MJev-Spatial: Single-Pass Classification Beats Autoregressive VLMs on Spatial Reasoning
mindlab-research · hf · 2026-10-16
Mind Lab introduces MJev-Spatial, a System One multimodal decision model for spatial intelligence. The authors note spatial tasks—relation holding, object shape, placement—have compact output spaces (one of a few options, a scalar, a 2D point), yet VLMs pay for autoregressive decoding and answer parsing.
- Method: Unifies spatial reasoning into a single-pass classification framework, treating each candidate as a class scored with the backbone's pretrained query/key projections—no added parameters. Numerical and point predictions use value bins, and image regions are refined via coarse-to-fine classification. Each decision takes one forward pass and generates no answer tokens.
- Training: Built on Qwen3.5-4B, trained on 12.3M QA pairs converted from existing spatial datasets; all coarse-to-fine stages are supervised jointly with teacher forcing in one forward pass, with length-aware global batching and cross-GPU load balancing raising throughput by 30.8%.
- Results: Improves both speed and accuracy over autoregressive vision-language models.
More from Multimodal
- Kandinsky Lab open-sources S2V 1.0: audio-driven avatar video from one still, MIT licensed — lmoroney · 2026-10-16
- Meme: Flux 3 Klein 5.6B steals the spotlight from Flux 2 Klein 9B — xb1n0ry · 2026-10-16
- Pichai amplifies user demo: Gemini Argon builds Asuka-period pagoda in 11 hours — sundarpichai · 2026-10-16
- Midjourney style code --sref 1127756965 yields bold collage-style illustrations — sergeantsref · 2026-10-16
- A cinematic ocean-to-cliff drone-style prompt worth trying on video models — umesh_ai · 2026-10-16
- Flux.3 klein vs Qwen Image 2.1 turbo on RTX 4070: 10s vs 3s per 1MP edit — Ant_6431 · 2026-10-16