Temporal Moment Generation: A New Recommendation Task and Cost-Efficient Distillation for Outbound Marketing at Scale

Abstract

Outbound marketing at scale relies on content-first programs and item-ranking pipelines that match a fixed catalog to recent activity, missing the long-horizon interests and seasonal context (e.g., back-to-school, Earth Day) that make a message relevant. Missing these signals has played a key role in engagement declines of up to 18% year-over-year, with 55% of push notifications rated irrelevant by users. We make two contributions. First, we formalize temporal moment generation as a new recommendation task, distinct from item ranking: rather than scoring a catalog against a customer, the system starts from why a specific customer should be reached now. Given a customer behavioral signal profile and reference time T, it generates temporally-grounded moments---each a name, trigger, bounded activation window, and message hook, plus tiered search phrases for product retrieval---feeding a downstream retrieval and messaging pipeline. Second, we present a cost-efficient distillation approach that makes this task viable at production scale: starting from Claude Sonnet as a teacher, we distill into a self-hosted Qwen3-32B model via supervised fine-tuning followed by ORPO preference optimization, reducing inference cost 14× while maintaining production-grade output reliability and achieving holistic quality 0.788---4% above the teacher. Together they replace the conventional item-ranking pipeline with more relevant, better-timed messaging. Offline evaluation against traditional recommendation algorithms as baseline shows 5.2× composite-score lift (conservative lower bound; see \S4). The system is implemented and its scaling has been validated offline at production scale; a live A/B experiment is underway, with initial results presented at the session.