“A paper from Meta on "Coda World Models" applies the concept of world models to LLMs, where the model learns to ensure the correctness of intermediate variables in a process, not just the final answer.”