MJev-Spatial · System One Spatial Intelligence

System One Spatial Intelligence
via Native Attention Matching

Luozheng Qin^*Yuhan Zhu^*Changhai Zhou^*Qian QiaoJun GaoYuhua ZhouHarry ZhangXiangyu ZengYuxuan LiuKaijie ChenAndrew ChenPony Ma

MindLab

* Equal contribution October 2026

Paper ↗ Code ↗ Model ↗

Overview

MJev-Spatial is a System 1 multimodal decision model for spatial intelligence, unifying option selection, numerical estimation, and image pointing through direct classification. Conventional vision-language models solve these compact-output tasks by generating and parsing answer text. We instead treat each candidate as a class and score it using the backbone’s pretrained query and key projections, without adding a prediction head. Numerical values and image locations are refined through coarse-to-fine classification. Each decision takes a single forward pass and generates no answer tokens.

Built on Qwen3.5-4B and trained on 12.3M question–answer pairs converted from existing spatial datasets, MJev-Spatial achieves the highest overall score (65.64%) among eight evaluated models, ranks first on seven of twelve benchmarks, and outperforms its autoregressive base model on all twelve. It runs 2.5× faster than the fastest autoregressive baseline and 5.6× faster than the base model. Joint supervision of all refinement stages, length-aware global batching, and cross-GPU load balancing improve training throughput by 30.8%.

65.64%

Mean score across 12 benchmarks

5.6× faster

vs. autoregressive Qwen3.5-4B

7 / 12 benchmarks led

Improves on the base model in all 12

Method

Direct spatial decisions through unified classification, native attention matching, and joint stage supervision.

Unified classification for spatial intelligence

Given visual observations and a spatial question, MJev-Spatial casts prediction as classification over a finite candidate set. Each class represents a supplied answer option, a numerical interval, or an image region. The shared scorer evaluates all legal candidates in one forward pass; a softmax gives their probabilities, and the highest-scoring class determines the answer or the next refinement stage.

Choice tasks finish in one decision. Numerical prediction uses up to two decisions to refine a coarse interval and decode an exact value or interval midpoint. Single-point prediction makes three successive 3 × 3 region selections. Between stages, the selected region is cropped from the original image and re-encoded to recover local detail, while the full image remains in context. The final cell center gives normalized coordinates on an effective 27 × 27 grid. Expanding multiple region paths extends the same predictor to multiple points without additional training.

Task formulation and coarse-to-fine inference from the paper.

Native attention matching

MJev-Spatial encodes visual tokens, the question, refinement history, and candidate descriptions in one causal sequence. Each candidate ends with an End marker, and a final Decide marker follows the candidate list. After the first 31 language layers of Qwen3.5-4B, the contextualized End states represent the candidates, while the Decide state represents the decision.

We reuse the final full-attention layer’s pretrained query and key projections: the decision state supplies a query, and each candidate state supplies a key. The scaled query–key inner products are averaged across attention heads and normalized with a softmax over the current candidates. Native normalization and rotary position transformations are retained. This reuses pretrained representations and matching projections without adding a prediction head or new trainable parameters.

MJev-Spatial architecture with native final-layer query and key projections.

Joint stage supervision and efficient training

We convert existing spatial annotations into 12.3M training QA pairs. Ground-truth answers determine the target classes and crops at every refinement stage, allowing teacher forcing to supervise all stages jointly in one causal forward pass. Stage cross-entropy losses are averaged per question so that multi-stage questions receive the same total weight as one-stage choices. We train all retained language parameters while freezing the vision encoder and merger. Global length grouping and cross-GPU load balancing reduce padding and workload imbalance, improving training throughput by 30.8%.

Results

MJev-Spatial achieves the strongest performance–latency trade-off among the eight evaluated models, supporting our System 1 approach to spatial intelligence. It obtains the highest mean score of 65.64%, ahead of 62.79% for Qwen3-VL-8B-Instruct, and ranks first on seven of twelve benchmarks. It also outperforms its autoregressive base model on all twelve, with gains of 22.89 percentage points on RoboSpatialVQA and 12.50 points on RefSpatial-Bench, spanning spatial question answering and point localization. On an NVIDIA B200, MJev-Spatial achieves the lowest mean query latency of 160.6 ms, yielding a 2.5× speedup over the fastest autoregressive baseline and a 5.6× speedup over its base model. These results support the central premise of our System 1 design: adapting pretrained multimodal knowledge into direct classification decisions enables accurate spatial judgments with fast responses, without autoregressive answer decoding or text parsing.

Scores on twelve benchmarks, overall score, and mean query latency for eight models. Bold marks the best value in each column.
Model	Benchmark scores (%) ↑												Overall (%) ↑	Mean latency (ms) ↓
Model	SAT	CV- Bench	Robo- VQ	BLINK	MMSI	Ref Spatial	Where 2Place	Robo- POI	Point- Bench	ERQA	Emb Spatial	VSI- Bench	Overall (%) ↑	Mean latency (ms) ↓
MJev-Spatial	79.33	88.82	79.06	67.17	45.60	58.50	67.96	55.74	64.38	46.75	78.46	55.93	65.64	160.6
Qwen3-VL-8B-Instruct	83.33	86.37	74.72	61.76	31.40	49.00	67.07	59.02	61.54	43.25	78.90	57.07	62.79	442.0
Spatial-SSRL-Qwen3VL-4B	79.67	85.52	65.26	60.39	26.70	48.75	66.54	61.48	60.60	41.25	78.10	46.65	60.08	397.4
Molmo2-ER	80.00	88.31	62.01	72.17	42.90	52.00	46.64	26.23	53.47	46.50	77.99	60.55	59.06	777.3
SenseNova-SI Qwen3-VL-8B	79.67	83.15	75.68	56.08	35.30	33.00	45.92	45.90	52.34	41.00	75.33	59.38	56.90	749.9
Qwen3.5-4B (base model)	43.00	85.56	56.17	63.11	33.90	46.00	55.98	50.00	62.65	45.50	75.82	39.93	54.80	894.1
InternVL3.5-8B	83.00	74.66	71.67	56.08	29.90	23.80	27.48	14.75	32.93	39.00	72.14	53.46	48.24	3,368.4
SpatialThinker-7B	78.00	78.17	64.94	56.02	27.80	0.50	2.66	13.93	14.46	40.25	69.89	29.56	39.68	406.7

Citation

@article{qin2026mjevspatial,
  title = {MJev-Spatial: System One Spatial Intelligence
           via Native Attention Matching},
  author = {Qin, Luozheng and Zhu, Yuhan and Zhou, Changhai
            and Qiao, Qian and Gao, Jun and Zhou, Yuhua
            and Zhang, Harry and Zeng, Xiangyu and Liu, Yuxuan
            and Chen, Kaijie and Chen, Andrew and Ma, Pony},
  journal = {arXiv preprint arXiv:2610.22276},
  year = {2026},
  url = {https://arxiv.org/abs/2610.22276}
}

Download .bib ↓