Great post and thanks for sharing the technical details.
I was wondering if you tried the same approach with an unknown dataset, e.g one of your in-house dataset, and if LLM still attains such a high level of accuracy.
LLM may be good in detecting outlier, but that may be due to some underlying knowledge of the input dataset (the medium house value dataset is quite common and widely used in the community)?