Guangtao nie

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.