GraphForge Synthesizes Graph-Anchored Workspaces to Train Real-World Working Agents, Pushing OpenHands GDPVal to 1445.7

核心背景与行业痛点

Modern working agents—such as Devin, OpenHands, and Claude Code—are tasked with reading heterogeneous files, orchestrating tools, and delivering functional outputs across enterprise workspaces. Training these agents demands tasks situated in rich multi-file workspaces paired with deterministic verification rubrics. Existing data generation recipes face an acute tradeoff: fully synthetic files generated by LLMs lack real-world messiness and diversity, while crawling organic repositories leaves tasks without automated, task-specific ground-truth verifiers, precluding reliable post-training alignment.

架构亮点与底层机制

Researchers from Fudan University and USTC introduce GraphForge (arXiv:2609.38923), grounding workspace synthesis in an evidence graph:

1. Occupation-Grounded Workspace Seeding: Assembles workspaces of authentic real files spanning code, structured spreadsheets, and configuration scripts seeded by occupational domain requirements.
2. Topological Evidence Graph Construction: Maps dependency references, variable schemas, and cross-file relationships into an explicit semantic evidence graph.
3. Dual-Anchored Formulation: Simultaneously generates task statements and deterministic evaluation rubrics directly from the evidence graph, anchoring every evaluation criterion to verifiable workspace ground-truths.
4. Initial Rollout & Revision Pipeline: Evaluates task feasibility with an execution probe and employs an automated revision agent to refine tasks and repair rubrics prior to trajectory collection.

权威 Benchmark 与实测跑分对比

Benchmarked across demanding interactive agent environments:

1. 65.7-Point Leap on OpenHands GDPVal: Fine-tuning Qwen3.6-27B on only 2,169 GraphForge trajectories pushes GDPVal to 1445.7 (+65.7) within the OpenHands harness.
2. Up to 13.7-Point Boost Under Claude Code: Catapults Workspace-Bench-Lite to 63.7 (+7.7) and SpreadsheetBench II to 24.0 (+13.7) in the Claude Code runtime.
3. Rejection Fine-Tuning (RFT) Gains: Rejection fine-tuning using evidence-anchored rubrics yields consistent secondary improvements, demonstrating the fidelity of GraphForge's automated reward signal.

开发者实战落地与开箱指南

GraphForge datasets, synthesis engines, and checkpoints are publicly accessible. For AI teams building enterprise automation agents and full-stack software engineers, GraphForge establishes a reproducible framework to synthesize high-fidelity multi-file environments, accelerating agentic capabilities without human verification bottlenecks.