ParetoJev: Integrating Decision Model Signals into Multi-Objective Video Recommendation

Decision models turn a user context and a finite set of alternatives into candidate probabilities. Their value for recommendation depends on whether these probabilities add useful information to an already trained behavioral ranker. We study this question for short-video recommendation with six engagement objectives and negative feedback. ParetoJev combines seven objective-specific JEV-27B distributions with a compact EMER-based reranker, using a mixture of feedback-count and task-balanced ranking losses and validation-based fallback to a matched behavioral model. The study follows a complete experimental progression: MicroLens complementarity probes, aligned behavioral baselines, prompt-definition experiments, 36,000-slate model development, and a frozen 10,000-slate evaluation on later KuaiRand dates. On the latter cohort, comprising 100,000 candidate exposures and 6,443 users, the six-objective mean NDCG@3 rises from 0.298391 to 0.302943, a 1.53% relative improvement; the paired user-bootstrap interval for the absolute gain is [0.001147, 0.008248]. Gains are concentrated in engagement, especially forwarding. Replacing JEV-27B with an English-catalogue-trained Kev-0.8B checkpoint reduces the gain to 0.50%, with an interval crossing zero, despite a 26.65 ms median seven-question inference time on an A800 GPU. Additional analyses connect this result to nearly uniform Kev probabilities and show a larger point gain for users absent from reranker development. These results support decision signals as a complementary input to multi-objective ranking, while identifying task definition, balanced supervision, and selective use as central design choices.