ZJU's Scene-Designer Beats GPT-5.2 by 61 Points on 4-Scene Ranking for Agentic 3D Scene Generation

Zhejiang University's Scene-Designer formalizes 'design intelligence' — the joint ability to build and critique 3D indoor scenes — in an agentic Qwen3-VL-8B framework. Trained on Scene-OTA-35K (35K Observation-Thought-Action trajectories) and Scene-OTC-15K (15K perturbation-ranked comparisons), it reaches 89% pairwise-critique accuracy vs 72% for GPT-5.2 and 76% on four-scene ranking vs just 15% for GPT-5.2. Code, project page and weights released.