[Paper Review] Large Language Models: Testing Their Capabilities to Understand and Explain Spatial Concepts (Short Paper)

Chaoyou Fu, Peixian Chen|arXiv (Cornell University)|Jun 23, 2023

Topic Modeling88 citations

TL;DR

This paper introduces MME, the first comprehensive benchmark for evaluating Multimodal Large Language Models (MLLMs) across 14 perception and cognition tasks. Designed with manually curated, concise instructions to ensure fair comparison and avoid data leakage, MME reveals that even top models like GPT-4V and Lion still have significant room for improvement, particularly in object position understanding and reasoning under strict instruction following.

ABSTRACT

Interest in applying Large Language Models (LLMs), which use natural language processing (NLP) to provide human-like responses to text-based questions, to geospatial tasks has grown rapidly. Research shows that LLMs can help generate software code and answer some types of geographic questions to varying degrees even without fine-tuning. However, further research is required to explore the types of spatial questions they answer correctly, their abilities to apply spatial reasoning, and the variability between models. In this paper we examine the ability of four LLM models (GPT3.5 and 4, LLAma2.0, Falcon40B) to answer spatial questions that range from basic calculations to more advanced geographic concepts. The intent of this comparison is twofold. First, we demonstrate an extensible method for evaluating LLM’s limitations to supporting spatial data science through correct calculations and code generation. Relatedly, we also consider how these models can aid geospatial learning by providing text-based explanations of spatial concepts and operations. Our research shows common strengths in more basic types of questions, and mixed results for questions relating to more advanced spatial concepts. These results provide insights that may be used to inform strategies for testing and fine-tuning these models to increase their understanding of key spatial concepts.

Motivation & Objective

• To address the lack of a comprehensive, fair, and quantitatively analyzable benchmark for evaluating Multimodal Large Language Models (MLLMs).
• To overcome limitations in existing evaluation methods, including data leakage from public datasets, inconsistent instruction design, and reliance on open-ended or subjective scoring.
• To evaluate both perception and cognition abilities of MLLMs in a unified, standardized, and reproducible manner.
• To identify persistent model weaknesses such as instruction non-compliance, perception errors, reasoning failures, and object hallucination.

Proposed method

• Designing a new benchmark, MME, with 14 subtasks covering perception (e.g., existence, count, position, color) and cognition (e.g., commonsense reasoning, numerical calculation, code reasoning).
• Creating all instruction-answer pairs manually to prevent data leakage from training data, ensuring evaluation integrity.
• Using a standardized, concise instruction format ('Please answer yes or no') to enable fair comparison across models and reduce prompt engineering bias.
• Implementing a binary classification format (yes/no) for all questions to allow for direct, quantitative statistical analysis of model performance.
• Evaluating 30 advanced MLLMs on the full benchmark to assess capabilities and identify failure patterns.
• Analyzing model behaviors through qualitative case studies to identify recurring problems like hallucination, misalignment, and reasoning breakdowns.

Experimental results

Research questions

• RQ1 How well do current MLLMs perform across a broad spectrum of perception and cognition tasks in a standardized, fair evaluation setting?
• RQ2 To what extent do MLLMs follow simple, concise instructions, and how does this affect their performance?
• RQ3 What are the most common failure modes in MLLMs when evaluating perception and reasoning under controlled, manual annotations?
• RQ4 How do different MLLMs compare in tasks requiring fine-grained visual understanding, such as landmark or artwork recognition?
• RQ5 Can a unified, manually curated benchmark reveal consistent, quantifiable performance gaps and guide future model optimization?

Key findings

• GPT-4V achieves the highest overall score (1621.66/2000) in perception and ranks first in cognition, but refuses to answer questions about individuals, resulting in a zero score in celebrity recognition.
• WeMM, InfMLLM, and SPHINX rank in the top three for perception, while GPT-4V, Lion, and WeMM lead in cognition, with GPT-4V excelling in code reasoning (170/200) and commonsense reasoning (142.14/200).
• Object position recognition is the weakest subtask, with models showing poor sensitivity to spatial layout, indicating a key limitation in spatial understanding.
• A significant number of models fail to follow simple 'yes or no' instructions, often generating free-form responses, which undermines reliability and evaluation fairness.
• Object hallucination is a prevalent issue: models frequently answer 'yes' to questions about non-existent objects, leading to ~50% accuracy and near-zero accuracy+ in such cases.
• Perception performance is highly sensitive to minor instruction variations—even a single-word change can cause contradictory responses, highlighting instability in visual grounding.