CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning
TLDR
Introduces CausalSpatial, a benchmark for object-centric causal spatial reasoning, showing MLLMs underperform humans, and proposes COW, a video-generation framework to ground reasoning.
Reasoning
The paper addresses an important gap and provides a diagnostic benchmark with clear human-model comparisons. Its main weakness is that the abstract does not specify whether the benchmark uses real or synthetic scenes, and the COW framework's effectiveness is not detailed.
Read-first score
Read-first score 62.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.
Field roles
Rank sensitivity
Stability: volatile; rank range: 310.
Keyword Scores
Deep Analysis
Innovations
- Introduces CausalSpatial, a diagnostic benchmark for object-centric causal spatial reasoning with four tasks: Collision, Compatibility, Occlusion, and Trajectory.
- Proposes the Causal Object World model (COW), a framework that externalizes simulation by generating videos of hypothetical dynamics to provide explicit visual cues of causality.
- Identifies a core MLLM failure mode: over-reliance on textual chain-of-thought reasoning that drifts from visual evidence and produces spatially ungrounded hallucinations.
Methodology
The paper introduces a benchmark evaluating whether models can anticipate consequences of object motions in 3D scenes across Collision, Compatibility, Occlusion, and Trajectory tasks. It compares human performance to multimodal large language models such as GPT-5. The proposed COW framework addresses observed failures by generating videos of hypothetical dynamics to ground reasoning in visual physical reality rather than linguistic priors.
Key Results
Humans score 84% on CausalSpatial while GPT-5 achieves only 54%, revealing a severe gap. Analysis indicates MLLMs over-rely on textual chain-of-thought reasoning and produce fluent but spatially ungrounded hallucinations.