MJev-Spatial Turns Spatial Understanding into Direct Decisions

Introduction

Vision-language models commonly answer spatial questions through autoregressive text generation. That approach is flexible, but many spatial decisions do not require open-ended language. Questions such as which relation holds, how far an object is, or where to place a pointer usually have a compact answer space: a small set of options, a numerical range, or a location in an image. Generating text and parsing it afterward adds latency to a task that may need to run at interaction speed.

MJev-Spatial proposes a different interface. Built on Qwen3.5-4B, it is a System 1 multimodal decision model designed to make spatial judgments through direct classification rather than answer generation.

What the model changes

• One framework for three task types. In multiple-choice spatial question answering, each provided option becomes a candidate class. Numerical estimation is represented with value bins, while image pointing is handled through image regions and coarse-to-fine refinement.
• Native attention matching. Instead of adding a separate parameter-heavy prediction mechanism, MJev-Spatial reuses the backbone’s pretrained Query and Key projections to score the relationship between the model representation and each candidate. The proposed design therefore uses the attention machinery already present in the base model.
• No answer-token decoding. Each classification decision is completed with a single forward pass and produces no answer tokens. For numerical and point prediction, the model progressively narrows the candidate space from coarse to fine levels.
• Joint training for heterogeneous data. The authors convert existing spatial datasets into 12.3 million question-answer pairs. All coarse-to-fine stages are supervised together with teacher forcing in one forward pass. Length-aware global batching and cross-GPU load balancing reportedly improve training throughput by 30.8% over the original batching strategy.

Results and implications

The project reports a 65.64% mean score across 12 benchmarks and the best result on 7 of them. It also reports a 5.6× runtime advantage over the autoregressive base model. These results support a broader engineering lesson: when the output space is naturally bounded, direct classification can be a better fit than general-purpose text decoding.

The contribution is not merely an optimization of token generation. It reframes how a multimodal model exposes its decision interface. A robot or visual agent may not need a sentence explaining a relation at every interaction step; it may need an option, a distance bin, or a pixel region immediately. A zero-token output path could therefore be useful for high-frequency perception and action loops.

There are also clear boundaries. The quality of a classification formulation depends on how candidates, value bins, and image regions are defined. The approach is naturally suited to structured tasks, while open-ended explanations and unconstrained descriptions still call for generative models. MJev-Spatial is best understood as a specialized fast decision path that can complement, rather than universally replace, autoregressive multimodal reasoning.

Source: Hugging Face Daily Papers