If an app builds a detailed psychological profile of you, which parts should never leave your own device?
What inner-shell user model fields should never leave the device?
This explores which parts of a layered personal profile (a stable identity core wrapped in psychological, cognitive, behavioral and social shells) are too sensitive to send to a cloud model, and whether the corpus names them.
This explores which parts of a layered personal profile are too sensitive to send to a cloud model, and the corpus doesn't give a field-by-field list. The Atomic User Model organizes a person as a stable identity nucleus wrapped in four interpretable shells: psychological, cognitive, behavioral and social Should personalization systems model stable personality traits?. The material here doesn't say which shell counts as "inner" or which fields must stay local. The closest statement is that retrieving a small subset of a user model "enables bounded disclosure per query while keeping sensitive fields local" Can a smaller user model subset match full model performance?. It doesn't say which fields those are.
That same finding reframes the question. Eight of 32 fields matched the full model's personalization quality, using 211 tokens instead of 915 Can a smaller user model subset match full model performance?. So for any single query, most of the model doesn't need to leave the device. You can start from the default that everything stays home and let each query pull only what it needs, instead of maintaining a blacklist. My own inference is that the psychological and cognitive shells are the natural candidates for the strictest rule, since they hold the most intimate inferences about a person. The corpus doesn't test that.
Several findings suggest what can safely stay behind. Profiles built only from a user's outputs match or beat complete profiles, while input-only profiles hurt performance, because personalization runs on style and preferences rather than semantic content Do user outputs outperform inputs for LLM personalization?. Abstract preference summaries also consistently beat retrieving specific past interactions Does abstract preference knowledge outperform specific interaction recall?. Raw episodes and the content of what someone asked are therefore the cheapest things to withhold, and they cost little in quality. Sending less may even help. A 13-model study found that personal context, especially user profiles, pushes models toward irrelevant personal references, narrower answers and excessive agreement Does personalization make large language models worse at their jobs?.
There are two architectural routes to keeping the model itself local. Lightweight adapters can hold a person's learned behavior as persistent local state on top of one shared base model Can lightweight adapters replace millions of personalized models?. Distilled user embeddings can replace long text histories in the prompt Can user embeddings personalize language models more efficiently than prompts?. The corpus frames both as efficiency wins, not privacy guarantees, and it doesn't say whether an embedding leaks what the raw text would.
A final caution is that "never leaves the device" can't rest on asking a remote model to be discreet. A filter judges one output at one moment, and an agent's risk spreads across memory, retrieved content and tool calls, so containment means controlling what the agent can touch Can a model-level filter truly contain an agent with environment access?. For a user model, that means the never-send fields need to be excluded before the request is built, not filtered afterward. Which fields belong on that list is an open question in this corpus.
Sources 8 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.
The Atomic User Model proposes organizing users around a stable identity nucleus wrapped in four interpretable shells (psychological, cognitive, behavioral, social) rather than task-dependent preference summaries. This structure avoids relearning the person when tasks change.
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.
Show all 8 sources
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.
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.
A filter judges a single output at one point in time; an agent's risk spreads across memory, retrieved content, tool calls, and environmental reach. Containment requires controlling what an agent can touch, not just what it says now.
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
- PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes
- Understanding the Role of User Profile in the Personalization of Large Language Models
- PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- User-LLM: Efficient LLM Contextualization with User Embeddings