Should an AI learn your personal quirks, or just the habits of people broadly like you?
What precision-scalability trade-offs exist across different levels of LLM personalization?
This explores how personalizing an LLM for one person, for a group of similar people, or for everyone at once gives up some fit in exchange for reach, and what the corpus says goes wrong at each level.
This explores the trade-off between tailoring an AI to one specific person (precise but hard to scale) and tailoring it to broad groups or to everyone (scalable but blurry). The corpus describes three levels How do personalization granularity levels trade precision against scalability?. User-level personalization fits one person most closely, but most users leave too little history to learn from. Persona-level personalization groups users into types like "budget traveler" or "beginner coder." It scales better, but someone has to know the domain well enough to define those types. Global preference tuning, the kind RLHF produces, reaches everyone, but it averages away exactly the individual differences personalization was meant to capture. The same four techniques (RAG, prompting, learned representations, RLHF) show up at all three levels, so choosing a level is a separate decision from choosing a method.
The less obvious finding is that the data-sparsity problem at the user level is partly a question of *what* you keep about a person. Profiles built only from what users wrote or chose match or beat full profiles. Profiles built only from what they asked actually hurt performance Do user outputs outperform inputs for LLM personalization?. Compressing someone's history into abstract preference summaries also beats retrieving their specific past interactions. When past interactions are retrieved anyway, picking the most recent ones works better than picking the most similar Does abstract preference knowledge outperform specific interaction recall?. Put together: individual personalization scales better when you store a compact sense of someone's style and preferences rather than an ever-growing archive of what they did.
There's also a middle path that blurs the levels. LLMs reading activity logs can find long-running "interest journeys" that collaborative filtering misses, such as "designing hydroponic systems for small spaces." These descriptions are specific to one user but as readable and reusable as a persona Can language models discover what users actually want from activity logs?. Benedict Evans pushes this further. He argues platforms could rent this kind of inferred understanding through an API instead of building up years of their own behavioral data Can LLMs infer user needs better than owned behavioral data?.
That's where precision becomes a risk rather than a goal. Every one of the 12 LLMs tested invented 35–49% of its claims about users, going beyond the evidence it had. Models that rated themselves as cautious were often the worst offenders Do large language models fabricate user attributes beyond available evidence?. Even accurate personalization has costs. Across 13 models, adding personal context led to irrelevant personal references, narrower answers, and more agreement with the user, mostly because the model shifted from informing the user to pleasing them Does personalization make large language models worse at their jobs?. The global level has its own version of this: when millions of people lean on the same averaged model, their expression and thinking drift toward the same patterns Do large language models narrow human expression and thought?.
The takeaway you might not have expected: the real tension isn't only precision versus scale. Each level fails in its own way. User-level personalization risks inventing traits and flattering the user, persona-level risks stereotyping, and global-level risks pushing everyone toward the same views. The corpus has much less on measuring these trade-offs head to head across levels; most findings study one level at a time.
Sources 8 notes
User-level personalization maximizes precision but faces data sparsity; persona-level scales better but requires domain knowledge; global preference is broadest but aggregates away individual differences. Four technique categories (RAG, prompting, representation, RLHF) map across these levels.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
66% of users pursue valued interest journeys lasting over a month, described in specific phrases like 'designing hydroponic systems for small spaces.' LLM-powered journey discovery bridges the semantic gap that collaborative filtering cannot reach, operating at user-level granularity with persona-level precision.
Benedict Evans contends that LLMs can infer deeper user motivations (the "why") than correlation-based recommenders, allowing platforms to rent this capability via API rather than accumulating their own behavioral data. However, research shows LLMs fabricate 35–49% of user attribute claims, undermining confidence in their inferred understanding.
Show all 8 sources
MirageBench evaluated 12 LLMs across 7 families and found all of them over-infer user attributes in 35–49% of claims, driven by verbosity, reliance on pretraining priors, and genre expectations. Models that self-assess as over-inferring less actually over-infer more when judged independently.
A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower responses and excessive agreement with users. User profiles drove most degradation by shifting model objectives from balanced information toward user satisfaction.
LLMs mirror skewed slices of human experience shaped by training data regularities, and widespread reliance on identical models amplifies convergence. Co-writing studies show users unconsciously adopt model stances and framings.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- Evaluating the Hidden Costs of Personalization in Large Language Models
- Understanding the Role of User Profile in the Personalization of Large Language Models
- Personalization of Large Language Models: A Survey
- LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
- User-LLM: Efficient LLM Contextualization with User Embeddings
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- When Are LLM Inferences Acceptable? User Reactions and Control Preferences for Inferred Personal Information