Personalization that keeps 22.8% of profile tokens beats RAG in 36 of 39 cells
arXiv 2608.09507·medium signal
AlignXada learns preference adaptation via verbal reinforcement learning, gaining an average 3.82 points across 13 tasks while retaining only 22.8% of the original profile tokens, and outperforming retrieval-augmented generation in 36 of 39 tested cells. The token reduction is the practical hook — most personalization stacks stuff full user profiles into every request. This suggests aggressive distillation of the profile is not just cheaper but actually better than retrieving profile fragments on demand.