MJev-Spatial: System One Spatial Intelligence via Native Attention Matching

MJev-Spatial. System One Spatial Intelligence via Native Attention Matching. By Luozheng Qin and eleven others, from Mind Lab. Published on the fifteenth of October. Its code repository is public and real: nineteen source files.

Spatial intelligence is what lets an agent perceive and act in the physical world, and many of its judgments have to be made at interaction rates. Which spatial relation holds. How far away an object is. Where to place something. Vision-language models answer questions like these by free-form text generation. That means paying for autoregressive decoding, and then for parsing an answer out of the text, even though the space of possible outputs is compact: one of a few options, a single scalar, or a point in two dimensions. The cost of the answer is set by the generation process, not by the size of what is being decided.

MJev-Spatial is a System One multimodal decision model. It handles all three kinds of spatial understanding task inside a single classification framework, where each candidate answer is treated as a class. The candidates are scored with the backbone's own pretrained query and key projections, so no parameters are added. What counts as a candidate depends on the task. For choice tasks, the candidates are the given options. For numerical prediction they are value bins, and for point prediction they are image regions, both refined through coarse-to-fine classification.

At inference, each classification decision takes a single forward pass and generates no answer tokens. Training raises its own efficiency problem, because the spatial data is heterogeneous. All of the coarse-to-fine stages are supervised jointly, with teacher forcing, in one forward pass. That is combined with length-aware global batching and cross-GPU load balancing, which together raise training throughput by thirty point eight percent over the original batching strategy.

The model is built on Qwen3.5-4B and trained on twelve point three million question-answer pairs, converted from existing spatial datasets. Among eight tested models it achieves the highest overall score, sixty-five point six four percent, and it ranks first on seven of the twelve benchmarks.

Against its own unmodified autoregressive base model, it comes out ahead on all twelve. It is also faster. It runs two point five times faster than the fastest autoregressive baseline, and five point six times faster than the base model it started from. Accuracy and speed improve together rather than being traded against each other.

The argument rests on a mismatch between the question and the machinery. When an answer is one of a few options, a number, or a point, generating text to express it spends decoding on something a single scoring pass can deliver. MJev-Spatial recasts every such answer as a choice among candidates, scores those candidates with attention projections the backbone already has, and in the reported results ends up both more accurate and faster than the model it began as. The phrase System One in the title names that ambition: quick, direct judgments, made at the pace an agent acting in the world requires.