Language models should be less confident when an input is ambiguous. We derive ambiguity vectors, and use them to improve confidence estimation in language models. We can also use them to encode these improvements into the model weights!
New work led by @themistoklisnk!
New work led by @themistoklisnk!
LLMs are highly overconfident, even when they don’t know something. We used interpretability to improve LLM confidence estimates!
We propose DART, which uses ambiguity steering vectors to guide fine-tuning, improving calibration across 7 models while preserving accuracy.
🧵View quoted post on X
