TL;DR Treating spatial reasoning as a classification problem over finite candidates, instead of generating text, speeds up a VLM by up to 5.6x while also improving accuracy. Title: MJev-Spatial: System One Spatial Intelligence via Native Attention Matching URL: Here are the key points. ・🎯 Choices, numeric values, and point coordinates are all unified as classification over a finite candidate set ・🧠 No new prediction head is added - it reuses the model's existing query-key attention projections for scoring ・📐 Point prediction uses a 3x3 grid refined through 3 rounds of re-cropping and re-encoding ・🚀 160.6ms average latency on a B200 GPU, 2.5x faster than the best autoregressive baseline and 5.6x faster than the base model ・📊 65.64% average across 12 benchmarks, beating Qwen3-VL-8B and others, ranking #1# on 7 of them ・🔢 Surprisingly, coarser distance bins actually beat finer ones in accuracy ・🌲 A probability-tree decoding scheme extends the single-point predictor to multi-point prediction with no extra training What's interesting is how this shift away from text generation toward a more "intuitive" classification approach delivers both speed and accuracy at once. #VLM# #SpatialReasoning#