Large language models (LLMs) have shown an impressive ability to generate text, translate languages, and even write different kinds of creative content. But how good are they at recognizing and fixing their own mistakes? New research explores the self-correction capabilities of LLMs, revealing a fascinating dynamic between confidence and critique. Researchers discovered that LLMs can be surprisingly good at sticking with correct answers (confidence) but struggle more when it comes to identifying and correcting errors (critique). They dug deep into how these models self-correct, categorizing their behavior into four scenarios: confident (right answer maintained), unconfident (right answer changed to wrong), critical (wrong answer corrected), and stubborn (wrong answer maintained). To quantify these behaviors, they introduced innovative metrics: Confidence Level (CL) and Critique Score (CS). These metrics measure the probability of an LLM getting the answer right after self-correction, given that the initial answer was right or wrong, respectively. Interestingly, accuracy after self-correction is a weighted sum of CL and CS. This suggests a trade-off: improving confidence can sometimes hurt the ability to critique and vice versa. The research team experimented with different ways to influence self-correction, including prompting techniques and in-context learning. However, it proved difficult to improve both confidence and critique simultaneously without retraining the model. To tackle this, they developed a new training approach called Confidence and Critique Improvement Tuning (CCT). CCT transforms training data to specifically teach LLMs to both maintain confidence in correct answers and refine incorrect ones. This technique significantly boosted both CL and CS, showing that it’s possible to train LLMs to be better self-correctors. While this research focused on simplified scenarios, it reveals important insights into the inner workings of LLMs. The next challenge is to understand why some models are more confident than others and how these behaviors arise during pre-training. This work opens new avenues for improving the reliability and trustworthiness of LLMs in critical tasks.