Humans in Space: Can MLLMs Construct Human-Centric Spatial Understanding via Active Exocentric Exploration?

Humans in Space. Can MLLMs Construct Human-Centric Spatial Understanding via Active Exocentric Exploration?. By Tianchen Guo and four others, from the University of Queensland, the Australian Institute for Machine Learning at Adelaide University, and two other institutions. Published on the twenty-fifth of October. Its code repository is public and real: about twenty-five source files.

Multimodal large language models are strong zero-shot reasoners, but their spatial abilities are mostly tested through static, passive observation. The model is handed a view and asked a question. In complex scenes full of people, occlusion often hides exactly the geometric evidence the question depends on. A passive benchmark then forces the model to guess, rather than to recognise that it cannot see enough. A right answer reached by guessing and a right answer reached by seeing look the same on such a test, and so does a model that has no idea it is missing information. What passive benchmarks cannot ask is whether the model would go looking if it could.

SATE-Bench is an interactive benchmark aimed at two things together. One is visual self-awareness: does the model know when its evidence is insufficient? The other is active exocentric view-exploration: can it go and get more? The name plays on the word sate. The model should keep looking until it has gathered enough evidence to fully satisfy the question, and then stop.

The benchmark holds three hundred and sixteen real-world scenes and two thousand five hundred and twenty-five multi-turn question-answer pairs. It runs in two settings. A Passive setting gives the model multiple views as a control. In the Active setting, the model starts from a severely occluded view and must iteratively query adjacent viewpoints until it has what it needs. Answers are scored with a step-penalized efficiency reward, so needless looking costs something too. The reward therefore pulls in two directions at once. Stopping too early risks a wrong answer, and wandering through every viewpoint costs steps.

State-of-the-art MLLMs show a consistent pattern. They end exploration prematurely while reporting high confidence, and they struggle with fine-grained 3D geometric tasks. Matched-view controls let the authors locate the cause of the gap between Active and Passive: it comes mainly from insufficient evidence acquisition. The models are not failing to reason over the right views so much as failing to collect them.

The authors then tried fine-tuning with Group Relative Policy Optimization, GRPO, rewarding view-seeking. That did increase active exploration, but overall accuracy went down. The models now over-explored, and failed to fuse the growing multi-view context across turns.

Put together, the two failures bracket the problem. Models as evaluated stop too soon, and models rewarded for looking look too much and lose track of what they saw. The authors conclude that scaling passive multi-view perception does not by itself produce active spatial reasoning. What is needed are models that know where to look, how to integrate what they see, and when to stop. Code and data are released.