PaperlayerView exact revision on arXiv

Source-linked AI summary

SpaceR: Reinforcing MLLMs in Video Spatial Reasoning

Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, Xu Sun

arXiv:2504.01805v2cs.CV

TL;DR

MLLMs struggle with video spatial reasoning because suitable datasets and effective reasoning-oriented post-training remain limited. SpaceR introduces a mixed dataset and SG-RLVR with map imagination, achieving strong spatial benchmark results while maintaining video understanding performance, including 45.6% on VSI-Bench and an 11.6% advantage over GPT-4o.

  • Problem

    MLLMs struggle with video spatial reasoning because high-quality task-specific datasets and post-training strategies sufficient for deep reasoning are limited.

  • Method

    SpaceR combines SpaceR-151k with SG-RLVR, which extends GRPO using task-specific verifiable rewards and map imagination for spatial-layout inference.

  • Results

    45.6% accuracy on VSI-Bench surpasses GPT-4o by 11.6% accuracy, alongside consistent gains on spatial benchmarks and promising video-understanding results.

  • Takeaways & Limitations

    SpaceR provides a dataset and reinforcement-learning framework that support spatial reasoning while preserving general video understanding performance.

Abstract

from arXiv · show

Video spatial reasoning, which involves inferring the underlying spatial structure from observed video frames, poses a significant challenge for existing Multimodal Large Language Models (MLLMs). This limitation stems primarily from 1) the absence of high-quality datasets for this task, and 2) the lack of effective training strategies to develop spatial reasoning capabilities. Motivated by the success of Reinforcement Learning with Verifiable Reward (RLVR) in unlocking LLM reasoning abilities, this work aims to improve MLLMs in video spatial reasoning through the RLVR paradigm. To this end, we introduce the $\textbf{SpaceR}$ framework. First, we present $\textbf{SpaceR-151k}$, a dataset with 91k questions spanning diverse spatial reasoning scenarios with verifiable answers, and 60k samples for maintaining general multimodal understanding. Second, we propose $\textbf{Spatially-Guided RLVR (SG-RLVR)}$, a novel reinforcement learning approach that extends Group Relative Policy Optimization (GRPO) with a novel map imagination mechanism, which encourages the model to infer spatial layouts in the thinking process, thereby facilitating more effective spatial reasoning. Extensive experiments demonstrate that SpaceR achieves state-of-the-art performance on spatial reasoning benchmarks (e.g., VSI-Bench, STI-Bench, and SPAR-Bench), while maintaining competitive results on video understanding benchmarks (e.g., Video-MME, TempCompass, and LongVideoBench). Remarkably, SpaceR surpasses the advanced GPT-4o by 11.6\% accuracy on VSI-Bench and is on par with the leading proprietary model Gemini-2.0-Flash, highlighting the effectiveness of our SpaceR-151k dataset and SG-RLVR in reinforcing spatial reasoning ability of MLLMs. Code, model, and dataset are available at https://github.com/OuyangKun10/SpaceR.

1 Introduction

Video spatial reasoning remains difficult for MLLMs because high-quality task-specific data and post-training strategies for deep reasoning are limited. SpaceR addresses these gaps with a verifiable dataset and spatially guided RLVR, achieving strong benchmark gains while retaining video understanding performance.

  • Video spatial reasoning requires reconstructing 3D spatial layouts from frames, beyond recalling or summarizing video content.
  • Existing MLLMs struggle because spatial-reasoning datasets are scarce and supervised fine-tuning is insufficient for deep reasoning.
  • SpaceR-151k combines 91k spatial reasoning QA pairs with 60k general multimodal understanding samples.
  • SG-RLVR extends GRPO with task-specific verifiable rewards and map imagination that encourages explicit spatial-layout generation.
  • 45.6% accuracy on VSI-Bench exceeds GPT-4o by 11.6% accuracy, while SpaceR maintains promising results on video understanding benchmarks.

2 Related Works

Video spatial reasoning differs from conventional video understanding because it requires inferring and reconstructing scene structure, while RL-based methods have recently strengthened language and multimodal reasoning.

  • Conventional video tasks primarily recall or summarize content through captioning, temporal grounding, and temporal perception.
  • Video spatial reasoning requires models to perceive visual content and reconstruct the spatial structure of entire scenes.
  • Spatial reasoning is described as important for developing world models and embodied agents, yet MLLMs continue to show shortcomings on related benchmarks.
  • RLVR and GRPO have shown potential for improving reasoning capabilities in LLMs and MLLMs across text and multimodal tasks.

3 Dataset Construction

SpaceR-151k combines ScanNet-derived spatial reasoning data with general video-understanding samples. Its construction parses scene metadata, generates verifiable QA pairs across spatial tasks, and filters them for quality and balance.

  • Dataset composition: SpaceR-151k contains 91k SR-91k spatial QA pairs and 60k QA instances resampled from Video-R1-260k.
  • Data collection: ScanNet videos are parsed into object metadata, resampled at 24 FPS, and represented with 10 × 10 maps of room object distributions.
  • QA generation: SR-91k generates multi-choice questions about relative distance, relative direction, and appearance order, plus numerical questions about size, distance, and counting.
  • Spatial task definitions: Relative distance uses minimum Euclidean distances, relative direction uses bounding-box centers, and appearance order uses first frame indices.
  • Numerical task definitions: Object size uses the longest point-cloud dimension, room size uses Alpha Shape estimation, absolute distance uses sampled bounding-box points, and counting uses ScanNet metadata.
  • Data filtering: Filtering limits QA pairs per video, shuffles multi-choice answer positions, and adjusts numerical distributions to improve diversity and reduce bias.
  • Data statistics: 151,310 samples span multi-choice, numerical, OCR, free-form, and regression QA types with verifiable answers.

4 Spatially-Guided Reinforcement Learning with Verifiable Reward

SG-RLVR extends GRPO with task-specific verifiable rewards and map imagination to reinforce spatially informed reasoning in video MLLMs. The framework combines answer correctness, spatial-map quality, and controlled reasoning length while constraining policy updates.

  • Framework: SG-RLVR extends GRPO with verifiable rewards tailored to multiple QA types and a map imagination mechanism for spatial reasoning.The framework supervises multi-choice, numerical, OCR, free-form, and regression outputs using task-specific criteria.
  • Verifiable Reward Functions: Format reward gives credit when reasoning and answers are enclosed in the required <think> and <answer> tags.This reward evaluates whether the model response follows the predefined output structure.
  • Verifiable Reward Functions: Multi-choice reward is binary under exact ground-truth matching, while numerical reward averages relative accuracy across confidence thresholds from 0.5 to 0.95.For general multimodal data, OCR, free-form, and regression rewards are additionally computed from WER, ROUGE scores, and relative numerical distance.
  • Map-Based Spatial Reasoning: Map imagination guides the model to generate an M × M scene map, and map reward quantifies predicted-object accuracy against ground-truth coordinates.The relative accuracy is averaged across objects to derive Rmap.
  • Reward Optimization: The augmented reward adds format, task, map, and length rewards when the task answer is correct, but omits map reward otherwise.Length reward applies only to correct answers within the defined range [lmin, lmax].
  • Reward Optimization: GRPO computes within-group response advantages and uses a clipped objective with KL regularization to limit deviation from the reference policy.The coefficients β and ϵ regulate policy divergence and update magnitude.

5 Experiment

Experiments evaluate SpaceR across spatial reasoning and video understanding benchmarks, compare SG-RLVR with SFT, analyze sampling and component ablations, and present qualitative examples.

  • Experimental Setups: SpaceR is evaluated on three spatial reasoning and three video understanding benchmarks.The benchmarks include VSI-Bench, STI-Bench, SPAR-Bench, Video-MME, TempCompass, and LongVideoBench.
  • Main Results: SpaceR consistently outperforms Qwen2.5-VL-7B-Instruct across spatial reasoning and video understanding benchmarks.The reported gains extend across Video-MME, TempCompass, and LongVideoBench, while SpaceR surpasses GPT-4o on spatial reasoning benchmarks.
  • Main Results: SG-RLVR improves performance across spatial reasoning and video understanding benchmarks, whereas SFT produces localized gains and degradation elsewhere.The comparison frames SG-RLVR as more generalizable than SFT.
  • Main Results: Targeted resampling removes all-correct and all-wrong samples to retain examples with higher learning utility.The strategy uses eight generated responses per sample to identify response-consistency categories.
  • Ablation Study: Ablations associate map imagination with spatial reasoning gains, general data with video understanding gains, and SR-91k with spatial reasoning performance.SpaceR outperforms the corresponding variants w/o-map imagination, w/o-general data, and w/o-SR data.
  • Qualitative Analysis: In a VSI-Bench distance example, SpaceR selects the table using explicit spatial mapping, while Qwen2.5-VL-7B-Instruct selects the stool.The SpaceR response represents object locations in a cognitive map and answers C.

6 Conclusion

SpaceR combines the SpaceR-151k dataset with SG-RLVR, which uses task-specific rewards and map imagination to support structured video spatial reasoning. Evaluations report effectiveness and generalizability, while adaptive control of reasoning depth remains an open limitation.

