Beyond VLM-Based Rewards: Diffusion-Native Latent Reward Modeling | ScienceToStartup | ScienceToStartup