Apple researchers enabled agents to self-improve through internal feedback

Apple researchers recently introduced a technique called RLTL;DR to help AI models improve their performance on tasks where they previously struggled. The study focused on challenging coding and tool-calling problems that left standard AI models stuck with a success rate near zero.

Usually, AI models learn by repeating a task and focusing on the successful attempts. However, this method fails when the tasks are so hard that the model rarely succeeds, leaving it with little data to learn from. The new approach changes this by forcing the model to reflect after every failed attempt.

Instead of just retrying, the model acts as its own critic. It generates a short, written insight about why it failed. These insights are kept in the model's memory for the next attempt. This allows the AI to learn from its mistakes incrementally until it completes the goal.

Testing showed that a Qwen 3.5 9B Thinking model saw its success rate jump from near zero to as high as 31% using this method. The researchers also found that training the model on these self-generated insights alone was almost as effective as training on the full attempt data.

Imagine you are learning a new language and keep a diary of your mistakes so you never repeat them. Do you think this method of internalizing feedback could make AI models more reliable in your daily work?

This technique currently relies on the model having a specific capacity to generate and use these insights during the training process. The research does not show if this approach works for tasks outside of coding or specialized tool-calling, nor does it address long-term stability in open-ended creative scenarios.