3 ms·
Competing Biases Underlie Overconfidence and Underconfidence in LLMs
- not_a_feature 7mo agoIntresting research that i experience myself when using LLMs. Models usually cling to their initial responses but they also overreact to contradictory feedback, deviating from optimal reasoning. These two competing biases, choice-supportive and contradiction-sensitive, should be kept in mind when using LLMs.