With a proper reward choice, CURIO models achieve personalization without compromising coherence and overall quality. The baseline is trained to optimize conversation quality using exactly the same prompt for eval. DiffLogAcc outperforms the baseline and all other models. (7/9)
