Research map / Vision-language and multimodal / Multimodal LLMs

CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CausalSpatial, a synthetic Blender-rendered benchmark for object-centric causal spatial reasoning (collision, compatibility, occlusion, trajectory), and a video-simulation augmentation (CausalSpatial-style world model via ATI) that gives small gains to Qwen3-VL-30B-A3B.

Problem

Current MLLMs struggle to answer "what-if" questions about how moving an object in a static 3D scene will affect other objects, relying on textual chain-of-thought that drifts from visual evidence.

Approach

A multiple-choice benchmark with four object-centric causal tasks, a "Not Sure" option with NSR metric, physics-simulated ground truth, and a pipeline that generates trajectory-controlled videos via ATI to supply visual simulation cues to an MLLM at inference.

Key results

• GPT-5 achieves 54.17% average accuracy versus 84.49% for humans on CausalSpatial.
• Qwen3-VL-30B-A3B-Thinking reaches 50.50% on Level 2, the best open-source result; Qwen3-VL-2B (40.60%) outperforms Qwen2.5-VL-7B (39.61%) and LLaVA-OneVision (34.62%) on average.
• Adding three frames of generated video to Qwen3-VL-30B-A3B-Instruct improves Collision by 2.40% and Trajectory by 2.20%, with a slight drop on Occlusion.

Surprising findings

• Scaling Qwen3-VL from 4B to 30B-A3B yields little accuracy gain (44.76% to 45.80%), and spatially tuned models (Spatial-VLM, Spatial-MLLM) underperform general models.
• Thinking mode regresses performance for Qwen3-VL-8B (43.53% to 42.65%) while lowering NSR, indicating decisiveness without accuracy gains.

Limitations

• Video-augmentation gains are small (about 2%) and tested on only one model; Occlusion declines.
• Benchmark is synthetic Blender scenes with limited object categories (over 50), and human evaluation uses only three evaluators.
• Generated video does not guarantee consistency for volume, shape deformation, or inter-object collision dynamics.

Experimental setup

Datasets: CausalSpatial (synthetic Blender scenes)

Benchmarks: CausalSpatial

Metrics: accuracy (TPR) Not Sure Rate (NSR)

Baselines: Spatial-VLM Spatial-MLLM LLaVA-OneVision-1.5-8B Qwen2.5-VL-7B Qwen3-VL-8B Gemini 2.5 Pro Gemini 2.5 Flash Sonnet 3.7

Builds on: ATI chain-of-thought prompting

Closest papers by meaning

• Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.90)
• Thinking in Dynamics: How Multimodal Large Language Models Perceive, Track, and Reason Dynamics in Physical 4D World (CVPR 2026 · similarity 0.90)
• OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding (NeurIPS 2025 · similarity 0.89)
• SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs (ICLR 2026 · similarity 0.89)
• Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests (NeurIPS 2026 · similarity 0.89)
• Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models (ICLR 2026 · similarity 0.89)