MJev-Spatial: spatial AI that picks answers instead of typing them
Why it matters
Mind Lab researchers' MJev-Spatial (arXiv, Oct 15) answers spatial questions without generating answer text: each candidate answer (an option, a distance bin or an image region) is treated as a class and scored with the backbone's own pretrained query and key projections, adding no new parameters. Multiple choice takes 1 decision, distances 2 (coarse then fine bins), pointing 3 (zooming into a 3x3 grid), and each decision is one forward pass with no answer tokens. Built on Qwen3.5-4B and trained on 12.3M spatial Q&A pairs, it had the highest average of 8 tested models (65.64% vs 62.79% for Qwen3-VL-8B-Instruct), ranked first on 7 of 12 benchmarks and beat its base model on all 12. Mean latency was 160.6 ms per query on one Nvidia B200 at batch size 1: 2.5x faster than the fastest autoregressive baseline and 5.6x faster than its base model. Apache 2.0 code; the 4B model is on Hugging Face. Caveat: a preprint with the authors' own results and timings, not independent tests.
Read the source: arXiv (Mind Lab) (opens in a new tab)
All hits from 19 Oct 2026More VisionGet the daily hits by email