Researchers at Concordia University have introduced a novel artificial intelligence framework, the Proximal Policy Optimization-based Cascaded Inference System, or PPO-CIS, engineered to enhance the detection and filtering of toxic online content. Published in the journal Knowledge-Based Systems in 2026, the system addresses the escalating computational burden of moderating the massive volume of user-generated material posted across digital platforms daily. The architecture departs from traditional single-stage moderation by implementing a multi-tiered cascade optimized through reinforcement learning. The framework employs a structured reward and penalty mechanism that trains the AI agent to dynamically balance classification accuracy with processing speed while adapting to shifting content patterns. PPO-CIS organizes its scanning pipeline into distinct layers: an initial rapid model filters incoming data and instantly discards benign posts. Content flagged as potentially harmful is escalated to a secondary, more computationally intensive classifier for deeper analysis. Any remaining ambiguous material is then routed to human moderators for final adjudication. This hierarchical approach enables platforms to calibrate moderation thresholds according to their specific community standards and regional regulatory mandates. Lead author Arezo Bodaghi, a doctoral graduate from Concordia’s Institute for Information Systems Engineering, highlighted the system’s configurability, noting that the algorithm can be adjusted to prioritize specific toxicity definitions established by individual platforms. The research team, which includes associate professor Ketra Schmitt and McGill University’s Benjamin Fung, embedded multiple base classifiers within the cascade to maximize individual model strengths while compensating for inherent limitations. According to the authors, this constitutes the first application of a cascaded reinforcement learning methodology specifically designed for online toxicity detection. Performance benchmarks conducted across the proprietary AugmenToxic dataset and the widely adopted ToxiGen dataset demonstrated clear advantages over contemporary moderation technologies. PPO-CIS achieved a two-point-one percent improvement in identification accuracy while substantially increasing processing throughput. The system processed 384 samples per second, compared to approximately 43 samples per second for comparable existing models. Additionally, the framework outperformed CETRA, an earlier reinforcement-learning architecture originally developed for malware detection. The convergence of heightened precision and rapid execution positions PPO-CIS as a practical solution for digital platforms subject to stringent compliance requirements. By accelerating the identification and removal of harmful material, the architecture provides a scalable moderation pipeline for services operating under strict jurisdictional deadlines. As online ecosystems grow increasingly complex, the integration of cascaded inference and adaptive reinforcement learning establishes a more efficient foundation for automated digital safety protocols.