Most people interact with a tiny slice of what an app offers, so how can it personalize with so little data?

This explores what goes wrong when a system has only a thin trail of data about each individual user, and how the corpus's work on recommenders, personalized reward models and LLM judges tries to cope. The core problem is that personalization is a small-data problem in disguise. Recommendation systems serve millions of users, but each person touches less than 1% of the catalog, so the per-user signal is tiny even when the total dataset is huge. Why does collaborative filtering struggle with sparse user data? describes the standard fix, which is to share statistical strength across users so that a sparse individual signal becomes informative by borrowing from people who behave similarly.

Sparsity does more than reduce accuracy. It can make the system unreliable. Why do LLM judges fail at predicting sparse user preferences? finds that LLM judges asked to predict a specific person's preferences from a thin persona description often can't do it, because the description doesn't carry enough information to predict specific choices. The proposed remedy is to let the judge abstain. When the model states its own uncertainty verbally, filtering to the high-certainty cases recovers reliability above 80%. A related risk shows up in Why do similar user profiles produce worse personalization errors?. When a profile is nearly but not exactly right, models apply the wrong preferences confidently. Those near-miss errors turned out worse than obvious mismatches, and the corpus describes this as an uncanny valley. So filling a gap with a 'similar enough' user can backfire.

One response is to make each data point carry more information. Can user preferences be learned from just ten questions? represents a person as a mix of shared base preferences, so about ten well-chosen adaptive questions are enough to estimate that person's coefficients. Another response is to compress history into something denser. Does abstract preference knowledge outperform specific interaction recall? finds that abstracted preference summaries beat retrieving specific past interactions. Can text summaries beat embeddings for personalized reward models? adds that learned text summaries condition reward models better than embedding vectors, and they stay readable to the user. A thin history may go further if it is distilled into a few sentences of what someone likes than if it is stored as raw episodes.

Which part of the data you keep also matters. Do user outputs outperform inputs for LLM personalization? shows that profiles built only from what users wrote match or beat full profiles, while input-only profiles hurt. That suggests sparse data should be spent on style and preference signals, not on the topics of past queries. For recommenders whose explanations fail on sparse users, Can retrieval enhancement fix explainable recommendations for sparse users? adds outside signal by retrieving reviews and picking the aspects each user cares about, because embedding-based methods can't create signal that isn't there. Can modeling multiple user personas improve recommendation accuracy? points at a subtler problem, that a single user vector may blur several distinct tastes together. Its multiple-persona approach weights latent personas by the candidate item, so a sparse history isn't averaged into mush.

The corpus also warns that richer personal context isn't free. Does personalization make large language models worse at their jobs? found in 13 models that adding profile information pushed responses toward irrelevant personal references, narrower answers and excess agreement. So the sparsity fix has two sides. You need enough signal to know the person, but not so much raw profile that the model starts optimizing for pleasing them.