Imagine teaching an AI to play classic Atari games without giving it any points or scores. How would it know what to do? That’s the challenge researchers tackled in "OCALM: Object-Centric Assessment with Language Models." Instead of using traditional reward systems, they used the power of Large Language Models (LLMs) to define the goals. Think of it like giving the AI a coach that can explain the game's objective. This coach, the LLM, analyzes the game and provides feedback based on the relationships between objects, like "avoid colliding with cars" or "hit the ball past the opponent." This approach, called OCALM, transforms natural language descriptions into a reward system that an AI agent can easily understand. The researchers tested OCALM on four iconic Atari games: Pong, Freeway, Skiing, and Seaquest. The results were impressive. Even without access to the game’s score, the AI agents learned to play effectively using the LLM-generated reward system. In some games, like Freeway and Seaquest, the AI performed almost as well as agents trained with traditional point-based rewards. The key innovation lies in OCALM's ability to define rewards based on the relationships between objects, a more human-like way of understanding tasks. This approach makes the reward system more transparent and easier to debug. The study also highlighted the importance of this object-centric approach. AI agents that learned with these relational rewards outperformed those trained with simpler reward systems. This research opens exciting possibilities for training AI agents in complex environments without explicitly defined goals, such as robotics and autonomous driving. Imagine an LLM explaining to a robot how to navigate a cluttered room or describing driving rules to a self-driving car – the future of AI training might look very different than we expected.