Megadose The AI curve, compressed.

SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models

· HF Daily Papers ·
SpatialCORE uses a model’s confidence in its own bounding-box grounding as the reward signal for better spatial reasoning.

The paper targets LVLM failures on questions that require reasoning over visual space. Its post-training method rewards grounding that is both accurate and confident, weighting box matches by coordinate-token confidence. A separate answer gate keeps that grounding reward tied to final-answer correctness. The authors report state-of-the-art results among open-source and specialized spatial reasoning models, plus zero-shot transfer to unseen data distributions. HF Daily Papers' note

score 4

Categories: Research