Exploring LLM Response Logits for Bias and Model Insights

Hi folks, did you experiment with using LLM response logits distribution to obtain insights about model behaviour with respect to biases, hallucinations and other undesired effects? I understand that this applies to open-source models only where you have an access to logits. I am curious in general whether this direction has a potential from your perspective? Let me know if u did some research and feel free to link relevant papers if any!