CONSONANCE.for your information
Monday, 5 October 2026frenvi

Worth reading closely

01 — training 144 upvotes

GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

QUESTION — How can evidence-graph based workspace data be synthesized from real files to train working agents effectively?

They present GraphForge, an evidence-graph based framework that grounds both tasks and verification in real files to train working agents. Starting from occupation-grounded seeds, GraphForge assembles a workspace of real files, builds an evidence graph, and employs a revision agent to repair tasks before trajectory collection. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories brings GDPVal to 1445.7 (+65.7) under OpenHands, and Workspace-Bench-Lite and SpreadsheetBench II to 63.7 (+7.7) and 24.0 (+13.7) under Claude Code. Rejection fine-tuning on the SFT model's own rollouts yields further improvements on all three benchmarks.

Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories brings GDPVal to 1445.7 (+65.7) under OpenHands.

Workspace-Bench-Lite and SpreadsheetBench II reach 63.7 (+7.7) and 24.0 (+13.7) under Claude Code, respectively.

Rejection fine-tuning on the SFT model's own rollouts yields further improvements on all three benchmarks.

groundhogLLM · 30 Sept 2026 read the original ↗
↑