AI
GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 02/10/2026, 11:53:57 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 02:18:24 UTC.
Briefings
GraphForge is an evidence-graph framework that grounds tasks and verification rubrics in real files, improving agent performance on benchmarks. Training Qwen3.6-27B on 2,169 trajectories improved GDPVal by 65.7 points under OpenHands and Workspace-Bench-Lite by 7.7 points under Claude Code. Rejection fine-tuning further enhanced results, indicating rubrics act as selection signals. The data and models are available.
Key points
- GraphForge uses occupation-grounded seeds to assemble a workspace of real files for each task seed, creating an evidence graph over file relations.
- Task statements and rubrics are derived from this graph, anchoring criteria to specific files needed for verification and ensuring task requirements are backed by workspace files.
Context and limitations
The study relies on a single model (Qwen3.6-27B) and limited dataset size (2,169 trajectories). Results are from a single rollout and revision agent, with no long-term stability or generalizability to other models or environments confirmed.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 11:53:57 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiebrg7tv4dtthjmuwkmp7i53gs75kjei5t2m7cvfeuw2zzwr376iq