OOCO: Objective-Oriented Content Optimization on Discrete Symbolic Manifolds

We reframe controllable text revision as Objective-Oriented Content Optimization, a constrained stochastic optimization problem over a discrete sequence space. Given a source text and context, we maximize expected reader behavioral utility subject to semantic fidelity and perplexity constraints. To bypass non-differentiable token spaces, we compute directional derivatives of a surrogate objective function in continuous embedding space, decoupling gradient evaluation from sequence generation. We propose two gradient-guided search frameworks to map continuous ascent directions onto discrete transformation operators: Gradient-to-Strategy Mapping, which projects continuous derivatives via cosine alignment or a trained classifier, and Gradient-Guided Empirical Search, which scouts latent continuous trajectories to prune the operator space prior to local search. Evaluations on news virality and email engagement benchmarks demonstrate that gradient projection yields up to a 98.1 percent mean reward increase with competitive transfer to independent evaluators. Furthermore, hybrid trajectory scouting achieves competitive convergence while reducing empirical sample complexity by over thirtyfold compared to prompt-space optimization baselines.