Understanding and simulating how subjective cognition and mental states arise, evolve, and influence behavior is essential for advancing artificial intelligence toward deeper social cognition and general intelligence. Existing large language models (LLM) and Theory of Mind (ToM) approaches mainly infer mental states from linguistic cues, lacking a unified mechanism for constructing subjective worlds, modeling cognition‑psychology interactions, handling dynamic evolution, and preserving historical dependencies. This paper introduces the Mental World Model (MWM) and its concrete implementation, the Multi‑Branch Mental Evolution Architecture (MMEA). MMEA explicitly builds an agent’s subjective world, models the bidirectional influence between cognition and mental states, and simulates continuous mental evolution shaped by past experiences. On the DynToM and RecToM benchmarks, MMEA achieves 96.48% and 94.00% accuracy respectively, surpassing leading LLMs and specialized ToM methods. Additional experiments show improvements of 28.7, 39.3, and 16.6 percentage points in mental state and behavior prediction, long‑sequence prediction, and implicit mental state inference. These results demonstrate that MMEA offers an interpretable framework for continuous mental modeling, enabling AI to move beyond surface‑level semantic representations to infer underlying mental states from language and dynamically simulate internal cognitive processes over time. MWM also provides a promising foundation for causal reasoning, personalized AI, sustained human‑AI interaction, and future progress toward artificial general intelligence (AGI).
Review