SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models
SpatialCORE uses grounding confidence as a reward so vision-language models reason from accurately localized objects.
SpatialCORE is a post-training method for large vision-language models that rewards bounding boxes which are both accurate and high-confidence. A self-regulating spatial reward weights box-matching quality by coordinate-token confidence, and an answer gate links grounding optimization to final-answer correctness. The authors report state-of-the-art results among open-source and specialized spatial-reasoning models, with zero-shot transfer to unseen distributions. Source code is released on GitHub.
- Rewards accurate, confident object grounding rather than final answers alone
- Spatial reward weights box quality by coordinate-token confidence
- An answer gate ties grounding updates to final-answer correctness
- Reports SOTA among open spatial-reasoning models and zero-shot transfer
From the original article
1 min read at huggingface.co
Large Vision-Language Models (LVLMs) have made remarkable progress across visual perception tasks, yet spatial reasoning remains a persistent weakness, especially for questions that require reasoning over visual space. Recent spatial-reasoning methods incorporate generated grounding, where models predict bounding boxes, masks, or other localization outputs for task-relevant objects as part of their reasoning trace.
By Rafi Ibn Sultan, Xiangyu Zhou, Md. Sajid Alam Chowdhury · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.