Your past writing and choices reveal what you want from an AI better than the request you just typed.
Why does personalization depend more on user history than query semantics?
This explores why knowing what a person has done and written before predicts what they want better than the literal meaning of their current request does.
This explores why a person's past behavior tells a model more about what they want than the words of their current request. The corpus suggests a query says what someone is asking about, while history says how they want it answered, and personalization is mostly about the second. The sharpest evidence is that user profiles built only from a person's past *outputs* (what they wrote, chose or produced) matched or beat full profiles. Profiles built only from their past *inputs*, the queries, made performance worse (Do user outputs outperform inputs for LLM personalization?). What carries over between tasks is style and taste, not subject matter.
This also explains why matching on meaning is a poor way to use history. In one study, simply recalling the most *recent* interactions beat retrieving the *most similar* ones, and short summaries of a user's preferences beat replaying specific past episodes (Does abstract preference knowledge outperform specific interaction recall?). Personalization models that lean on surface semantic overlap also break down when the cues in a profile and the preferences behind a query live in different concept spaces, and closing that gap takes explicit reasoning about concepts, not better retrieval (Why do personalized language models fail when profiles and preferences diverge?).
History also holds patterns no single query can show. One analysis found that about 66% of users pursue an interest for over a month, such as 'designing hydroponic systems for small spaces'. Recommenders miss these journeys because they only see individual clicks (Can language models discover what users actually want from activity logs?). Embeddings distilled from long interaction histories beat pasting that history into a prompt as text, and they cost less (Can user embeddings personalize language models more efficiently than prompts?). The useful signal is also compact. Eight of 32 user-model fields matched the full model at about a quarter of the tokens (Can a smaller user model subset match full model performance?), and ten well-chosen questions were enough to infer a person's reward preferences (Can user preferences be learned from just ten questions?). A few durable traits do most of the work.
History-heavy personalization has costs. The more a profile resembles you without actually being you, the worse it gets: near-matching profiles produced bigger errors than obviously wrong ones, because the model confidently applied the wrong preferences (Why do similar user profiles produce worse personalization errors?). Across 13 models, adding personal context led to irrelevant personal references, narrower answers and more agreement with the user, because the profile shifted the model's goal from balanced information toward user satisfaction (Does personalization make large language models worse at their jobs?). Even reasoning has to use the profile. Generic chain-of-thought underperforms on personalization tasks because it ignores user context (Why does chain-of-thought reasoning fail for personalization?).
History wins only when there is enough of it. For sparse users, retrieval augmentation can supply richer signal (Can retrieval enhancement fix explainable recommendations for sparse users?). Aggregated clicks from other users can also expose article relationships that no single person's short history reveals (Can cross-user behavior reveal news relations that individual histories miss?). So the answer isn't that queries don't matter. The query says what to talk about, and history is the only place to learn how this particular person wants it said, with the crowd filling in when the history runs thin.
Sources 12 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.
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.
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.
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.
Show all 12 sources
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.
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.
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.
Generic chain-of-thought underperforms for personalization because it ignores user context. Fine-tuning destroys reasoning capacity entirely. Self-distillation lets models generate customized thinking traces that maintain both depth and relevance.
ERRA combines model-agnostic review retrieval with personalized aspect selection to address data sparsity that embedded methods cannot solve. Retrieval augmentation provides richer signal when user history is sparse, while aspect personalization ensures explanations match user context rather than generic defaults.
GLORY constructs a global news graph from aggregated user clicks to discover article relationships invisible in any single user's sparse history. This population-level behavioral structure enables recommendations even when direct textual or per-user similarity fails.
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
- PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes
- Personalization of Large Language Models: A Survey
- Understanding the Role of User Profile in the Personalization of Large Language Models
- User-LLM: Efficient LLM Contextualization with User Embeddings
- VIBE-Bench: Evaluating Personalized Large Language Models When Profiles Don't Mean Preferences
- Learning Pluralistic User Preferences through Reinforcement Learning Fine-tuned Summaries