FlowLPS: Langevin-Proximal Sampling for Measurement-Feedback-Guided Generative Inference

World feedback offers measurable, non-human signals for aligning learning systems with real-world consequences. In imaging inverse problems, such feedback is available in a particularly explicit form: a known measurement process defines a physical consistency signal that evaluates whether a candidate reconstruction explains the observed measurement. This turns degradation-aware inverse solving into an inference-time alignment problem: generate samples that remain on the pretrained data prior while maximizing measurement-derived feedback. We propose FlowLPS, a training-free latent flow solver that uses this feedback under a finite compute budget. FlowLPS combines posterior-oriented Langevin exploration with local proximal feedback refinement: Langevin updates produce feedback-aware posterior-oriented candidates around the model trajectory, while proximal refinement rapidly improves measurement feedback from these candidates. A controlled pCN-style re-noising step stabilizes the reverse trajectory while preserving stochasticity. Experiments on FFHQ and DIV2K across five inverse problems show that FlowLPS achieves a strong balance between feedback satisfaction, measured by reconstruction fidelity, and perceptual realism, demonstrating how explicit world feedback can guide pretrained generative models without human labels or task-specific training.