Huimu Wang

Naive linear reward scalarization inevitably triggers destructive gradient tug-of-wars, but decoupling marginal objectives via orthogonal projections and localized constraints stabilizes training across both recommender systems and complex LLM alignment tasks.

Structure-grounded reasoning in generative recommendation can dramatically enhance performance, especially when traditional methods struggle with complex item relationships.

LLMs can act as effective action-level supervisors in reinforcement learning, dramatically boosting the sample efficiency of SAC without sacrificing convergence guarantees.