are getting better at managing it — zeeg · 2026-10-15