Order Matters: Exploring Order Sensitivity in Multimodal Large Language Models

TL;DR

This work demonstrates that popular MLLMs pay special attention to certain multimodal context positions, particularly the beginning and end, and proposes a new metric, Position-Invariant Accuracy (PIA), to address order bias in MLLM evaluation.

Abstract

Multimodal Large Language Models (MLLMs) utilize multimodal contexts consisting of text, images, or videos to solve various multimodal tasks. However, we find that changing the order of multimodal input can cause the model's performance to fluctuate between advanced performance and random guessing. This phenomenon exists in both single-modality (text-only or image-only) and mixed-modality (image-text-pair) contexts. Furthermore, we demonstrate that popular MLLMs pay special attention to certain multimodal context positions, particularly the beginning and end. Leveraging this special attention, we place key video frames and important image/text content in special positions within the context and submit them to the MLLM for inference. This method results in average performance gains of 14.7% for video-caption matching and 17.8% for visual question answering tasks. Additionally, we propose a new metric, Position-Invariant Accuracy (PIA), to address order bias in MLLM evaluation. Our research findings contribute to a better understanding of Multi-Modal In-Context Learning (MMICL) and provide practical strategies for enhancing MLLM performance without increasing computational costs.

Figures (14)

• Figure 1: Certain multimodal context orders elicit accurate responses, while others lead to erroneous outputs. The green demonstration indicates good order, while the yellow and blue demonstrations indicate bad order.
• Figure 2: The performance of OpenFlamingo Awadalla2023OpenFlamingoAO (3B-base, 3B-instruct, 4B-base, 4B-instruct, 9B) with different multimodal context orders in the image captioning task.
• Figure 3: Spearman correlation coefficients of multimodal context order performance across OpenFlamingo models with varying parameter scales.
• Figure 4: Performance of Qwen-VL-Chat-7B Qwen-VL, DeepSeek-VL-7B lu2024deepseekvl, IDEFICS-9B-Instruct laurencon2023obelics and IDEFICS-v2-8B-Instruct Laurenon2024WhatMW on image captioning task for modified COCO dataset.
• Figure 5: Performance of Qwen-VL-Chat-7B, DeepSeek-VL-7B, IDEFICS-9B-Instruct and IDEFICS-v2-8B-Instruct on image captioning task for modified COCO dataset, with different option positions.