Tracing Belief-Driven Thoughts with Theory-of-Mind Agents: An Opinion Analysis Framework

Opinion Analysis (OA) is a critical task in Natural Language Processing (NLP), aimed at identifying stance, sentiment, or hate speech toward specific targets in social media text. With the advancement of Large Language Models (LLMs), researchers have begun exploring LLM-based agent frameworks for OA. However, existing LLM-based agent frameworks rely on heuristic simulations and lack the capacity for emotional resonance, which is achieved by understanding others' internal thoughts, intentions, and beliefs during interactions-a cognitive ability rooted in Theory of Mind (ToM). To bridge this gap, we introduce OpinionToM, a multi-agent framework that shifts from heuristic simulation to ToM-driven cognitive reasoning. This shift requires a formal process capable of capturing key aspects of ToM, particularly the uncertainty each agent has about the hidden mental states of others. To this end, we model multi-agent reasoning in OA as a Social Partially Observable Markov Decision Process (Social POMDP). The Social POMDP is designed to track agents' continuous belief states by dynamically generating beliefs and adjusting corresponding weights based on observations. The weight adjustment mechanism draws inspiration from Bayesian inverse planning, leveraging LLMs as a computational backend to perform approximate probabilistic inference over competing beliefs, conditioned on the agents' received perceptions. We evaluate OpinionToM on six benchmarks across three distinct opinion analysis tasks, demonstrating significant performance improvements compared to baselines. We release the code and dataset at https://github.com/wenjt/OpinionToM.