OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind
Researchers have developed OSCToM, a novel approach to enhance Theory of Mind (ToM) reasoning in Large Language Models (LLMs), particularly in complex social scenarios involving nested belief conflicts. This method utilizes reinforcement learning and a specialized domain-specific language to generate challenging observer-self conflicts, pushing LLMs beyond simple perspective-taking. Experiments show that OSCToM-8B significantly improves performance on benchmarks like FANToM, achieving 76% accuracy compared to previous results, and demonstrates a more efficient data-synthesis procedure. AI
IMPACT Enhances LLM capabilities in complex social reasoning, potentially improving their application in interactive and strategic AI systems.