When Rabbits Chase Tigers: Unpacking AI's Visual Understanding

Multimodal Large Language Models (MLLMs) have shown impressive capabilities in visual tasks. However, their effectiveness in handling scenarios contradicting everyday logic remains largely untested. This gap is what the CAIT benchmark aims to fill.

CAIT: Testing the Limits of Common Sense

CAIT, a newly introduced benchmark, consists of 400 synthetic scenes depicting counter-intuitive actions. Imagine a rabbit chasing a tiger. It's a scenario that defies common sense, deliberately crafted to test the limits of visual understanding in AI. The paper, published in Japanese, reveals that humans almost perfectly identify these scenarios, achieving about 95% accuracy.

Proprietary models like Claude and Gemini also perform well, showing up to 88% accuracy. Compare these numbers side by side. Open-source instruction-tuned models, however, barely rise above random guessing. The benchmark results speak for themselves.

The Challenge of Language Priors

Why do these models fail where others succeed? The data shows that standard models succumb to strong language priors. Instead of trusting the visual input, they lean on statistically common text descriptions, overriding the visual evidence. It's a fascinating but frustrating quirk of language-driven AI.

While Chain-of-Thought reasoning mechanisms offer a potential solution by boosting accuracy, they introduce a new issue: overthinking. Models sometimes refuse to accept the visual content simply because it clashes with real-world physical laws. Is it a case of AI becoming too smart for its own good?

Fine-Tuning: A Path Forward

There's hope yet. The research suggests that targeted fine-tuning and structured prompting can help models focus on the actual visual evidence. This could pave the way for more accurate and reliable open-source MLLMs. Western coverage has largely overlooked this, but it could redefine how we train and trust AI models.

Shouldn't we prioritize closing this performance gap? As AI continues to integrate into various sectors, we can't afford to ignore its shortcomings. Fine-tuning might seem like a minor technical detail now, but it could be the key to unlocking AI's potential in understanding our world in all its complexity.