Language models can write hundreds of lines of functional code, but fail to recognise critical user information just a few conversation turns back. A collaborative project between our group and
@FLAIR_Ox demonstrates this and other biases in LLM reasoning. Check it out!
Lize Alberts@lizegeezer
📢 Excited to introduce CURATe: A multi-turn benchmark assessing personalised alignment in LLMs through 1.6k conversations across 5 distinct scenarios We show systematic biases across 10 leading models in handling user-specific risk contexts ☠️
English
