Unexplored flaws in multiple-choice VQA evaluations

2025-11-27

AI Summary

This paper identifies a previously overlooked “semantic-neutral prompt formatting bias” in multiple-choice visual question answering (VQA) benchmarks: minor syntactic variations—such as punctuation, line breaks, or option layout—that preserve semantic equivalence nonetheless significantly degrade multimodal large language models’ (MLLMs) answer accuracy, independently of answer ordering and model confidence. Method: We systematically categorize and model three classes of prompt variation factors; conduct large-scale experiments across seven state-of-the-art MLLMs and five VQA datasets, evaluating 48 distinct prompt variants via controlled ablation studies and confidence-consistency analysis. Contribution/Results: We demonstrate that existing debiasing methods are entirely ineffective against this bias. Our findings expose a fundamental methodological flaw in current multimodal evaluation frameworks and underscore the urgent need to redesign prompt robustness assessment paradigms for reliable MLLM evaluation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Question Answering

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Multimodal Large Language Models (MLLMs) demonstrate strong capabilities in handling image-text inputs. A common way to assess this ability is through multiple-choice Visual Question Answering (VQA). Earlier works have already revealed that these benchmarks are sensitive to answer choice order, a limitation that can be mitigated through careful design. Yet, we highlight additional, unexplored biases in prompt formatting that question the reliability of current MLLM evaluations. Specifically, we identify three key variation factors in prompt formatting and analyze their impact through a large-scale study involving $mathbf{ ext{seven}}$ MLLMs and $mathbf{ ext{five}}$ VQA datasets, spanning $mathbf{48}$ distinct $mathbf{ ext{prompt format variations}}$. Our findings reveal that multiple-choice VQA is highly sensitive to minor prompt format changes, even when these changes are semantically neutral. We further demonstrate that these biases persist independently of known order biases or the MLLM's confidence in the correct answer. Finally, we demonstrate that existing bias mitigation strategies fail to address these newly identified biases.

Problem

• Identifies unexplored prompt formatting biases in MLLM evaluations
• Analyzes sensitivity of multiple-choice VQA to minor format changes
• Shows existing mitigation strategies fail to address these new biases

Innovation

• Identified unexplored prompt formatting biases in VQA
• Analyzed impact through large-scale multi-model dataset study
• Demonstrated existing mitigation strategies fail for new biases