Intresting 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.
