To make an AI sound like you, is it better to show it your past writing or your past questions?
Do user outputs drive personalization more effectively than input queries?
This explores whether what a user has written or produced (their past responses and outputs) tells an LLM more about how to personalize than what they've asked it (their past queries and inputs).
This explores whether what a user has produced tells an LLM more about how to personalize than what they've asked it. In the corpus, the answer is yes, and the result is stronger than expected. User profiles built from outputs alone matched or beat complete profiles across multiple tasks, while profiles built from inputs alone made performance worse Do user outputs outperform inputs for LLM personalization?. The paper's explanation is that personalization runs on style and preferences, not semantic content. Your questions show what you were curious about that day. Your writing shows how you sound, and the second signal is the one that transfers to a new task.
The same pattern shows up in other work: distilled beats raw. The PRIME framework found that abstract preference summaries beat retrieving specific past interactions. Recency-based recall also beat similarity-based recall Does abstract preference knowledge outperform specific interaction recall?. Text summaries of preferences conditioned reward models better than embedding vectors, and users could still read them Can text summaries beat embeddings for personalized reward models?. Eight of 32 fields in a user model matched the full model at about a quarter of the token cost Can a smaller user model subset match full model performance?. Ten well-chosen questions were enough to infer a person's reward coefficients Can user preferences be learned from just ten questions?. Across these, the model needs a small, well-chosen signal of what you prefer, not a full transcript of what you did.
More personal context is also not automatically better, and it can backfire. A 13-model evaluation found that personal context pushed models toward irrelevant references to the user, narrower answers, and too much agreement. User profiles drove most of that damage by shifting the model from giving balanced information toward pleasing the user Does personalization make large language models worse at their jobs?. Near-matches are dangerous too. Swapping in a profile that was almost, but not quite, the user's produced the worst errors, because the model confidently applied the wrong preferences Why do similar user profiles produce worse personalization errors?. So the choice of signal matters as much as the amount.
Outputs are not the whole story, though, and this is my reading of how the pieces fit. The outputs result is about style and preference, which is the 'how'. Inputs still carry the 'what'. LLMs reading activity logs found persistent interest journeys, such as 'designing hydroponic systems for small spaces', that recommenders missed entirely Can language models discover what users actually want from activity logs?. A field study of movie recommenders found that context supplied by the user mattered more than prompt engineering Do LLM movie recommenders actually personalize to individual users?. When histories get long, compact user embeddings beat pasting text into the prompt Can user embeddings personalize language models more efficiently than prompts?. One proposal goes further and anchors everything on a stable personality core, so the system doesn't have to relearn the person for each task Should personalization systems model stable personality traits?.
The takeaway is that outputs win for matching voice, format and taste, and inputs are still the better source for topical interests. A raw dump of past queries is the weakest option, because it is mostly content that doesn't transfer. The strongest option is a short summary of how someone writes and what they value.
Sources 11 notes
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.
PLUS trains summarizers and reward models jointly, learning that text-based preference summaries capture dimensions zero-shot summaries miss. These summaries transfer to GPT-4 for zero-shot personalization and remain interpretable to users.
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.
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.
Show all 11 sources
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.
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.
A 160-user field study found LLMs deliver strong explainability yet lack personalization, diversity, and user trust. User-provided context matters more than prompt engineering, and LLMs perform better on niche items than mainstream ones.
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.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Evaluating the Hidden Costs of Personalization in Large Language Models
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- 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
- Learning Pluralistic User Preferences through Reinforcement Learning Fine-tuned Summaries
- PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time