Stochastic Primal-Dual Decoding for Multiobjective Generative Recommender Systems

Short summary

The authors propose a lightweight inference-time decoding layer for autoregressive generative recommender systems that supports multiobjective slate generation without retraining. Decoding is formulated as online constrained optimization, balancing relevance and auxiliary objectives dynamically via a stochastic primal-dual scheme. A large-scale online A/B test shows +1.8% gains in auxiliary objectives at zero cost to user satisfaction, with theoretical guarantees on constraint violation and regret.

• •Plug-and-play decoding layer adds multiobjective support to generative recommender systems without retraining
• •Stochastic primal-dual approximation dynamically balances relevance vs. auxiliary constraints during generation
• •Online A/B test confirms +1.8% auxiliary objective gain with no user satisfaction cost

Generated with AI, which can make mistakes.