Mind the Perspective: Let's Reason Recursively for Theory of Mind

The article presents RecToM, an inference-time framework designed to enhance Theory of Mind (ToM) reasoning in large language models (LLMs) by explicitly modeling nested beliefs through recursive perspective construction. This approach reduces complex higher-order belief questions to simpler actual-world questions, demonstrating superior performance on ToM benchmarks such as Hi-ToM, Big-ToM, and FanToM, achieving 100% accuracy with models like GPT-5.4 and Qwen3.5. RecToM's methodology offers a significant advancement for practitioners aiming to improve LLMs' reasoning capabilities in understanding agents' beliefs from limited observations.