Research map / Vision-language and multimodal / Multimodal LLMs

Keep it SymPL: Symbolic Projective Layout for Allocentric Spatial Reasoning in Vision-Language Models

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.

Proposes SymPL, a training-free framework that reformulates allocentric spatial questions into symbolic top/front-view layouts with colored abstract dots and colored partitions, answered by Qwen2.5-VL as a localization question.

Problem

VLMs perform well on egocentric spatial reasoning but degrade sharply when questions must be answered from an object-centered (allocentric) viewpoint.

Approach

Reformulate allocentric questions via 3D estimation and orthogonal projection into abstract colored-symbol layouts with bipartitioned color regions, converting relational questions into color-based localization questions that VLMs handle well, without any training.

Key results

• On COMFORT#, SymPL reaches 69.00% left/right, 97.33% closer, 91.41% visibility and 91.50% facing, outperforming all prior methods.
• On 3DSRBench, SymPL achieves 79.94% left/right and 75.00% visibility, the best among compared methods.
• On COCOSPATIAL egocentric left/right and above/below, SymPL reaches 89.83% and 94.33%, above the best baselines.

Surprising findings

• Most general-purpose VLMs perform at or below random chance on allocentric left/right questions, despite strong egocentric performance.
• Spatial-reasoning-specialized models such as SpatialVLM and SpatialRGPT fine-tuned for egocentric reasoning degrade on allocentric questions.

Limitations

• Pipeline depends on foundation models (GroundingDINO, DepthPro, OrientAnything); errors in facing direction estimation are the most frequent failure source.
• Evaluated on synthetic and a limited set of 3DSRBench/COCOSPATIAL categories rather than full benchmarks.
• Reasoning steps require multiple VLM and foundation-model calls, adding inference cost.

Experimental setup

Datasets: COMFORT# 3DSRBench COCOSPATIAL COMFORT VI COMFORT Multi

Benchmarks: COMFORT# 3DSRBench COCOSPATIAL COMFORT VI COMFORT Multi

Metrics: accuracy success rate

Baselines: GPT-5 Gemini-2.5-Flash Qwen2.5-VL LLaVA-NeXT LLaVA-OneVision Molmo Cambrian-1 Qwen2.5-VL + CoT Qwen2.5-VL + SoM Qwen2.5-VL + SCAFFOLD SpatialVLM SpatialRGPT

Builds on: APC SoM SCAFFOLD chain-of-thought

Closest papers by meaning

• Allocentric Perceiver: Disentangling Allocentric Reasoning from Egocentric Visual Priors via Frame Instantiation (NeurIPS 2026 · similarity 0.92)
• 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (ICML 2026 · similarity 0.90)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
• Pursuing Minimal Sufficiency in Spatial Reasoning (ICLR 2026 · similarity 0.89)
• SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs (ICLR 2026 · similarity 0.89)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.89)