Reinforcement learning (RL) is a powerful tool for training AI agents, but designing the right reward function – the signal that tells the AI what constitutes "good" behavior – is notoriously difficult. Imagine trying to teach a self-driving car to navigate safely: how do you quantify and encode all the nuances of human driving into a reward? Researchers are exploring how to harness the power of large language models (LLMs) to create reward functions that capture human knowledge. In a new approach called REvolve (Reward Evolution), researchers are using LLMs not just to generate rewards but to evolve them over time based on human feedback. This method treats human preferences as a fitness function, guiding an evolutionary search for the best rewards. REvolve works by first generating multiple reward function candidates using an LLM like GPT-4. These candidates are then used to train different AI agents in a simulated environment. Human evaluators then watch videos of these agents performing the task (e.g., driving) and provide feedback, selecting which behaviors they prefer. This feedback is used to score the reward functions. The highest-scoring reward functions are then “mutated” or “crossed over” by the LLM, similar to genetic algorithms, to create a new generation of reward functions. This process repeats, with the reward functions continuously improving based on human input. The researchers tested REvolve in three challenging scenarios: autonomous driving, humanoid robot locomotion, and robotic hand manipulation. In all cases, REvolve outperformed traditional methods for designing reward functions and even came close to mirroring human expert performance in driving simulations. This research suggests LLMs can translate human preferences into a language AI can understand, making it a crucial tool for building more aligned and human-centric AI systems. While promising, REvolve currently relies on computationally expensive reinforcement learning and the closed-source GPT-4 model. Future research directions include making it compatible with open-source LLMs and testing its real-world applications by transferring the simulated training to real-world robotics.