SATE-Bench Tests Active Spatial Reasoning in Multimodal LLMs
The paper “Humans in Space,” submitted to arXiv on October 25, 2026, introduces SATE-Bench, an interactive benchmark with 316 real-world scenes and 2,525 multi-turn question-answer pairs. It evaluates whether multimodal large language models can explore adjacent viewpoints to overcome occluded visual evidence, using a step-penalized efficiency reward. The authors report that evaluated models often stopped exploring prematurely with high confidence, while GRPO fine-tuning increased exploration but reduced overall accuracy because models over-explored and struggled to fuse multi-view context.
SATE-Bench provides a concrete test of whether multimodal large language models can acquire evidence through active viewing rather than relying only on passive multi-view perception.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.34149 (cs)
[Submitted on 25 Oct 2026]
Title:Humans in Space: Can MLLMs Construct Human-Centric Spatial Understanding via Active Exocentric Exploration?
Authors:Tianchen Guo, Tengyue Zhang, Chen Liu, Ling Chen, Xin Yu
View a PDF of the paper titled Humans in Space: Can MLLMs Construct Human-Centric Spatial Understanding via Active Exocentric Exploration?, by Tianchen Guo and 4 other authors
#SATE-Bench#Multimodal Large Language Models#Group Relative Policy Optimization