Geometry as Position
for Spatial Reasoning
in Vision-Language Models

Han Su^1,2Meng Tian²Tianyu Huang³Mingyu Gao¹Lennox Wei¹
Dongxu Li¹Jianhua Han²Hang Xu²Xiaohe Wu¹Wangmeng Zuo¹

¹ Harbin Institute of Technology² Yinwang Intelligent Technology³ Huawei Technologies

Abstract

Vision-language models still struggle to understand 3D spatial relations across views and update spatial judgments as the observer moves. A common approach fuses 3D reconstruction features into visual tokens, but aligning these features with pretrained semantic representations remains challenging under limited spatial supervision. We propose DK-VLM with Dual-Kernel Attention, which computes attention from both a visual token's input position and its geometric position in the scene. A semantic kernel retains the pretrained model's native attention and positional encoding, while a separate geometric kernel uses estimated 3D points and camera poses to directly modulate attention through point displacements and relative view transformations. Then, a learned gate fuses their outputs. Geometry thus enters attention as position rather than as fused reconstruction features. We further introduce Reference-Frame Composition (RC) supervision to train the model to update its reference frame across successive judgments. It provides stepwise turn and heading targets derived from verified single-step annotations. Among the compared methods, DK-VLM-8B achieves the highest average score across six spatial benchmarks and the highest overall scores on VSI-Bench (74.33) and ReVSI-32 (58.28). RC supervision further improves multi-step spatial reasoning, while the full model preserves general video understanding.

One token. Two positions.

DK-VLM overview: 2D image locations do not fully capture 3D relations. A semantic kernel and a geometric kernel model two kinds of position. — Keep the native positional computation. Let scene geometry shape attention in parallel.

How DK-VLM Works

Dual-Kernel Attention models geometric relations. Reference-Frame Composition supervision teaches the model to use them across successive judgments.

Dual-Kernel Attention

Architecture

Supervision

Reference-Frame
Composition

Every move changes the reference frame for the next judgment. RC connects verified single-step annotations into chains and provides a target at each step.

1.6MRC samples constructed

150Kselected for training

Training uses two- and three-step chains; four-step chains are held out for evaluation.

RC illustration: a robot starts at a chair facing a window and follows a route to a table, bookshelf and door with stepwise right, left and straight targets. — Current position & headingNext stop

Experimental Results

Evaluated across seven spatial benchmarks

Experiments with Qwen2.5-VL-7B and Qwen3-VL-8B evaluate spatial understanding, stepwise composition, and general video understanding.

VSI-Bench74.33Overall · DK-VLM-8B

ReVSI-3258.28Overall · DK-VLM-8B

Six-benchmark average64.09Highest among compared methods

Spatial Reasoning Demos

VSI-BENCH01 / 04

Playback

BibTeX

@misc{su2026dkvlmgeometrypositionspatial,
  title         = {DK-VLM: Geometry as Position for Spatial Reasoning in Vision-Language Models},
  author        = {Han Su and Meng Tian and Tianyu Huang and Mingyu Gao and Lennox Wei and Dongxu Li and Jianhua Han and Hang Xu and Xiaohe Wu and Wangmeng Zuo},
  year          = {2026},
  eprint        = {2610.22356},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.22356},
}

Geometry as Positionfor Spatial Reasoning in Vision-Language Models