PaperlayerView exact revision on arXiv

Source-linked AI summary

Creating an Atomic User Model for Personality-Aware Large Language Model Interaction

B. Sankar, Deepthika S, Pawni Yadav, Amogh A S

arXiv:2609.12086v2cs.HCcs.AIcs.CL

TL;DR

LLM personalization commonly stores users’ preferences, although preferences are situational and personality is comparatively stable. The paper proposes AUM, a readable structured profile used for task-conditioned retrieval; eight fields matched whole-model style fidelity with 23% of the context and improved on preference notes. The study used simulated participants, and the authors identify how AUM’s fields are populated as an open problem.

  • Problem

    LLM personalization stores observed actions or preferences rather than the comparatively stable personality that produces task-specific preferences.

    Source: §1 · p. 2

  • Method

    AUM is a structured, human-readable personality model used as an index, with task identification and budgeted retrieval selecting relevant fields.

    Source: §4.2 · p. 11, §5.1 · p. 14, §11 · p. 45

  • Results

    Eight retrieved fields matched whole-model style fidelity at 23% of the context, improved on preference notes by 0.24 points, and raised own-voice identification from 14.9% to 42.7% against 25% chance.

    Source: §11 · p. 45

  • Takeaways & Limitations

    A small, readable, task-conditioned payload was enough in the study, and personalisation benefited most the participants whose voices the default assistant served worst.

    Source: §11 · p. 45

    AUM specifies what fields a user representation should contain but not how they are populated, leaving acquisition an open problem for deployment.

    Source: §9.5 · p. 41

Abstract

from arXiv · show

Assistants built on large language models are expected to write in their users' own voice. Most systems summarise the user's preferences and include the summary in the prompt. This is the wrong way round. Preferences are only the surface of a person and change with the task, while the underlying personality stays the same, so storing preferences alone means relearning the user afresh whenever the task changes. This paper makes four contributions. First, we describe an effect we call personality seepage: the wording of a prompt carries traces of the writer's personality, which the assistant copies without knowing the writer. Second, we propose the Atomic User Model (AUM), a readable profile with a stable identity core surrounded by four layers covering psychological, cognitive, experiential, behavioral, and social details, plus notes on inner conflict and authenticity. Third, instead of inserting the entire profile, we use AUM as a searchable index, in which a task classifier, a selection step, and a budgeted retriever pass along only a few relevant fields. Fourth, we test the pipeline with 16 simulated users, 6 style-sensitive tasks, and 3 seeds. Eight retrieved fields matched the writing quality of the whole profile, while using only 23 percent of the context (211 tokens instead of 915). They scored 0.24 points higher than a plain preference note on a five-point scale. Accuracy in picking a user's own writing from four samples rose from 14.9 to 42.7 percent, where guessing gives 25 percent. Four pre-registered controls showed no effect, so the gain comes from the profile's structure rather than the search method. Personalization helps most for the users for whom a generic assistant imitates them the worst.

1. Introduction

The introduction argues that storing task-specific preferences misses the comparatively stable personality behind them. It proposes the Atomic User Model (AUM) and task-conditioned retrieval, and reports gains over preference notes with a smaller context payload.

1. Introduction

  • Preference notes store what users have done or said they want, not the person from whom those wants emerge.The paper argues that personality is comparatively stable while preferences are situational, so storing preferences alone leaves unseen tasks uncovered.Source: §1 · p. 2
  • Personality seepage describes how prompts carry users’ linguistic and behavioural patterns, which assistants mirror without access to the personality behind them.The paper proposes AUM as a structured, readable representation and treats it as an index from which task-relevant fields are retrieved.Source: §1 · pp. 2–3
  • Eight retrieved fields matched whole-model style fidelity using 23% of the context, improved on preference notes by 0.24 points, and raised own-voice identification from 14.9% to 42.7% against 25% chance.The identification increase is reported for forced-choice selection of the participant’s own voice.Source: §1 · p. 3
  • Four controls returned null, supporting the conclusion that the positive results measure the structured representation rather than the search method.The paper reports that the swarm retriever did not beat forward greedy; at the simulated operating point, greedy was within 0.1% of the exhaustive optimum.Source: §1 · p. 3
  • The reported benefit of personalisation was largest for participants whose voice the un-personalised assistant reproduced worst.The supplementary analysis reports Spearman ρ = −0.61, p = 0.013.Source: §1 · p. 3

2. Related Work

The related work places AUM at the intersection of explicit user schemas, personalization memory, and retrieval. It distinguishes the paper’s representation and evaluation choices from adjacent work on personality, privacy, writing assistance, and subset selection.

