4 ms·
What’s more interesting to me here are the calibration graphs: • LLMs, at least GPT models, tend to overstate their confidence. • A frequency-based approach a
by ggnore7452 2y ago
What’s more interesting to me here are the calibration graphs:
• LLMs, at least GPT models, tend to overstate their confidence.
• A frequency-based approach appears to achieve calibration closer to the ideal.
This kinda passes my vibe test. That said, I wonder—rather than running 100 trials, could we approximate this by using something like a log-probability ratio? This would especially apply in cases where answers are yes or no, assuming the output spans more than one token.
- GaggiX 2y agoyeah, this is by far the most interesting part of this page, the fact that LLMs can know what they know is not a trivial fact.
- ALittleLight 2y agoIf you imagine a future where LLMs get faster and cheaper even without getting better it means we'd be able to automatically repeat questions 100x and every answer could come with a pretty good confidence measure.