This paper introduces thought-tracing, an inference-time reasoning algorithm inspired by sequential Monte Carlo and Bayesian Theory of Mind. The method is designed to track the mental states of agents by generating and weighting hypotheses based on observations, without relying on ground-truth answers. The authors evaluate the approach across diverse Theory-of-Mind benchmarks and report improvements over baseline prompting and reasoning methods. The work is relevant for social reasoning evaluation, model interpretability, and structured reasoning in settings without simple verification rules.

Author of the paper: Hyunwoo Kim, Melanie Sclar, Tan Zhi-Xuan, Lance Ying, Sydney Levine, Yang Liu, Joshua B. Tenenbaum, Yejin Choi

Publisher or journal of publication: arXiv / COLM 2025

The paper is available at the following link.