Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models
Akshay Paruchuri, Sanmi Koyejo, Ehsan Adeli
Read on arXiv →
Key claim
All tested MLLMs exhibit significant order-invariance failures.
In plain English
Imagine you're building a system that combines text and images, like a chatbot that can also analyze photos. You want it to give consistent answers regardless of how you present the information. However, current benchmarks often assume that the order of inputs doesn't matter, which isn't true. When you shuffle the order of questions or images, the model's answers can change significantly, leading to unreliable results. This inconsistency is a problem because it means the model might not be as robust as you thought, especially in real-world applications where input order can vary widely. This issue is known as order-invariance failure.
To tackle this, the authors developed a new auditing tool called Facet-Probe. It evaluates models based on five different aspects of input order, helping to identify how much the order affects the model's responses. They used a Bayesian item-response model to separate the noise from the actual performance of the models, which allows for a clearer understanding of where the models struggle. Their findings showed that all 18 models they tested had significant order-related inconsistencies, with some flipping answers up to 50% of the time depending on input order.
This work is important because it highlights a major gap in how we evaluate multimodal models. By proposing a new standard for reporting order-related performance, it encourages future research to focus on making models more robust to input variations. For builders, this means that relying on traditional evaluation methods might lead to deploying models that perform poorly in real-world scenarios, and they should consider these new metrics when assessing model reliability.
The paper introduces a new auditing framework for evaluating multimodal large language models, addressing a significant gap in current evaluation methods.
The findings are supported by a robust Bayesian model and extensive testing across multiple models, providing solid evidence for the claims made.
Deep reliability assessment
The methodology supports a strong claim that the audited MLLMs are order-sensitive across the five tested ordering facets, with ODI helping separate item-level ordering noise from systematic facet bias. It is weaker as a universal deployment claim because the same-ordering decoder-noise control is mainly Gemini-based, mixed-modality scoring relies on LLM judges, and some facets or regimes are demoted due to heterogeneous or degenerate scoring.
Reproducibility
Yes, code and audit artifacts are linked on GitHub. The paper says it will not redistribute dataset content, but will provide permutation indices, model outputs, and aggregation scripts to reproduce the audit against upstream datasets.
Key figure
Figure 1 shows the same evidence presented in different semantically equivalent orders producing different MLLM answers, alongside a capability-versus-flip-rate plot where stronger models flip less but none become order-invariant.