Research map / Vision-language and multimodal / Multimodal LLMs

InfiniBench: Infinite Benchmarking for Visual Spatial Reasoning with Customizable Scene Complexity

CVPR 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.

Presents InfiniBench, an automated benchmark generator that turns natural-language scene descriptions into customizable, physically plausible 3D scenes and camera videos for diagnosing VLM spatial reasoning.

Problem

Evaluating VLM spatial reasoning requires benchmarks that are diverse, scalable and customizable along compositional, relational and observational scene complexity, which existing benchmarks and 3D scene generators do not provide.

Approach

An LLM agent (Gemini-2.5-Pro) iteratively refines procedural scene constraints using error reports from a cluster-based layout optimizer (movable clusters of parent and child objects), and a frontier-inspired, task-aware camera trajectory optimizer renders the scene into videos with full object coverage.

Key results

• In the high-object setting InfiniBench reaches fidelity 0.98 and near-zero collisions/out-of-boundary objects (#CN 0.0, #OB 0.1), while LayoutGPT has 13.5 collisions and Infinigen fidelity drops to 0.64.
• Ablation: full InfiniBench raises fidelity from 0.64 (base Infinigen) to 0.92, versus 0.71 for constraint refinement alone and 0.68 for cluster optimization alone.
• Gemini-2.5-Pro achieves the highest accuracy on most generated tasks (e.g., 87.9 on low-distractor spatiotemporal), and VLM performance declines as object count rises from 5 to 50.

Surprising findings

• Bird's-eye-view camera views do not uniformly help: GPT5 measurement accuracy is nearly unchanged between views, and some models perform worse on spatiotemporal tasks from ego views only by modest margins.
• Models tend to overestimate object counts in dense scenes, which the authors attribute to repetitive counting of the same object across video frames.

Limitations

• Scene generation is built on Infinigen's asset library and depends on its procedural APIs.
• Iterative LLM refinement requires multiple rounds of LLM calls and optimization runs, adding cost.
• Camera optimization is restricted to a 4-DOF horizontal-plane search with fixed height and zero roll.

Experimental setup

Datasets: Infinigen asset library

Benchmarks: InfiniBench-generated measurement, perspective-taking and spatiotemporal benchmarks

Metrics: prompt fidelity CLIP score layout realism #OB (out-of-boundary objects) #CN (collisions) mean relative accuracy exact match accuracy

Baselines: I-Design Holodeck LayoutGPT Luminous Infinigen

Builds on: Infinigen ProcTHOR frontier-based exploration Chain-of-Thought prompting Holodeck

Closest papers by meaning

• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.90)
• MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.90)
• From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)
• ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.90)
• MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.89)