Personalizing an AI doesn't mean sending everything you know about a user — so how little can you send?
How much of a user model must be sent per request for effective personalization?
This explores how small a slice of a user's profile or history an LLM needs to see on each request to personalize well, and whether the kind of content sent matters more than the amount.
This explores how small a slice of a user's profile or history an LLM needs on each request to personalize well, and whether the kind of content matters more than the amount. The corpus has one direct measurement. Retrieving 8 of 32 user-model fields matched the personalization quality of the full model, at 211 tokens instead of 915, about a quarter of the cost (Can a smaller user model subset match full model performance?). That is one study on one 32-field schema, so read it as "far less than everything" and not as a universal ratio. It also means you can decide per query what leaves the device, and sensitive fields can stay local.
Which quarter you send matters as much as how big it is. Profiles built only from what a user *produced* (their past outputs) matched or beat complete profiles across several tasks, while profiles built only from what they *asked* made results worse (Do user outputs outperform inputs for LLM personalization?). Personalization seems to run on style and preference, not on topic. A related finding points the same way: distilled preference summaries beat retrieved snippets of specific past interactions, and recent history beat merely similar history (Does abstract preference knowledge outperform specific interaction recall?). So a short, distilled statement of how someone writes and what they favor is worth more than a long transcript of what they said before.
Sending more can also backfire. A 13-model evaluation found that personal context pushes models toward irrelevant references to the user, narrower answers and too much agreement. The user profile caused most of this, because it shifts the model's goal from a balanced answer toward pleasing the user (Does personalization make large language models worse at their jobs?). A wrong slice is worse than a small one. When a replacement profile is nearly but not exactly right, the model applies the wrong preferences with confidence, and that does more damage than an obvious mismatch (Why do similar user profiles produce worse personalization errors?). This creates a tension with the 8-field result. Picking the right fields by semantic similarity can fail when the user's profile cues and the query sit in different concept spaces, and closing that gap takes explicit reasoning about concepts, not just retrieval (Why do personalized language models fail when profiles and preferences diverge?).
Some approaches skip sending text about the user altogether. Embeddings distilled from interaction history and fed in through cross-attention beat text prompts on long histories and cost less to run (Can user embeddings personalize language models more efficiently than prompts?). Lightweight adapters (PEFT) can hold a user's learned behavior as persistent local state, so one shared base model serves millions of people and the per-request payload is closer to "which adapter" than "which facts" (Can lightweight adapters replace millions of personalized models?). Other options need very little input. Ten well-chosen questions were enough to infer a user's reward coefficients, with no change to the model's weights (Can user preferences be learned from just ten questions?). And a compact description of a user's month-long interest, like "designing hydroponic systems for small spaces," can serve as the unit of a user model, one that recommenders built on click logs miss (Can language models discover what users actually want from activity logs?).
Sending less also helps with privacy. Longitudinal work finds that personalization raises trust and privacy concern together, and each good interaction lifts expectations for the next (Does chatbot personalization build trust or expose privacy risks?). The best current answer is a compact, distilled, output-based slice chosen per query, on the order of a quarter of a full model, or none at all if the personalization sits in an adapter or embedding. The open problem is choosing the slice reliably, because a wrong one is worse than a small one.
Sources 11 notes
Researchers found that retrieving only 8 of 32 user model fields achieved equivalent personalization performance while using 211 tokens instead of 915. This enables bounded disclosure per query while keeping sensitive fields local.
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.
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.
PRIME shows a U-shaped error curve where most-similar profile replacements cause steepest performance drops. The model confidently applies wrong preferences when profiles are nearly but not truly matched, an uncanny valley effect more harmful than obvious mismatch.
Show all 11 sources
Current personalized LLMs rely on shallow semantic correlations and fail when user profile cues and query preferences occupy different concept spaces. VIBE-Bench demonstrates this gap requires explicit concept-aware reasoning to bridge, not semantic retrieval alone.
User-LLM distills embeddings from diverse user interactions via self-supervised learning, then integrates them through cross-attention and soft-prompting. This approach outperforms text-based personalization on long-sequence and deep-understanding tasks while being computationally cheaper and preserving general knowledge.
PEFT adapters function as durable behavioral deltas carrying learned user experience, enabling a single strong base plus millions of lightweight adapters to replace millions of full models—but only when scale-up, scale-down, and scale-out reinforce simultaneously.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
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.
Longitudinal research shows personalization enhances trust and anthropomorphism but also amplifies privacy concerns and escalating user expectations. One-shot studies miss these temporal dynamics—each interaction raises the baseline, making failures more disappointing.
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
- Personalization of Large Language Models: A Survey
- Understanding the Role of User Profile in the Personalization of Large Language Models
- PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes
- User-LLM: Efficient LLM Contextualization with User Embeddings
- VIBE-Bench: Evaluating Personalized Large Language Models When Profiles Don't Mean Preferences
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization