Howard Chu @ Symas

Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

https://aim-uofa.github.io/reasonmatch/

Outperforms all other open- and closed-source spatial reasoning models, but still inferior to human accuracy. Uses #LMDB

ReasonMatch — Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

CVPR 2026 · ReasonMatch-Bench (2,810 image pairs) and DCRL reach 70.5 F1, outperforming evaluated open- and closed-source MLLM baselines.

Jun 7 at 1:45amWeb