Organizations: State Key Lab of CAD & CG, Zhejiang University, China · Zhejiang University of Technology · Hangzhou Dianzi University, China · Xiaomi Corporation, China
Metric reasoning is a critical and challenging task for Vision Language Models (VLMs), playing a pivotal role in embodied AI tasks such as robotic manipulation and autonomous navigation. However, current spatial reasoning remains bottlenecked by rigid pixel-level supervision; such localized optimization often compromises general multimodal intelligence, triggering performance degradation or catastrophic forgetting of broad reasoning capabilities. To address these limitations, we introduce Metric-Bench, a focused benchmark designed to guide metric-spatial reasoning using contextual information. By incorporating in-image reference objects with known physical dimensions, Metric-Bench guides models to implicitly learn the 2D-to-3D mapping without camera intrinsics. We further present MetricReasoner, a task-adapted reinforcement fine-tuning recipe for reference-grounded metric reasoning, using structured prompts and verifiable numerical rewards. Extensive experiments on Metric-Bench demonstrate that our approach significantly enhances spatial metric understanding, outperforming existing and even larger proprietary models by 43.1%, while improving downstream embodied performance over a spatial-specialized counterpart by 30.4% on RoboSpatial overall accuracy and 9.3% on ERQA, and additionally delivering consistent gains on general benchmarks (15.9% on V⋆Bench, 88.9% on BLINK), indicating that the proposed adaptation does not necessarily compromise general VLM capabilities.
Figures & tables
Figure 1: We propose Metric-Bench, a benchmark inspired by human perception of 3D space: given the true physical measurements of a few reference objects, the model learns to reason about relative positions and perspective relationships in 3D space, thereby inferring precise metric measurements of other objects in the scene. Furthermore, building upon this benchmark, we employ a spatial-guided reinforcement fine-tuning strategy, without compromising the model’s generalizable reasoning capability, and achieves state-of-the-art performance in metric scale perception, surpassing existing multimodal large language models.
Figure 2: Data construction pipeline and statistics in Metric-Bench.
MRA ↑
RMSE ↓
δ1↑
Avg.
Size
Pos.
Dep.
Dis.
Box
Human
76.83
81.51
77.76
76.33
79.35
82.09
0.2671
0.7920
Proprietary Models (API)-Mini
Gemini-2.5-flash
33.95
34.74
31.96
61.64
18.87
29.09
0.7886
0.3254
GPT-5-mini
37.84
38.47
33.06
58.88
22.73
33.40
0.7842
0.4014
Open-source Models
Table 1: The metric estimation performance of public and our fine-tuned models on Metric-Bench. All public models are evaluated under zero-shot settings. Proprietary models and human performance are evaluated using a mini dataset, which contains 400 queries randomly sampled from the bench. Pos., Dep., Dis. are abbreviations for position, depth and distance. Length, width, height, elevation and dimension are all included in Size. The unit of RMSE is meters.
Figure 3: Performance of MetricReasoner across all categories in Metric-Bench.
Table 5
Models
ERQA ↑
RoboSpatial ↑
CountBench ↑
V ⋆ Bench ↑
BLINK ↑
Qwen3-VL-8B
42.25
53.14
0.917
61.78%
58.61%
VST-7B-RL
40.25
42.29
0.891
59.16%
43.32%
MetricReasoner
44.00
55.14
0.920
68.59%
81.82%
Table 4: General performance of spatial-specialized model VST-7B-RL, our MetricReasoner and corresponding base model Qwen3-VL-8B on ERQA, RoboSpatial, and general benchmarks, including CountBench, V ⋆ Bench and BLINK.
Strategy
RoboSpatial
CountBench
V ⋆ Bench
BLINK
RL(MetricReasoner)
55.14
0.9204
68.59%
81.82%
RL w/ COT&SFT
45.14
0.9190
62.83%
34.53%
SFT
40.86
0.9093
63.35%
51.26%
Table 5: Results on RoboSpatial, CountBench, V ⋆ Bench, and BLINK of different fine-tuning strategies. RL w/ COT&SFT denotes SFT on CoT-annotated data followed by reinforcement fine-tuning.
