GraphForge Released: Evidence-Graph Based Framework for Training Working Agents
Summary:GraphForge is an innovative evidence-graph based framework designed to address the challenges of data synthesis and task verification in agent training. By assembling workspaces of real files and building evidence graphs over their relationships, GraphForge ensures that task requirements are grounded in verifiable data. Initial experiments show that fine-tuning Qwen3.6-27B on trajectories generated by GraphForge leads to significant performance improvements across benchmarks like OpenHands, Work
Core Breakthroughs
GraphForge is an evidence-graph based framework for training working agents, addressing the challenges of data synthesis and task verification. Its key features include:
- Real File Integration: By integrating real files into the task environment, GraphForge ensures the diversity and authenticity of training data.
- Evidence Graph Construction: It builds evidence graphs to describe the relationships between files and generate task statements and rubrics, ensuring the consistency of task requirements and verification standards.
- Task Executability Testing: It tests the executability of tasks in the initial stage and optimizes tasks and rubrics through a revision agent.
- Rejection Fine-Tuning: It performs rejection fine-tuning on the model's own outputs to further improve model performance.
Technical Highlights
- Evidence-Graph Driven: Both task and rubrics are derived from evidence graphs, ensuring the verifiability and consistency of task requirements.
- Diverse Data Generation: It generates diverse workspace files through occupation-grounded seeds, enhancing the authenticity and diversity of training data.
- Intelligent Revision Mechanism: It introduces a revision agent to optimize tasks and rubrics, ensuring task executability and rubric rationality.
- Significant Performance Improvements: Experiments show that fine-tuning Qwen3.6-27B on trajectories generated by GraphForge leads to significant performance improvements across benchmarks like OpenHands, Workspace-Bench-Lite, and SpreadsheetBench II.
Industry Impact
GraphForge represents a significant advancement in the field of agent training, particularly in scenarios requiring complex task handling and multi-file collaboration. Its evidence-graph based framework not only improves the quality of training data but also provides reliable task verification. Developers can leverage GraphForge to build more efficient and reliable agent training pipelines, driving the application and development of agent technology in various industries.
Developer Recommendations
- Explore Application Scenarios: Consider applying GraphForge to scenarios that involve complex file handling and multi-task collaboration, such as document processing and automated office tasks.
- Optimize Training Pipelines: Utilize GraphForge's evidence-graph driven mechanism to optimize agent training pipelines and improve model performance.
- Engage with the Community: Stay updated with GraphForge's community activities, participate in discussions, and access more technical support and resources.
Conclusion
The release of GraphForge marks an important milestone in the field of agent training. Its evidence-graph based design provides a new solution for agent training, enabling developers to build more efficient and reliable agents and promoting the deep application of AI technology in various industries.
— END —Source: Hugging Face Daily Papers (2026-09-30)
Tags: #GraphForge #Agent Training #Evidence Graph #AI Framework #Multimodal
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments