SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks

Researchers introduce SpatialWorld, a comprehensive benchmark for evaluating multimodal AI agents' ability to understand and navigate physical spaces in real-world tasks. Testing 15 advanced models reveals significant limitations: GPT-5 achieves only 17.4% task success while open-source alternatives lag further, exposing critical gaps in spatial reasoning and long-horizon planning capabilities.

SpatialWorld addresses a fundamental gap in AI evaluation methodology by moving beyond static, passive benchmarks toward dynamic, interactive spatial reasoning tasks. The benchmark integrates eight different simulation backends under a unified protocol, enabling standardized assessment of how well multimodal agents perceive environments, gather visual information actively, and execute complex instructions—capabilities essential for embodied AI systems deployed in real-world scenarios.

The surprisingly low performance metrics across state-of-the-art models signal that current multimodal large language models lack robust spatial understanding despite their general capabilities. Even GPT-5's 17.4% success rate indicates that spatial reasoning remains a critical bottleneck in AI development. The performance gap between execution efficiency and task success suggests models struggle with planning horizons and resource optimization, not merely perception.

For the AI industry, these findings underscore the substantial work required before deploying agents in practical robotic or autonomous applications. Developers cannot rely on general-purpose LLMs for spatial tasks without specialized training or architectural modifications. The research validates the need for domain-specific benchmarks that expose model weaknesses under real operational constraints, guiding investment in core capability improvements rather than incremental gains.

Looking forward, the field must address three key challenges: improving active exploration strategies, extending planning horizons, and generalizing spatial understanding across diverse environments. The benchmark itself will likely become a standard testing ground, influencing how companies prioritize AI development roadmaps and where resources flow within the broader AI research ecosystem.

• →GPT-5 achieves only 17.4% success on interactive spatial tasks, revealing fundamental limitations in current multimodal agents.
• →SpatialWorld's unified protocol across eight simulators enables standardized, rigorous evaluation of spatial reasoning in realistic scenarios.
• →Performance-efficiency mismatch indicates models struggle with long-horizon planning and active exploration, not just perception.
• →Open-source models lag commercial alternatives significantly, creating a competitive moat for well-resourced AI companies.
• →Benchmark findings will likely shape AI development priorities, particularly for robotics and autonomous systems deployment.