New LBS model trains vision-language models for active visual observation

Researchers have developed a new method called LBS (Look Beyond Screens) that trains vision-language models to perform active, task-oriented visual observation. This approach uses GUI grounding trajectories as a source of spatial supervision, initially through supervised fine-tuning and then through reinforcement learning for multi-turn inspection decisions. Models trained with LBS, including those based on Qwen3.5-35B-A3B and Qwen3.8-27B, show improved performance and generalize to benchmarks like PerceptionBench, HR-Bench, and ChartQAPro, demonstrating effective observation strategies beyond simple imitation. AI

IMPACT This research advances transferable visual perception in AI, potentially improving agent capabilities in complex interactive environments.

RANK_REASON The cluster contains an academic paper detailing a new model and methodology for visual observation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• ChartQAPro
• HR-Bench
• Hugging Face
• PerceptionBench
• Qwen3.5-35B-A3B
• Qwen3.8-27B

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →