If an AI's memory of you is slightly wrong rather than totally wrong, does that make its mistakes worse?
What happens when a personalization graph contains stale or conflicting information?
This explores what goes wrong when the stored picture of a user (their profile, memory, or preference graph) is out of date or contradicts itself, and how personalization systems handle or mishandle that mismatch.
This explores what goes wrong when a system's stored picture of you is outdated or contradicts itself. The corpus has little work on graph-structured user memory going stale. What it does have, from several directions, is a surprising answer: bad information about a user usually doesn't fail loudly. It fails confidently. The clearest evidence comes from experiments that swap one user's profile for another's. When the swapped-in profile is obviously different, the damage is modest. When it is *nearly* right, performance drops the most, because the model applies the wrong preferences with full conviction. The researchers describe this as an uncanny valley Why do similar user profiles produce worse personalization errors?. A stale profile is exactly this kind of near-miss: it's still about you, just a version of you that has moved on.
That same line of work suggests one practical defense. When a system pulls up past interactions, recency-based recall beats similarity-based recall, and compact preference summaries beat retrieving raw episodes Does abstract preference knowledge outperform specific interaction recall?. Put simply, what you did lately is a better guide than what you once did that looks relevant. Recommender-systems research gives a sobering baseline, though: even 'fresh' data conflicts with itself. The same user rates the same item differently across sessions, sometimes by several stars, because of mood, anchoring, and personal rating habits Why do the same users rate items differently each time?. Some contradiction in a preference store isn't corruption. It's how people actually behave, and a system that treats every signal as stable truth will overfit to noise.
Conflict also happens between *kinds* of information, not just between old and new. Personalized LLMs struggle when profile cues (who you are) and query preferences (what you want right now) point in different directions. They lean on surface-level word similarity rather than reasoning about which signal should win Why do personalized language models fail when profiles and preferences diverge?. One proposed structural fix is to separate layers by how fast they change: a stable identity core wrapped in shells for behavior, cognition, and social context, so that a shift in short-term preferences doesn't force the system to relearn the whole person Should personalization systems model stable personality traits?. On the graph side specifically, knowledge-graph recommenders like KGAT spread signals across many hops of connections Can graphs unify collaborative filtering and side information?. The corpus doesn't test this directly, but by the same logic one outdated link in such a graph could spread its influence well beyond the node it sits on.
The part you may not have expected is that stale or thin information gets filled in rather than flagged. Every one of 12 LLMs tested invented between 35% and 49% of its claims about users, going beyond the evidence it had. The models that rated themselves as more careful were actually worse Do large language models fabricate user attributes beyond available evidence?. Stored context also leaks into unrelated tasks. ChatGPT changed how it framed politically neutral answers based on a political orientation it had inferred from memory Does ChatGPT shift responses based on inferred political views?, and across 13 models, personal context produced irrelevant callbacks, narrower answers, and excess agreement Does personalization make large language models worse at their jobs?. Combine these findings and a stale profile becomes a self-reinforcing trap: the system tailors to an outdated version of you, and per-user reward models then learn to please that version, which risks the echo-chamber dynamics already seen in recommender systems Does personalizing reward models amplify user echo chambers?.
The takeaway: the danger isn't that the system forgets you. It's that it remembers a slightly wrong you, acts on it with confidence, and has no built-in reason to notice. Weighting recent signals, separating stable traits from passing preferences, and treating contradictions as normal human variation are the strongest leads the corpus offers. Explicit methods for detecting and resolving conflicts inside a personalization graph remain a gap in this collection.
Sources 10 notes
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.
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.
Amatriain et al. found that the same user gives substantially different ratings to the same item across sessions, shifting by multiple stars. This noise stems from temporal inconsistency, rater-specific biases, and anchoring effects—making ratings reflect both preference and rating-behavior rather than stable preference alone.
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.
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.
Show all 10 sources
KGAT merges user-item interaction graphs with item knowledge graphs into a Collaborative Knowledge Graph, using attention-based propagation to capture both user-similarity and attribute-similarity signals simultaneously—including high-order connections that standard supervised learning methods miss.
MirageBench evaluated 12 LLMs across 7 families and found all of them over-infer user attributes in 35–49% of claims, driven by verbosity, reliance on pretraining priors, and genre expectations. Models that self-assess as over-inferring less actually over-infer more when judged independently.
A study of three GPT-4o personas found responses to politically neutral questions shifted systematically with inferred political views conveyed through memory or custom instructions. Republican-coded personas used economy and local framing; Democratic-coded personas used democracy and global framing.
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.
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.
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
- Tailored to you: longitudinal effects of personalising language models
- PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes
- Understanding the Role of User Profile in the Personalization of Large Language Models
- LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
- Interaction Context Often Increases Sycophancy in LLMs