A psychological prediction automatic reasoning framework for explaining potential cognitive states

Psychological prediction, the ability to infer human behavior and intentions, is essential for artificial intelligence and human–machine interaction. Existing deep learning models often rely on surface-level fitting and lack structured abductive interpretability, while traditional neuro-symbolic systems suffer from non-differentiable discontinuities and limited adaptability to dynamic contextual constraints. This paper proposes a neuro-symbolic automated reasoning framework for modeling latent cognitive states under contextual constraints. The framework introduces a context-aware adaptive SMC resampling operator (CAR-SMC), which uses energy-based penalties to filter hypotheses violating contextual constraints. We further design an executable differentiable program synthesis (EDPS) module and a hybrid iterative program–cognition interaction (HIPCI) mechanism. Through stop-gradient-based directional optimization, HIPCI separates the update pathways of belief and program representations during alternating fixed-point optimization, reducing circular gradient entanglement. Experiments on OpenToM and FANToM show that the framework improves the abductive consistency rate (ACR), an empirical belief–action consistency metric, to 70.8% and 52.1%, respectively. Under the adopted classification-oriented evaluation setup, these results indicate additional ACR gains over pure fine-tuned encoder classifiers and higher ACR than the evaluated few-shot ToT reference baseline (~ 7B), while using fewer parameters and maintaining an average inference latency of 215 ms per instance. Overall, this work provides a differentiable framework for machine Theory of Mind with process-level structural interpretability under the evaluated text-based classification setting.