Today's AI models have different types of errors labeled as 'hallucinations,' making it challenging to tackle without advanced methods, says Sebastian Farquhar, a senior research fellow at the University of Oxford.
Farquhar's team developed a method to detect 'arbitrary and incorrect answers' named confabulations, focusing on computing uncertainty at the meaning level rather than specific word sequences, which outperforms similar methods with 79% accuracy.
The approach can identify confabulations by comparing responses' meaning, where different answers like Paris, Rome, and Berlin indicate likely errors, but it doesn't address biases due to flawed training data.
Collection
[
|
...
]