Interesting probing at how LLMs "think". Read their comment about running the model at temperature 0 to get raw logits