Generative Divergence and Discriminative Convergence: A Collaborative Framework for Event Prediction

Event prediction aims to infer plausible future events from historical event sequences and plays an important role in public risk early warning and emergency decision support. Despite recent progress, generative event prediction models often suffer from two fundamental issues: semantic degeneration, where generated events become overly generic or stereotypical, and probability bias, which may lead to logically inconsistent or weakly grounded predictions. To address these challenges, we propose MSD‐DV, a collaborative framework that integrates generative divergence and discriminative convergence for event prediction. In the generative divergence stage, a multiview semantic decoupling mechanism disentangles heterogeneous behavioural and contextual dimensions, encouraging the model to explore complementary candidate events across diverse semantic subspaces and mitigating semantic degeneration. In the discriminative convergence stage, a dual verification mechanism evaluates forward semantic consistency and backward information specificity, progressively filtering logically inconsistent or under‐specified candidates to alleviate probability bias. Extensive experiments are conducted on two datasets, ROCStories and OpenForecast. The results show that MSD‐DV achieves consistent improvements over traditional sequence generation models and prompt‐based large language models across automatic and LLM‐based evaluation metrics. Compared with prompting and Chain‐of‐Thought baselines, MSD‐DV achieves improved semantic quality while generating predictions with better causal coherence and temporal consistency. Furthermore, quantitative analyses based on semantic informativeness and contextual contradiction demonstrate that MSD‐DV can effectively alleviate semantic degeneration and reduce contextually inconsistent predictions. The performance gains remain stable across different LLM backbones, demonstrating the robustness and model‐agnostic applicability of the proposed framework. These results indicate that combining structured semantic generation with explicit verification provides an effective paradigm for reliable open‐domain event prediction.