SpatialWorld: A New Benchmark for Interactive Spatial Reasoning Capabilities of Multimodal Agents

SpatialWorld is a new interactive spatial reasoning benchmark for multimodal agents, integrating 8 heterogeneous simulation backends (covering home environments, travel scenarios, social collaboration, etc.) and containing 760 manually annotated tasks. Evaluation results show that even the current strongest closed-source model GPT-5 has an average task success rate of only 17.4%, revealing significant bottlenecks in the active exploration and long-term planning capabilities of multimodal agents. This benchmark comes from a paper published on arXiv on June 8, 2026 (link: http://arxiv.org/abs/2606.09669v1).

Spatial reasoning is a fundamental ability for Multimodal Large Language Models (MLLMs) to perceive and interact with the physical world, but existing evaluation benchmarks have obvious flaws: they mainly rely on passive evaluations (such as static visual question answering) or pipelines tied to specific simulators, which cannot truly assess an agent's spatial understanding ability in dynamic, interactive environments. This approach is like letting someone learn to drive by looking at photos without ever actually operating a car, making it difficult to reflect real spatial reasoning levels.

SpatialWorld adopts a unified evaluation framework for multimodal agents, with a design philosophy that aligns with the complexity of the real world:

1. Multi-simulator Backend Integration: Covers 8 heterogeneous scenarios (home, travel, social collaboration, etc.) and is uniformly scheduled via a simulator-agnostic protocol;
2. Key Task Features:
   • Partial Observability: Agents can only access limited visual information and need to actively explore the environment;
   • Unified Text Action Interface: Aligns with the native capabilities of MLLMs, no need to learn low-level control commands;
   • Reliable Evaluation Mechanism: Manual verification of initial states, reference trajectories, and terminal state validators to ensure repeatable and trustworthy results.

The research team evaluated 15 advanced agent models, and the results show:

• The closed-source model GPT-5 has an average Task Success Rate (TSR) of only 17.4%;
• The open-source model Qwen-3.5 has a success rate of 14.1%; Further analysis found that there is a disconnect between task success rate and execution efficiency (some models succeed but take many steps, others are efficient but have low success rates); performance varies significantly across different domains (home/travel/social), indicating insufficient cross-domain generalization capabilities.

SpatialWorld provides a rigorous testing platform for spatial intelligence research, revealing the core bottlenecks that current multimodal agents need to break through: active exploration and long-term planning capabilities. Future research can focus on:

• Improving the efficiency and success rate of active exploration and long-term planning;
• Enhancing cross-domain generalization capabilities; With the advancement of simulation technology and model capabilities, we expect agents' performance on this benchmark to gradually improve, ultimately achieving reliable interaction with the real physical world.