We present a production LLM-based taxonomy classification system deployed at The Washington Post that tags news content across five schemas (Subject, Person, Company, Organization, Geography) using a proprietary taxonomy of ∼ 20,400 entries across seven hierarchical levels. For the Subject schema, we employ embedding-based candidate filtering followed by LLM selection. For other schemas, we combine LLM-based named entity extraction with fuzzy n-gram matching, followed by LLM selection. Comparison of post-production F1 scores against commercial vendor baselines demonstrates significant improvements across all five schemas, with the most substantial gain in Subject schema (+29.3%, p < 0.001). The system processes hundreds to thousands of articles and news items daily with a mean latency of 3–4 seconds per request and supports zero-downtime taxonomy updates.

We describe how The Washington Post deployed an offline reinforcement learning (RL) system to optimize paywall decisioning at production scale. We cast each non-subscriber article access attempt as a sequential decision with three actions: free access, registration wall, or subscription paywall, and learn policies from logged data collected via a small-traffic randomized controlled trial and subsequent production logging. We iterated from a tabular Q-learning baseline to a deep offline RL model trained with Conservative Q-Learning (CQL), using off-policy evaluation primarily to screen and rank candidates before online testing. The system was rolled out with guardrails and a persistent randomized holdout to manage risk in a revenue-critical setting. In year-long online experiments, the learned policies outperformed the legacy rules-based metering policy and improved a stakeholder-weighted value metric; the CQL policy delivered a +3% lift versus the randomized baseline while increasing subscriptions (+6%) and reducing the registration gap relative to earlier RL iterations. This case study highlights the practical steps needed to safely train, evaluate, and deploy offline RL for high-stakes personalization.

Publishers increasingly rely on access gates to drive registrations and subscriptions. Determining when to present these gates is a sequential decision problem well suited to reinforcement learning (RL). However, online exploration is costly and risky due to delayed conversion signals. We introduce Uncertainty-Aware Advantage-Weighted Actor–Critic (UA-AWAC), an offline RL method that learns from logged traffic to produce conversion-ready policies. UA-AWAC optimizes a multi-objective reward incorporating subscriptions, registrations, and engagement, while mitigating distribution shift through epistemic uncertainty modeling and pessimistic value targets. The policy is trained using advantage-weighted behavioral cloning with Kullback–Leibler (KL) regularization to remain close to historical gating behavior. Experiments show that UA-AWAC improves subscription rate by up to 10% and registration rate by 62% compared to baseline and state-of-the-art offline RL methods, demonstrating a practical and stable solution for intelligent content gating where exploration risks are high.

Session identification is a common strategy used to develop metrics for web analytics and perform behavioral analyses of user-facing systems. Past work has argued that session identification strategies based on an inactivity threshold is inherently arbitrary or has advocated that thresholds be set at about 30 minutes. In this work, we demonstrate a strong regularity in the temporal rhythms of user initiated events across several different domains of online activity (incl. video gaming, search, page views and volunteer contributions). We describe a methodology for identifying clusters of user activity and argue that the regularity with which these activity clusters appear implies a good rule-of-thumb inactivity threshold of about 1 hour. We conclude with implications that these temporal rhythms may have for system design based on our observations and theories of goal-directed human activity.