MJev-Spatial: System One Spatial Intelligence via Native Attention Matching
MJev-Spatial answers spatial questions without generating answer tokens, reporting a 5.6× speedup over its base model.
Built on Qwen3.5-4B, it classifies choices, numerical values and image locations using existing attention projections, adding no parameters. Numerical and point predictions use coarse-to-fine classification, with each decision requiring one forward pass. Trained on 12.3 million question-answer pairs from existing spatial datasets, it scored 65.64% overall—the highest among eight tested models—and ranked first on seven of twelve benchmarks. It beat its unmodified base model on all twelve benchmarks and ran 2.5× faster than the fastest autoregressive baseline tested.
HF Daily Papers' note
Built on Qwen3.5-4B, it classifies choices, numerical values and image locations using existing attention projections, adding no parameters. Numerical and point predictions use coarse-to-fine classification, with each decision requiring one forward pass. Trained on 12.3 million question-answer pairs from existing spatial datasets, it scored 65.64% overall—the highest among eight tested models—and ranked first on seven of twelve benchmarks. It beat its unmodified base model on all twelve benchmarks and ran 2.5× faster than the fastest autoregressive baseline tested.
HF Daily Papers' note
score 5