Should the profile an app builds of you be readable, like a short bio, rather than a black box of numbers?

This explores whether the model of the user inside a personalization system should be something a person can read and check, such as text summaries, named personas or labeled traits, rather than an opaque vector. The corpus leans yes. The surprise is that readable representations often aren't a price paid for transparency. In several cases they perform better.

The strongest case is in the preference-learning work. When a summarizer and a reward model are trained together, the resulting text summaries condition the reward model more effectively than embedding vectors. They capture dimensions that zero-shot summaries miss, transfer to GPT-4 for zero-shot personalization, and users can still read them Can text summaries beat embeddings for personalized reward models?. A related finding is that abstract preference summaries beat retrieving specific past interactions Does abstract preference knowledge outperform specific interaction recall?. Recommendation research points the same way. Treating a user as several attention-weighted personas makes each suggestion traceable to the taste it satisfies, improves accuracy, and removes the need for a separate diversity step Can attention mechanisms reveal which user taste explains each recommendation? Can modeling multiple user personas improve recommendation accuracy?. Two other proposals show what the readable units could be. One is a stable identity core wrapped in four labeled shells, so the system doesn't relearn the person when the task changes Should personalization systems model stable personality traits?. The other is plain-language interest journeys like 'designing hydroponic systems for small spaces', which collaborative filtering never sees Can language models discover what users actually want from activity logs?.

The opaque side wins on cost and scale. User embeddings fed in through cross-attention beat text prompts on long histories and are cheaper to run Can user embeddings personalize language models more efficiently than prompts?. Reward factorization is compact too. It infers a user's coefficients from about ten adaptive questions without touching model weights Can user preferences be learned from just ten questions?, though the summary doesn't say whether those coefficients mean anything to a human. So the corpus shows a trade-off in efficiency, not a clear accuracy penalty for readability. It also doesn't force a choice, since a readable profile and an embedding can sit side by side.

The better argument for readability is safety. Personal context pushes models toward irrelevant personal references, narrower answers and too much agreement, and the user profile drives most of that shift Does personalization make large language models worse at their jobs?. Per-user reward models can also learn sycophancy and echo chambers once the averaging effect of aggregate models is gone Does personalizing reward models amplify user echo chambers?. With an opaque vector, nobody can point to what is pushing the model toward flattery. With a readable profile, someone can at least find the line and remove it.

Readable does not mean correct. Every one of 12 tested LLMs fabricated 35 to 49 percent of its claims about users, and models that rated themselves as over-inferring less actually over-inferred more Do large language models fabricate user attributes beyond available evidence?. An LLM-written profile is a fluent place for invented traits to hide, so interpretability only pays off if users can inspect and edit it. Another finding suggests what to put in it: profiles built from a user's outputs match or beat full profiles, while input-only profiles hurt, which means personalization runs on style and preference rather than topic Do user outputs outperform inputs for LLM personalization?. A useful readable profile should describe how someone writes and what they prefer, not just what they asked about.