Zhejiang Univ Introduces ProVisE: Evaluating Spatial Cognition via Pixels
机器之心 · wechat · 2026-08-08
The OmniAI team at Zhejiang University introduced the ProVisE framework, shifting away from traditional coordinate or text-based evaluations of AI spatial cognition by having generative image models directly "draw" their answers.
ProVisE Framework
- Visual Protocols: Models are guided to mark targets, generate depth maps, or draw trajectories directly on images. A parser then converts these visual answers into structured data for scoring.
- Agentic Builder: An automated framework that scans task inputs and scoring rules to construct and validate corresponding "generation-parsing" protocols.
SpatialGen-Bench
The team built a comprehensive benchmark covering 14 sub-tasks across four levels: perception, understanding, reasoning, and interaction.
Key Findings
- Generative Strengths: Generative models exhibit stronger "spatial intuition" in tasks requiring direct visual perception, showing significant complementarity with text-based models.
- VLM Strengths: Text-output VLMs maintain an edge in tasks requiring abstract reasoning, such as counting or geometric feasibility checks.
Related event: Zhejiang Univ Introduces ProVisE for AI Spatial Reasoning(2 posts)→
More from Multimodal
- Kandinsky Lab open-sources S2V 1.0: audio-driven avatar video from one still, MIT licensed — lmoroney · 2026-10-16
- Meme: Flux 3 Klein 5.6B steals the spotlight from Flux 2 Klein 9B — xb1n0ry · 2026-10-16
- Pichai amplifies user demo: Gemini Argon builds Asuka-period pagoda in 11 hours — sundarpichai · 2026-10-16
- Midjourney style code --sref 1127756965 yields bold collage-style illustrations — sergeantsref · 2026-10-16
- A cinematic ocean-to-cliff drone-style prompt worth trying on video models — umesh_ai · 2026-10-16
- Flux.3 klein vs Qwen Image 2.1 turbo on RTX 4070: 10s vs 3s per 1MP edit — Ant_6431 · 2026-10-16