  • 6 Conclusion: SpaceR-151k contains 91k spatial reasoning QA pairs and 60k general video-understanding samples.
  • 6 Conclusion: SG-RLVR integrates task-specific reward functions and map imagination to guide spatial layout inference and structured reasoning.
  • 6 Conclusion: Evaluations across spatial reasoning and video understanding benchmarks validate SpaceR’s effectiveness and generalizability.
  • 6 Conclusion: The framework lacks adaptive control over reasoning depth, which may affect inference efficiency in practice.

A More Details for Data Construction

SR-91k constructs spatial-reasoning QA pairs from ScanNet-derived object information using task-specific templates, then filters noisy or unreliable examples and analyzes video durations.

  • QA templates: Relative-direction templates ask about left, right, front, and back relations using an observer facing the positive y-axis.Back means turning at least 135 degrees, and directions correspond to Cartesian-plane quadrants.
  • QA generation: Generated QA pairs insert identified objects into predefined question templates shown in Figure 6.
  • Data filtering: The construction removes questions involving noisy objects and numerical questions whose objects are too small to identify.
  • Evaluation preparation: Overlapping ScanNet-derived videos are excluded from SR-91k to ensure fair evaluation on VSI-Bench.
  • Data statistics: The SR-91k video-duration distribution is visualized in Figure 7.

B.1 Benchmarks Description

The evaluation covers spatial reasoning, general video understanding, temporal perception, and long-context multimodal understanding, using both open-source and proprietary MLLM baselines.

  • Spatial reasoning benchmarks: VSI-Bench contains over 5,000 QA pairs across 288 indoor-scene videos and evaluates MLLM visual-spatial intelligence.Its videos span homes, offices, factories, and other real-world indoor environments.
  • Spatial reasoning benchmarks: STI-Bench evaluates spatial understanding across desktop, indoor, and outdoor videos, with SR_sub covering six spatial-reasoning subtasks and over 2,000 QA pairs.
  • Spatial reasoning benchmarks: SPAR-Bench contains over 7,000 QA pairs spanning basic perception to complex spatial reasoning in single-view and multi-view settings.
  • Video understanding benchmarks: Video-MME, TempCompass, and LongVideoBench evaluate general video understanding, temporal dynamics, and long-context multimodal understanding, respectively.Video-MME evaluation excludes subtitles, and LongVideoBench uses its validation set without subtitles.
  • Compared models: The comparison models include GPT-4o, Gemini-family MLLMs, and several open-source MLLMs based on Qwen, SigLIP, InternViT, MoonViT, or related backbones.
  • Implementation: Training uses 8 L20 80 GiB GPUs or 4 A800 80 GiB GPUs, while evaluation uses 4 L20 80 GiB GPUs.

C More Empirical Results and Analyses

Additional analyses examine inference-time thinking, model-size scalability, and data-scale effects across spatial-reasoning and video-understanding benchmarks.

  • Data-scale analysis: The data-scale analysis evaluates SpaceR across progressively larger SpaceR-151k subsets and reports results in Figure 8.
  • Thinking modes: Table 3 compares non-think and think modes for SpaceR-Tiny, Qwen2.5-VL models, and SpaceR variants.Non-think outputs answers directly, whereas think includes an explicit reasoning process; average tokens measure generated thinking tokens.
  • Thinking modes: Models not explicitly trained to reason show significant performance drops across most benchmarks in think mode.The comparison includes Qwen2.5-VL-3B-Instruct, Qwen2.5-VL-7B-Instruct, and SpaceR-Tiny SFT.
  • Limitations: SG-RLVR optimizes answer accuracy without directly supervising reasoning traces, which can make reasoning suboptimal or inconsistent.
  • Limitations: Reasoning may be redundant or misleading for video-understanding questions answerable directly, because incorrect reasoning can introduce spurious information.

C.3 Impact of Data Scale on Model Performance.

SpaceR gains spatial-reasoning performance with limited data and shows a larger improvement when trained on the complete SpaceR-151k dataset, while retaining the released model and dataset setup.

  • Data efficiency: 5k–15k training subsets already produce notable gains on VSI-Bench and SPAR-Bench Single View, indicating strong data efficiency.
  • Scaling behavior: Performance generally plateaus or fluctuates slightly through 30k samples before a substantial jump on the full 151k dataset.
  • Scaling behavior: Full-dataset training pushes VSI-Bench accuracy above 45%.
  • Release: SpaceR uses the Qwen2.5-VL-7B-Instruct architecture, and its dataset and model are released under CC BY-NC 4.0 for academic research.
Loading 2504.01805v2…