Figure 4: Qualitative results of MetricReasoner.
Figure 5: Results of open-source VLMs(InternVL3.5-38B and Qwen3-VL-32B) on Metric-Bench.
Figure 6: Visualization of MetricReasoner on RoboTwin2.0. Points on images are the projection of the predicted results.
Visual Language Models (VLMs) have increasingly become the main paradigm for understanding indoor scenes, but they still struggle with metric and spatial reasoning. Current approaches rely on end-to-end video understanding or large-scale spatial question answering fine-tuning, inherently coupling perception and reasoning. In this paper, we investigate whether decoupling perception and reasoning leads to improved spatial reasoning. We propose an agentic framework for static 3D indoor scene reasoning that grounds an LLM in an explicit 3D scene graph (3DSG). Rather than ingesting videos directly, each scene is represented as a persistent 3DSG constructed by a dedicated perception module. To isolate reasoning performance, we instantiate the 3DSG from ground-truth annotations. The agent interacts with the scene exclusively through structured geometric tools that expose fundamental properties such as object dimensions, distances, poses, and spatial relationships. The results we obtain on the static split of VSI-Bench provide an upper bound under ideal perceptual conditions on the spatial reasoning performance, and we find that it is significantly higher than previous works, by up to 16%, without task specific fine-tuning. Compared to base VLMs, our agentic variant achieves significantly better performance, with average improvements between 33% to 50%. These findings indicate that explicit geometric grounding substantially improves spatial reasoning performance, and suggest that structured representations offer a compelling alternative to purely end-to-end visual reasoning.
Spatial reasoning is foundational for Vision-Language Models (VLMs), particularly when deployed as Vision-Language-Action (VLA) agents in physical environments. However, existing benchmarks predominantly focus on elementary, single-hop relations, neglecting the multi-hop compositional reasoning and precise visual grounding essential for real-world scenarios. To address this, we introduce MultihopSpatial, offering three key contributions: (1) A comprehensive benchmark designed for multi-hop and compositional spatial reasoning, featuring 1- to 3-hop complex queries across diverse spatial perspectives. (2) Acc@50IoU, a complementary metric that simultaneously evaluates reasoning and visual grounding by requiring both answer selection and precise bounding box prediction - capabilities vital for robust VLA deployment. (3) MultihopSpatial-Train, a dedicated large-scale training corpus to foster spatial intelligence. Extensive evaluation of 37 state-of-the-art VLMs yields eight key insights, revealing that compositional spatial reasoning remains a formidable challenge. Finally, we demonstrate that reinforcement learning post-training on our corpus enhances both intrinsic VLM spatial reasoning and downstream embodied manipulation performance.
Youngwan Lee, Soojin Jang, Yoorhim Cho +3
Electronics and Telecommunications Research Institute, South Korea · Korea Advanced Institute of Science and Technology, South Korea · Sungkyunkwan University, South Korea +1
Vision-Language Models (VLMs) have been increasingly applied in real-world scenarios due to their outstanding understanding and reasoning capabilities. Although VLMs have already demonstrated impressive capabilities in common visual question answering and logical reasoning, they still lack the ability to make reasonable decisions in complex real-world environments. We define this ability as spatial logical reasoning, which not only requires understanding the spatial relationships among objects in complex scenes, but also the logical dependencies between steps in multi-step tasks. To bridge this gap, we introduce Spatial Logical Question Answering (SpatiaLQA), a benchmark designed to evaluate the spatial logical reasoning capabilities of VLMs. SpatiaLQA consists of 9,605 question answer pairs derived from 241 real-world indoor scenes. We conduct extensive experiments on 41 mainstream VLMs, and the results show that even the most advanced models still struggle with spatial logical reasoning. To address this issue, we propose a method called recursive scene graph assisted reasoning, which leverages visual foundation models to progressively decompose complex scenes into task-relevant scene graphs, thereby enhancing the spatial logical reasoning ability of VLMs, outperforming all previous methods. Code and dataset are available at https://github.com/xieyc99/SpatiaLQA.
Yuechen Xie, Xiaoyan Zhang, Yicheng Shan +6
Zhejiang University · The University of Sydney · ManyCore +2