How do personalization granularity levels trade precision against scalability?

The Personalization of LLMs survey consolidates a taxonomy of personalization granularity that cuts across all implementation approaches:

User-level personalization — targets individual users via their specific history, interactions, and preferences (often identified by user ID). Highest precision and engagement. Faces data sparsity: new users have no history, infrequent users have thin profiles. Scaling is challenging because each user requires sufficient data.

Persona-level personalization — targets groups of users sharing characteristics or preferences. More scalable (groups have more data) and representative (captures common patterns). Less granular — individual deviations from group norms are missed. Requires domain knowledge to define meaningful personas. This connects to Why do LLM judges fail at predicting sparse user preferences? — persona-level information is often too sparse to predict specific preferences.

Global preference personalization — targets widely shared norms and standards. Broadest applicability and simplest to implement. Least specific — individual and group differences are flattened. Aggregation introduces noise from diverse populations. This is where Should AI alignment target preferences or social role norms? offers a critique: aggregation constitutes epistemic injustice when it silences minority perspectives.

Four technique categories map to these levels:

1. RAG — retrieves user-specific information from external knowledge base via embedding similarity
2. Prompting — incorporates user context into prompts for in-context learning
3. Representation learning — encodes user information into model embeddings or parameters
4. RLHF — uses user-specific feedback as reward signal for alignment

The survey reveals that direct personalized text generation and downstream task personalization appear distinct but share underlying components. Both retrieve user data, construct personalized prompts/embeddings, and leverage these for output. The key difference is evaluation: text generation evaluates against user-written ground truth; downstream tasks evaluate specific task metrics.

The formalization defines user documents (written content), user attributes (static demographics), user interactions (dynamic behaviors), and pair-wise preferences (explicit feedback) as distinct data types that personalization systems consume differently. Since Does chatbot personalization build trust or expose privacy risks?, the data types each carry different privacy implications — behavioral data is less visible to users than explicit preference queries.

Source: Assistants Personalization