MJev-Spatial: System 1 Spatial AI via Attention

Background and Context

Spatial intelligence lets an agent perceive and act in the physical world. Many of the judgments involved must happen at interaction speed: whether a spatial relation holds, how far apart two objects are, and where an item should be placed. Vision-language models (VLMs) usually answer such questions with free-form text. That choice adds the cost of autoregressive decoding and the extra step of parsing the generated answer.

The output space of these tasks is small. The answer is one option from a short list, one scalar, or one 2D point in an image. Generating a full text passage to deliver such a compact answer is a mismatch between the task and the tool. MJev-Spatial, an arXiv paper, is built around this observation.

The paper describes MJev-Spatial as a System 1 style multimodal decision model. One classification framework covers three task types: multiple-choice selection, numeric prediction and point prediction. The authors list three contributions. First, each candidate is treated as a class and scored with the backbone's own pretrained query and key projections, so no parameters are added. Second, coarse-to-fine classification handles numbers and points. Third, a joint-supervision and batching recipe makes training on mixed spatial data more efficient.

Deep Analysis

How does MJev-Spatial answer spatial questions without generating text? It uses native attention matching. For every candidate, the model computes a match score from the query and key projections that the backbone already has. The highest-scoring candidate is the prediction. Because the weights are pretrained, there is no classification head to learn from scratch. Each decision needs one forward pass and produces no answer tokens. The candidates change with the task. For choice questions, they are the options given in the question. For numeric prediction, they are value bins. For point prediction, they are regions of the image. Precision comes from stepwise refinement. The model first picks a coarse bin or region, then classifies again inside it, until the required precision is reached. Training uses teacher forcing. A single forward pass supervises all coarse-to-fine stages together, so the stages are not recomputed one by one. The authors add length-aware global batching and load balancing across GPUs. Together these reduce wasted compute when samples of very different lengths are mixed. The paper reports that this raises training throughput by 30.8%.

The model is built on Qwen3.5-4B. It was trained on 12.3M question-answer pairs, all converted from existing datasets. This shows the method does not depend on newly collected data. The source summary gives no per-component ablation numbers, so the size of each design choice's contribution cannot be judged from it.

Industry Impact

The evaluation covers 12 spatial benchmarks and eight models, including the original backbone. MJev-Spatial scores 65.64% overall, the highest of the eight. It ranks first on 7 of the 12 benchmarks. It also beats the unmodified autoregressive base model on all 12. The last result carries the most weight. The two models share a backbone. They differ only in how they answer and how they are trained. So the comparison suggests that replacing text generation with classification scoring did not cost accuracy on these tasks. It improved accuracy.

Speed is the second result. MJev-Spatial runs 2.5x faster than the fastest autoregressive baseline and 5.6x faster than the base model. For robot manipulation, embodied agents and real-time visual assistants, low latency is a hard requirement. Decoding delay and fragile answer parsing are real obstacles in those settings. When the output space is compact, treating a VLM as a classifier is both faster and, on these benchmarks, more accurate. It also removes the need to parse free text. The design may also transfer well. The method adds no parameters and reuses existing query and key projections, so it may port to other backbones with modest effort. That point is an inference from the design. The paper's reported results use Qwen3.5-4B.

Outlook

The coarse-to-fine scheme gives continuous outputs such as distances and coordinates a path that avoids text generation. It could extend to other vision tasks that need precise localization or estimation. The reported training cost is also moderate: existing datasets supplied all 12.3M pairs, and the batching changes raised throughput by 30.8%.

Open questions remain. The source does not report results on larger backbones, on more task types, or inside a live interactive system. Those are the natural next tests. A fair check would also compare the fine-grained precision of numeric and point outputs under different bin settings, which the summary does not detail.

For teams that build agents with tight latency budgets, the practical message is narrow but clear. When the answer is a choice, a number or a point, a single scoring pass over candidates can replace token-by-token generation. MJev-Spatial reports a 65.64% overall score and a 5.6x speedup over its own base model as evidence for that claim.