TARS: MinMax Token-Adaptive Preference Strategy for Hallucination Reduction in MLLMs

Summary

This paper talks about TARS, a new method that helps multimodal large language models (MLLMs) reduce mistakes called hallucinations, where the model says things that sound right but are actually wrong or unrelated to the visuals.

What's the problem?

The problem is that these AI models often generate confident but incorrect answers because they learn from fixed preference data that can cause them to focus on superficial language patterns instead of truly understanding the meaning behind images and text.

What's the solution?

TARS fixes this by using a special training approach that changes certain parts of the input tokens in a controlled way to simulate different meanings while keeping the main idea the same. It then trains the model to stay accurate despite these changes, balancing between exploring variations and staying true to correct answers.

Why it matters?

This matters because it makes AI systems more reliable and trustworthy, especially for tasks that combine vision and language, by greatly lowering the chance of giving false or misleading information.