SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models
Rafi Ibn Sultan, Xiangyu Zhou, Md. Sajid Alam Chowdhury, Chengyin Li, Prashant Khanduri, Marco Brocanelli, Dongxiao Zhu
Wayne State University · Henry Ford Health · The Ohio State University
Published September 30, 2026 · on the feed October 1, 2026 · ▲ 3 · arXiv 2609.38716
The abstract
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. However, these approaches typically optimize final-answer correctness alone, allowing correct answers to be rewarded even when the model does not reason from confidently localized task-relevant objects. We introduce SpatialCORE (Spatially COnfident REasoning), a post-training framework that turns the model's own confidence in generated grounding into a learning signal for spatial reasoning. Its central idea is to reinforce grounding that is both accurate and confident, encouraging the model to reason from confidently localized task-relevant objects. SpatialCORE realizes this through a self-regulating spatial reward that weights each predicted bounding box's matching quality by its coordinate-token confidence. An answer gate further ties grounding optimization to final-answer correctness. SpatialCORE achieves state-of-the-art results among open-source and specialized spatial reasoning models across diverse benchmarks, and transfers effectively in zero-shot settings to unseen data distributions. The source code is available at https://github.com/rafiibnsultan/SpatialCORE.
The Brief
2:16A short podcast summary of the paper. Written and voiced by AI.
Transcript
SpatialCORE. Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models. By Rafi Ibn Sultan and six others, from Wayne State University, Henry Ford Health, and one other institution. Published on the thirtieth of September. Its code repository is a placeholder: a readme, and nothing to run yet.
Large vision-language models have advanced quickly on visual perception, but spatial reasoning is still a persistent weakness, particularly for questions that require reasoning over visual space. A recent line of methods adds generated grounding to the reasoning trace: the model predicts bounding boxes, masks, or other localization outputs for the objects the task depends on. Those methods, however, usually optimize only whether the final answer is right. A correct answer then earns its reward even if the model never reasoned from task-relevant objects it had confidently localized. The answer is rewarded, and the grounding behind it goes unexamined.
SpatialCORE, short for Spatially COnfident REasoning, is a post-training framework that turns the model's own confidence in its generated grounding into a learning signal. The central idea is to reinforce grounding that is both accurate and confident, pushing the model to reason from objects it has localized with confidence rather than from wherever it happened to land.
That idea is carried by a self-regulating spatial reward. For each predicted bounding box, the reward takes the quality of its match and weights it by the model's confidence in the coordinate tokens it produced. A box that matches well but was produced hesitantly counts for less than one that matches well and was produced with confidence. On top of that sits an answer gate, which ties the optimization of grounding to the correctness of the final answer, so the two objectives are not pursued in isolation from each other.
Across diverse benchmarks, SpatialCORE reaches state-of-the-art results among open-source models and specialized spatial reasoning models. It also transfers effectively in zero-shot settings to data distributions it has not seen. The abstract states these outcomes without specific figures.
One way to hold this paper in mind is as a correction to what gets rewarded. Answer-only training cannot tell a well-grounded correct answer from a lucky one. By pricing confidence into the grounding reward, and gating that reward on the answer being right, SpatialCORE asks for correct answers that rest on objects the model was sure it had found.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A right answer built on wrong boxes. That is the crack this quest starts from.
You are teaching a vision-language model to answer questions like: from the girl's point of view, where is the lamp? It writes bounding boxes for the objects it uses while it thinks.
The road ahead
- 1 · The problem ~7 min When training rewards only the final answer, a lucky right answer pays off even if the boxes the model drew while thinking were poor.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Reward each box by how well it matches the object, weighted by how confident the model was in writing it, and tie that to the answer.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The model writes boxes while reasoning; each is matched to an object, weighted by its confidence, and the result is gated on the answer.
- 6 · The experiment ~6 min You’ll guess: Against open-source models and specialized spatial reasoning models, where did SpatialCORE land across diverse benchmarks?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.