arXiv 2610.21783
Learning to Think and Explore in 3D World
AI89 upvotes
Authors: Yijiang Li, Bingyang Wang, Nuno Vasconcelos
Overview
The authors train vision-language models to reason about 3D scenes by giving them one first-person image at a time and letting them navigate a 3D environment through actions such as moving left. The model is first finetuned on trajectories, then trained with reinforcement learning that is rewarded mainly by final-answer correctness, with no direct supervision on which viewpoints to visit. The resulting model, ACT3D, learns its own exploration policy.
Introduction
Vision-language models natively process single images, but many 3D questions require evidence from viewpoints that the initial view does not show.
Results
ACT3D reaches an average accuracy of 53.2% across six 3D benchmarks, improving its base Qwen model by 22.0% and surpassing proprietary Gemini 3.6 Flash. Its gains over the base model grow from 0.6% on questions answerable from the initial view to 16.3% on those requiring extensive exploration.
Conclusion
Outcome-based RL can teach VLMs not only to reason over visual evidence but also when and how to acquire it.