Can preference dimensions extracted from outputs replace topic-based user summaries?
This explores whether you can personalize an AI by reading the *dimensions* of a user's taste (their style, their preferences, the 'how' of what they like) pulled from what they actually do — instead of building a topic summary of *what* they've talked about.
This explores whether you can personalize an AI by reading the dimensions of a user's taste — pulled from their outputs — rather than summarizing the topics they engage with. The corpus leans hard toward yes, and the most striking result is that *where* you look matters more than you'd guess. One study finds that profiles built from a user's outputs alone match or beat profiles built from everything, while profiles built from their input queries actually make things worse Do user outputs outperform inputs for LLM personalization?. The interpretation is that personalization rides on style and preference, not on the semantic content of what someone asked about — which is exactly the case for preferring extracted dimensions over topic summaries.
The form those dimensions take is the next lever. Abstract preference knowledge — compact summaries and parametric encodings — consistently outperforms retrieving specific past interactions, the topic-recall approach Does abstract preference knowledge outperform specific interaction recall?. And when you ask what *kind* of abstraction works best, text wins over vectors: learned text summaries condition a reward model more effectively than embeddings, capturing dimensions zero-shot summaries miss while staying readable to the user Can text summaries beat embeddings for personalized reward models?. So 'preference dimensions' isn't just an abstraction — it can be human-legible language about how you like things, not a black-box topic profile.
Pushed further, dimensions can become genuinely mathematical. One approach factorizes any user's reward as a linear combination over a shared set of base preference functions, then pins down a new person's coefficients with about ten well-chosen questions — no retraining, no topic history at all Can user preferences be learned from just ten questions?. That's the strongest version of the question's premise: a person reduced to weights on interpretable axes. The recommendation side reaches the same place from a different direction — modeling users as multiple weighted personas rather than one vector, with attention deciding which facet of taste explains each item, which delivers both accuracy and built-in explanations Can attention mechanisms reveal which user taste explains each recommendation? How can user vectors capture diverse interests without exploding in size?.
But 'replace' is too clean. Dimensions degrade when history is thin — sparse users force you to pull in outside review text to reconstruct enough signal to even produce an explanation Can retrieval enhancement fix explainable recommendations for sparse users?. And there's a quieter cost the corpus flags: the more sharply you fit one person's preference axes, the more you lose the averaging that aggregate models provide — and reward models tuned per user start amplifying sycophancy and echo chambers Does personalizing reward models amplify user echo chambers?. The thing that makes extracted dimensions powerful — they capture *you* specifically — is also what makes them able to wall you in.
The thing you didn't know you wanted to know: the field is quietly converging on the idea that *how* you express yourself is a better fingerprint than *what* you talk about — and a user can even be reconstructed from negatives, with an LLM flipping a critique like 'not right for a date' into the positive preference 'more romantic' that a retriever can actually use Can language models bridge the gap between critique and preference?.
Sources 9 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.
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.
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.
Deep Interest Network weights historical behaviors against each candidate ad, activating only relevant interests dynamically. This preserves dimension efficiency while expressing diverse tastes without lossy compression.
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.
Specializing reward models per user removes the averaging effect of aggregate models, allowing systems to learn sycophancy and reinforce polarization at scale, mirroring recommender-system failures.
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.