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

> MJev-Spatial unifies spatial tasks as candidate classification using pretrained attention projections, achieving state-of-the-art accuracy with 5.6x faster inference than autoregressive baselines.

- **Source:** [arXiv](https://arxiv.org/abs/2610.22276)
- **Published:** 2026-10-17
- **Permalink:** https://picx.dev/p/k6e35W
- **Whiteboard:** https://picx.dev/p/k6e35W/image

## Summary

## Summary (Overview)

- **MJev-Spatial** introduces a System 1 multimodal decision model for spatial understanding that unifies three task types—option selection, numerical estimation, and image pointing—within a single classification framework, eliminating autoregressive text generation.
- The model achieves the highest overall score (65.64%) among eight evaluated models, ranks first on seven of twelve spatial benchmarks, and outperforms its autoregressive base model (Qwen3.5-4B) on all twelve benchmarks.
- MJev-Spatial achieves 2.5× faster inference than the fastest autoregressive baseline and 5.6× faster than its base model, with a mean query latency of 160.6 ms on a B200 GPU.
- The approach uses pretrained query/key projections from the backbone's final attention layer for candidate scoring, requiring no additional prediction heads or trainable parameters.
- Training on 12.3M QA pairs converted from existing spatial datasets uses joint stage supervision with teacher forcing, plus length-aware global batching and cross-GPU load balancing that improves training throughput by 30.8%.

## Introduction and Theoretical Foundation

Spatial intelligence enables agents to perceive and reason about geometry and spatial relationships, requiring timely judgments about spatial relations, distances, and object placement. Traditional vision-language models (VLMs) handle these tasks through free-form text generation, which introduces two key inefficiencies:

1. **Supervision mismatch**: Learning the spatial answer is coupled with learning its textual form (formatting, boilerplate), adding supervision beyond the task-relevant prediction.
2. **Inference overhead**: Autoregressive decoding and answer parsing add latency even though the output space is compact (one of a few options, a scalar, or a 2D point).

The paper observes that all three common spatial output types can be represented as **finite candidate sets**: answer options are already finite, while numerical values and image locations can be discretized within specified ranges and resolutions. This motivates a System 1 approach inspired by Jev (TypeSafe AI, 2026), which returns probabilities over specified candidates without generating answer text.

The theoretical foundation builds on several insights:
- **Candidate classification**: All spatial predictions can be cast as classification over finite candidate sets (options, numerical bins, or image regions).
- **Pretrained feature reuse**: The backbone's pretrained query and key projections can serve as a classifier without additional parameters.
- **Coarse-to-fine refinement**: Hierarchical classification (coarse intervals/regions → fine cells) provides precision without enumerating large candidate sets.

## Methodology

### Unified Classification Framework

MJev-Spatial treats each candidate as a class and scores it using the backbone's pretrained query and key projections. The classification rule is:

$$p_{\theta}(i \mid I, q, h_{<t}, \mathcal{C}_t) = \frac{\exp z_{t,i}}{\sum_{j=1}^{K_t} \exp z_{t,j}}, \qquad \hat{a}_t = \underset{i}{\arg \max} \, p_{\theta}(i \mid I, q, h_{<t}, \mathcal{C}_t) \tag{1}$$

**Task-specific candidate structures** (Table 1):

| Output | Candidates per stage | Stages | Decoded answer |
|--------|---------------------|--------|----------------|
| Choice | Supplied option contents | 1 | The selected option |
| Length (m) | 32 roots; 4 or 8 fine cells | 1–2 | Zero, overflow, or fine-cell midpoint |
| Count | 32 coarse groups; 32 integers | 2 | $32a_1 + a_2$, covering 0–1023 exactly |
| Angle (deg) | 14 roots; 32 fine cells | 1–2 | Zero, overflow, or degree interval midpoint |
| Point | 9 spatial cells at each level | 3 | Center of the final cell in original image coordinates |

**Length codebook** uses log1p-spaced coarse intervals with $U = 16384 \, \text{m}$, $s = 0.01 \, \text{m}$:

$$e_j = s \left[ \exp\left(\frac{j}{30} \log(1 + U/s)\right) - 1 \right], \qquad j = 0, \ldots, 30 \tag{2}$$

Fine cells split each interval linearly with adaptive count $F_j$ (4 or 8), satisfying:

$$\frac{e_{j+1} - e_j}{2F_j} \leq \max(0.01 \, \text{m}, 0.05 e_j) \tag{3}$$

Decoding uses the fine-cell midpoint:

$$\hat{y} = e_j + \frac{k + 1/2}{F_j}(e_{j+1} - e_j) \tag{4}$$

**Point prediction** uses three 3×3 grid decisions with crop re-encoding between stages. The rectangle update rule:

$$B_{t+1} = \left[ x_t + \frac{c w_t}{3}, \, y_t + \frac{r h_t}{3}, \, x_t + \frac{(c+1) w_t}{3}, \, y_t + \frac{(r+1) h_t}{3} \right] \tag{5}$$

This yields an effective $27 \times 27$ grid with a correct-cell center error of at most 1/54 of image dimensions per axis.

### Model Architecture

Built on **Qwen3.5-4B**, MJev-Spatial:
- Encodes candidates as text with an option-end marker, followed by a decision marker
- Uses the hidden state at each option-end marker as a candidate key
- Uses the hidden state at the decision marker as the query
- Computes scores **before** the final MLP (motivated by findings that the final MLP shifts representations toward lexical prediction)

The scoring mechanism (Equation 6):

$$z_{t,i} = \frac{1}{H} \sum_{h=0}^{H-1} \frac{Q_{t,h}^{\top} K_{t,i,g(h)}}{\sqrt{d}}, \quad H = 16, \quad d = 256, \quad g(h) = \lfloor h/4 \rfloor \tag{6}$$

Each group of four query heads shares one of four key heads. The vision encoder and merger remain frozen.

