When a vision-language model answers a spatial question — "where is this object in the image" — it takes a clumsy path: autoregressively generate a text string, then parse it back into an option, a scalar, or a 2D point. The output space is often just a handful of options, yet the model pays the full cost of decoding and parsing. MJev-Spatial, submitted to arXiv on October 15, argues this class of questions should be answered by classification, not generation (arXiv:2610.22276).
Spatial Questions as Multiple Choice
The paper unifies three types of spatial understanding — spatial relation QA, numerical estimation, and image pointing — into a single classification framework. Each candidate answer is treated as a class, scored directly with the backbone's pretrained query and key projections, adding zero parameters. For choice tasks the candidates are the given options; numerical estimation uses coarse-to-fine bins over the value range; pointing refines image regions step by step. First author Luozheng Qin and 11 co-authors; the paper runs 37 pages with 27 tables.
One Forward Pass, Zero Answer Tokens
At inference, each decision takes a single forward pass and generates no answer tokens. The official numbers: 5.6x faster than its unmodified autoregressive base model, and 2.5x faster than the fastest autoregressive baseline. For robotics and embodied agents whose spatial judgments must run at interaction rates, this is a qualitative gap, not a marginal optimization.
Scores and Release
The base model is Qwen3.5-4B, trained on 12.3M question-answer pairs converted from existing spatial datasets. In the authors' evaluation, the model achieves the highest overall score (65.64%) among eight tested models across 12 benchmarks, ranks first on seven, and beats its unmodified base on all twelve. Training supervises all coarse-to-fine stages jointly with teacher forcing in one forward pass, and combines length-aware global batching with cross-GPU load balancing, raising training throughput by 30.8%. The delivery is complete: Apache-2.0 code, weights on Hugging Face (mindlab-research/MJev-Spatial-4B), and a GitHub repo with the runtime, training loop, and a public evaluator.
Caveats
The 65.64% figure is a comparison within the authors' own eight-model evaluation, not a public leaderboard rank, and there is no independent replication yet. The deeper limit is scope: classification only works where the output space is compact — open-ended spatial description and reasoning still need generative models. The System 1 fast lane is built; the System 2 road remains where it was.
The takeaway for embodied-AI builders is direct: when the answer space is enumerable, generation is waste. Asking "does this question really need generation?" saves more than upgrading to a bigger model.