[kick_in_the_dor]

I think you make a good point, but my guess is that e.g. your Taylor Swift example, a well-grounded model would have a low likelihood of outputting multiple consecutive answers about her being a comedian, which isn't grounded in the training data.

For your Tom Cruise example, since all those phrases are true and grounded in the training data, the technique may fire off a false positive "hallucination decision".

However, the example they give in the paper seems to be for "single-answer" questions, e.g., "What is the receptor that this very specific medication acts on?", or "Where is the Eiffel Tower located?", in which case I think this approach could be helpful. So perhaps this technique is best-suited for those single-answer applications.

[dwighttk]

What’s the single-answer for where the Eiffel Tower is located?

[bhaney]

48.8582° N, 2.2945° E

[dontlikeyoueith]

The Milky Way Galaxy.