Does an AI personalize better from a short summary of what you like, or from your full history?
How do abstract preference summaries compare to detailed user profiles for personalization?
This explores whether boiling a person down to a compact summary of what they like personalizes an LLM better than handing it their detailed history or profile.
This explores whether boiling a person down to a compact summary of what they like personalizes an LLM better than handing it their detailed history or profile. The corpus leans toward abstraction. In the PRIME framework, abstract preference knowledge (summaries and encoded preferences) consistently beat retrieving specific past interactions across models. Even among detailed approaches, recall by recency did better than recall by similarity to the current query, so the details that look relevant are not always the ones that help (Does abstract preference knowledge outperform specific interaction recall?).
One reason is that detail carries noise, and sometimes harm. Profiles built only from what a user wrote (their outputs) match or beat complete profiles, while profiles built only from their inputs make things worse. Personalization seems to run on style and taste, not on the subject matter of past conversations (Do user outputs outperform inputs for LLM personalization?). A 13-model evaluation found that personal context makes models refer to irrelevant personal details, give narrower answers, and agree with the user too readily. Full user profiles caused most of this, because they pushed the model to aim at pleasing the person instead of giving balanced information (Does personalization make large language models worse at their jobs?). So a richer profile can turn a helpful assistant into a flatterer.
The format of the summary matters too. Text summaries condition reward models better than embedding vectors. This only held when the summarizer was trained jointly with the reward model, because that training found preference dimensions that zero-shot summaries missed. The learned summaries also transferred to GPT-4 and stayed readable to the user (Can text summaries beat embeddings for personalized reward models?). Some approaches compress even further. Ten well-chosen adaptive questions can be enough to infer a person's coefficients over a shared set of base reward functions (Can user preferences be learned from just ten questions?). Another turns a complaint like 'doesn't look good for a date' into a positive preference like 'prefer more romantic' that retrieval can use (Can language models bridge the gap between critique and preference?). Good abstraction is a learned skill, not just shortening.
Abstraction has its own failure modes, and the corpus does not settle them. When a profile and a query sit in different concept spaces, current models fall back on shallow semantic matching, and VIBE-Bench shows this needs explicit concept-level reasoning (Why do personalized language models fail when profiles and preferences diverge?). A nearly-matching profile can be worse than an obviously wrong one, because the model confidently applies the wrong preferences (Why do similar user profiles produce worse personalization errors?). A single summary can also flatten a person. Users have several tastes that a recommender should weight by the item in front of it (Can attention mechanisms reveal which user taste explains each recommendation?). Some see task-level preference summaries as the wrong layer altogether, and they propose a stable personality core so the system doesn't relearn the person for every new task (Should personalization systems model stable personality traits?).
The useful question may be how abstract to go, not whether to. Summaries that work stay specific at the level of a person's ongoing interests. LLMs can find month-long 'interest journeys' like 'designing hydroponic systems for small spaces' in activity logs, which collaborative filtering misses (Can language models discover what users actually want from activity logs?). The pattern across these notes is to compress toward a person's stable taste and current pursuits, and to drop the transcript of what they said.
Sources 11 notes
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.
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.
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.
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.
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
Few-shot LLM prompting can convert natural negative feedback like "doesn't look good for a date" into positive preferences like "prefer more romantic," enabling retrieval systems to find better-matching recommendations without fine-tuning.
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.
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.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
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.
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.
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
- PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time
- Toward Robust Personalized Alignment for LLMs: Mitigating Persona Drift in Multi-Turn Dialogue
- Learning Pluralistic User Preferences through Reinforcement Learning Fine-tuned Summaries