Can users be modeled as multiple personas instead of single vectors?
This explores whether representing a user as several distinct personas — rather than one fixed preference vector — captures their interests better, and what the corpus says that buys you.
This explores whether a user is better modeled as several distinct personas than as one fixed preference vector — and the corpus answers yes, with a clear account of why the single vector breaks down. The core problem is a bottleneck: a fixed-length user vector forces every interest a person has into one compressed representation, which loses information when tastes are genuinely diverse How can user vectors capture diverse interests without exploding in size?. The fix that recurs across the collection is candidate-conditional attention: instead of computing one user vector up front, you weight the user's history (or their latent personas) against the specific item being scored, so only the relevant interests activate at prediction time. Deep Interest Network does this with raw behaviors How can user vectors capture diverse interests without exploding in size?, while AMP-CF does it with explicit latent personas — and gets two things at once: better accuracy and built-in explanations, since each recommendation traces back to the particular persona it satisfies Can modeling multiple user personas improve recommendation accuracy? Can attention mechanisms reveal which user taste explains each recommendation?.
What's interesting is that the multi-persona idea shows up in two very different research conversations that don't usually talk to each other. One is recommendation, above. The other is the study of LLMs themselves, where 'persona' means the character a model adopts. There, researchers find a low-dimensional persona space whose dominant axis measures distance from the default 'Assistant' mode, and that conversations predictably push the model along it How stable is the trained Assistant personality in language models?. The shared insight across both worlds: an entity — a user or a model — is more faithfully described as a weighted mix of personas you can move between than as a single fixed point.
But a persona is only useful if it's stable, and the corpus is refreshingly honest about when it isn't. Run the same persona prompt repeatedly and the output varies as much across runs as it does across entirely different personas — meaning the model's own uncertainty, not durable knowledge, is driving the answers Why do LLM persona prompts produce inconsistent outputs across runs?. In multi-turn dialogue, simulated personas drift; training the simulator with consistency rewards cuts that drift by over half Can training user simulators reduce persona drift in dialogue?. So 'multiple personas' isn't free — you have to actively keep each one coherent.
The most forward-looking thread treats personas not as fixed slots but as something that evolves. PersonaAgent uses a persona as a living intermediary between a user's memory and their actions, tuning it at test time against recent interactions — and notably, the learned personas cluster into genuinely separate regions of latent space, evidence that they're capturing real per-user structure rather than noise Can personas evolve in real time to match what users actually want?. If you want to go further afield, the corpus also asks whether such personas are 'real' at all — one line of work argues post-training installs realized dispositions that survive adversarial pressure, not performances that collapse under a jailbreak Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?.
The thing you didn't know you wanted to know: modeling a user as multiple personas isn't just a accuracy trick — it's what makes a recommendation explainable. Once each suggestion is attributable to a specific persona, the system can tell you *why* it recommended something, and you can drop the separate, bolted-on step that usually forces diversity into results Can attention mechanisms reveal which user taste explains each recommendation?.
Sources 9 notes
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.
AMP-CF separates user representation into latent personas weighted by attention to the candidate item. This candidate-conditional approach improves accuracy by adapting the user representation at prediction time and produces inherent explanations for why items were recommended.
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.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
PersonaAgent uses structured personas to bridge episodic/semantic memory and personalized actions, optimizing them at test time by simulating recent interactions against textual feedback. Learned personas cluster meaningfully in latent space, suggesting genuine user-specific separation beyond standard post-training drift.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.