Large-scale language models (LLMs) excel at mathematical and code reasoning, but struggle with social reasoning tasks, exhibiting cognitive confusion, logical inconsistencies, and confusion between objective world states and subjective belief states. Analysis of the inference trajectories of DeepSeek-R1 reveals that LLMs frequently encounter inference blockages when handling scenarios with multiple participants and timelines, outputting contradictory terms like "tricky" and "confusing," leading to erroneous inferences or infinite loops. The fundamental problem lies in the inability to separate objective reality from the agent's subjective beliefs. To address this, we propose an adaptive world model enhancement inference mechanism that builds a dynamic textual world model that tracks entity states and temporal sequences. This mechanism dynamically monitors inference trajectories for confusion indicators and provides a clear world state description, helping the model resolve cognitive dilemmas. This mechanism mimics how humans use implicit world models to distinguish internal beliefs from external events. Significant improvements in accuracy (e.g., +10% on Hi-ToM) and reduced computational costs (up to 33.8% token reduction) were observed across three social benchmark evaluations.