### Training Recipe

- **12.3M QA pairs** converted from RefSpatial, SAT, SenseNova-SI, VSI-590K-image, and VST-P
- **Joint stage supervision** with teacher forcing in one forward pass; the loss is:

$$\mathcal{L} = \frac{1}{N_{\text{actual}}} \sum_q \frac{1}{S_q} \sum_{t=1}^{S_q} -\log p_{\theta}(a_{q,t}^{*} \mid I_q, q, h_{q,<t}^{*}, \mathcal{C}_{q,t}) \tag{7}$$

- **Length-aware global batching** and **cross-GPU load balancing** (longest-work-first assignment) to handle heterogeneous sequence lengths
- Full-parameter fine-tuning with AdamW: learning rates $5 \times 10^{-6}$ (backbone) and $2 \times 10^{-4}$ (readout), BF16 language computation with FP32 trainable weights

## Empirical Validation / Results

### Main Results (Table 2)

| Model | SAT | CV-Bench | Robo-VQ | BLINK | MMSI | RefSpatial | Where2Place | Robo-POI | Point-Bench | ERQA | EmbSpatial | VSI-full | **Overall** |
|-------|-----|----------|---------|-------|------|------------|-------------|----------|-------------|------|------------|----------|-------------|
| **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** |
| 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 |
| Qwen3.5-4B | 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 |
| 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 |

MJev-Spatial outperforms its base model (Qwen3.5-4B) on all twelve benchmarks, including gains of 22.89 points on RoboSpatial VQA and 12.50 points on RefSpatial-Bench.

### Key Ablation Findings

**Pretrained candidate scoring (Table 3)**: Native Q/K scoring achieves the best RefSpatial accuracy (51.00%) and lowest length MAE (0.3341 m), with no consistent winner across all tasks, favoring the minimal-change design.

**Numerical discretization (Table 4)**: Coarser 4/8-bin codebook (216 cells) reduces prediction MAE from 0.3277 to 0.3202 m despite higher quantization error (0.1540 vs. 0.0391 m). Argmax decoding (0.3202 m) outperforms conditional expectation (0.3300 m).

**Visual refinement (Table 5)**: Crop re-encoding achieves the highest RefSpatial (51.00%) and Point-Bench (56.84%) scores. Removing crops from the trained model collapses RefSpatial accuracy to 9.00%; retraining without crops recovers only 46.50%.

**Multi-point prediction**: The probability tree raises counting success from 19.39% to 38.78%, with root pruning reducing decision nodes from 91 to 51.10 and inference time by 1.79×.

### Training Efficiency (Table 6)

| Layout | QA/s | Padding (%) | Step CV (%) | Tail (ms) |
|--------|------|-------------|-------------|-----------|
| Original order | 99.06 | 35.17 | 2.17 | 39.5 |
| Local length | 116.91 | 12.36 | 4.13 | 157.6 |
| **Global balanced** | **129.54** | **1.77** | 6.07 | 85.5 |

Global grouping with cross-GPU balancing achieves 30.78% higher throughput than the original strategy.

## Theoretical and Practical Implications

**Theoretical contributions**:
- Demonstrates that spatial understanding can be formulated as classification over finite candidate sets, challenging the assumption that free-form text generation is necessary for spatial reasoning tasks
- Shows that pretrained attention projections (Q/K) can serve as effective classifiers without architectural modifications, preserving pretrained knowledge
- Establishes that coarse-to-fine hierarchical classification can match or exceed finer-grained single-stage approaches, with coarser discretization sometimes improving prediction accuracy

**Practical implications**:
- System 1 spatial models enable real-time spatial judgments at interaction rates, critical for robotics, autonomous navigation, and embodied AI applications
- The 5.6× inference speedup over autoregressive baselines makes deployment feasible in latency-sensitive contexts
- The unified framework simplifies model design: one mechanism handles choice, numerical, and point prediction tasks
- The multi-point extension via tree expansion demonstrates that additional capabilities can be added without retraining
- The efficient batching strategy (30.8% throughput improvement) reduces training costs for heterogeneous multi-task data

## Conclusion

MJev-Spatial validates a System 1 approach to spatial understanding: using pretrained multimodal knowledge for direct classification decisions without generating answer text. Key achievements:

- **Unified framework**: Options, numerical values, and image locations share a candidate-classification mechanism with hierarchical refinement for precision
- **Superior performance**: Highest overall score (65.64%) among eight models, leading seven of twelve benchmarks
- **Efficiency**: 2.5× and 5.6× speedups over the fastest autoregressive baseline and base model respectively
- **Scalable training**: 12.3M QA corpus with joint stage supervision and efficient batching (30.8% throughput gain)

**Future directions** suggested by the work:
- Extending the classification framework to additional spatial and non-spatial task types
- Exploring the probability-tree approach for more complex multi-point and structured predictions
- Investigating whether the native Q/K scoring advantage generalizes to other backbone architectures
- Applying the efficient batching strategy to other heterogeneous multi-task training scenarios

The results support direct candidate decisions as a viable and efficient paradigm for spatial tasks, opening new possibilities for real-time spatial intelligence in embodied systems.

---

_Markdown view of https://picx.dev/p/k6e35W, served by PicX — AI-generated visual whiteboard summaries of research papers._