The Complexity of Multimodal AI Testing and Benchmarking

Understanding Multimodal AI Evaluation

• 💡 Multimodal AI systems can process various data types like images, natural language, and audio simultaneously, making testing more complex and potentially expensive.
• ❓ Defining "multimodal" is crucial, as it can encompass audio, video, documents, 2D images, or 3D images, requiring clarity on the specific modes involved.
• 🧠 Architectures like omni-models (e.g., GPT-4, LLaVA) project different data modes into text tokens, which is how models like GPT-4 handle image inputs and outputs.

Benchmarking Multimodal Models

• 🎯 The core goal of benchmarking remains the same: assessing if an AI can perform a specific task.
• 📊 Multimodal benchmarks, such as MMMU (a multimodal version of MMLU), often use multiple-choice questions with image inputs.
• ⚠️ The relevance of image input should be secondary to the AI's ability to perform the task, whether it's trivia or using image information for another purpose.
• 🛠️ Hundreds of multimodal benchmarks exist, covering video, audio, and primarily image-based tasks.

LLMs as Judges for Multimodal AI

• 🤖 New LLMs are being developed to act as judges for multimodal evaluations, like Lava Critic, which scores image, question, and answer pairs.
• ⚖️ Some judges use LLMs to compare two answers to the same prompt (image + question) and determine which is better, providing a rationale.

Challenges with Image Output Evaluation

• 🧐 Evaluating AI systems that output images introduces more ambiguity.
• 🤔 It becomes challenging to verify if the AI ingesting an image to evaluate it is sufficiently capable of understanding that image's content.
• 🔁 This creates a potential self-fulfilling prophecy: if an AI is tasked with outputting an image and another AI must ingest an image to evaluate it, how do we confirm the evaluator's competence?