AI

DeskForge: Dense Supervision from Desktop Environments for Computer-Use Agents

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 16:18:24 UTC.

Briefings

DeskForge introduces a controllable desktop environment to generate large-scale supervision for computer-use agents. By composing real applications, varying states, and fusing screenshots with accessibility data, the project created DeskForge-1M, a corpus of 1.2M annotated observations containing 159.7M element instances. Fine-tuning four vision-language models on 200K grounding examples improved performance across held-out conditions and five external benchmarks, with Qwen3.5-4B accuracy increasing by 11.51% on ScreenSpot-Pro and 10.11% on OSWorld-G. Long-horizon task completion also improved, with Qwen3.5-4B solving 50% more of 119 tasks and 15% more of 100 tasks under a fixed planner.

Key points

• DeskForge-1M is a corpus of 1.2M annotated desktop observations containing 159.7M element instances generated by varying application states, content, window layouts, and resolutions.
• Fine-tuning four vision-language models on 200K grounding examples from DeskForge-1M improved accuracy by 11.51 percentage points on ScreenSpot-Pro and 10.11 points on OSWorld-G for Qwen3.5-4B.

Context

Existing training data rarely pairs complex desktop scenes with dense annotations or varies them in a controlled way, making it difficult for agents to reliably ground action targets. DeskForge addresses this by providing a controllable environment that composes real applications to generate large-scale supervision for computer-use agents.

Limitations

The evidence covers a specific dataset and model performance on five external benchmarks; it does not include results from other datasets or models, nor does it address long-horizon task completion under different planning strategies or varying agent architectures.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/10/2026, 06:00:01 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiaxyhficfd54iqvqzxtrfivwtnglu7fuko7hnen6lg6jllcy6bneu