2.1. User models as components in adaptive systems; 2.2. Memory and personalisation in LLM assistants

  • AUM follows the tradition of separable, inspectable user schemas, while borrowing from information retrieval the need to select relevant, nonredundant profile fields.The paper frames existing LLM memory systems as addressing how to store and fetch user-relevant content, while AUM asks what should be stored and how it should be organized.Source: §2.1 · p. 4, §2.2 · p. 5

2.3. Context budgets and compression

  • Unlike prompt compression, AUM selects which fields enter a task-conditioned payload; compression could then shorten the selected fields.Source: §2.3 · p. 5

2.4. Personality: structure, stability, and its trace in language

  • Personality leaves measurable traces in language, but existing trait scores and latent-vector representations are neither modular nor inspectable; self-report also has known limits.The cited work contrasts self-knowledge of internal states with observers’ knowledge of externally visible traits.Source: §2.4 · pp. 5–6

2.5. Structured user representations and human digital twins

  • AUM is top-down, specifying slots for observations, unlike bottom-up representations; it targets style-sensitive task completion rather than identity substitution.Source: §2.5 · p. 6

2.6. Scrutability, control, and the privacy of a user model; 2.7. Writing assistance and the user’s voice

  • User scrutiny and control matter for user models, while AUM’s more intimate profile content sharpens privacy concerns; writing assistance makes whose voice appears consequential.The paper argues that if eight fields suffice, the other twenty-four need not leave the device.Source: §2.6 · p. 6, §2.7 · p. 7

2.8. Selecting a subset under a budget

  • The retrieval objective is not generally submodular, so the classical greedy guarantee does not transfer; at the simulated scale, the artificial fish swarm did not earn its cost.The paper identifies a regime in Section 8 where the answer changes.Source: §2.8 · p. 7

2.9. Simulated participants and model-based judgement

  • The study uses language-model-simulated participants as a feasibility method, not as evidence about human users, and describes controls for model-based judging.Source: §2.9 · p. 7

2.10. Cognitive architectures

  • AUM applies the cognitive-architecture case for explicit structure to representing users inside LLM-mediated systems.Source: §2.10 · p. 8

3. Personality, Preferences, Persona, and Seepage

The section distinguishes stable personality from task-dependent preferences and audience-dependent persona. It introduces personality seepage and argues that accumulated preference notes cannot interpret conflicts, stale assumptions, or underlying dispositions.

3.1. Three terms used in a specific way

  • Personality is a comparatively stable, task-independent set of traits; preferences are task-dependent outputs of personality in a decision context, while persona is its selective, audience-dependent projection.AUM models personality, derives preferences from personality and context at generation time, and may represent personas as fields inside Shell 4.Source: §3.1 · p. 8

3.2. Personality seepage

  • Personality seepage occurs when a prompt carries linguistic and behavioural patterns that the assistant mirrors without access to what produced them.The paper identifies missing prompt details, over-extrapolated surface features, and invisible stable habits as consequences.Source: §3.2 · p. 9

3.3. Why accumulating preferences treats the symptom

  • Accumulated preference notes address forgetting but not seepage: they are topically bound and lack structure for interpreting conflicts, stale notes, or shared underlying dispositions.The paper connects this limitation to preference notes being indistinguishable from chance on forced-choice identification of the user’s own voice.Source: §3.3 · pp. 9–10

4. The Atomic User Model

The Atomic User Model represents a person as a stable identity core surrounded by layers that differ in stability and observability. Its readable, fixed-slot fields support tier-specific updates and bounded disclosure, with cross-shell entries for conflict, growth, and authenticity.

4.1. The metaphor, and what it commits us to

  • The Nucleus changes over years while outer layers change over months or weeks; observability rises outward, encoding distinct refresh and exposure trade-offs.Treating all fields as equally durable risks wasteful or invasive core updates, or stale communication habits.Source: §4.1 · p. 10

4.2. Formal definition

  • AUM fields use short natural-language values, fixed schema slots, and nominal update periods τ that vary by tier.Users can inspect, edit, or delete values; the paper leaves the update mechanism unspecified.Source: §4.2 · pp. 10–11

4.3. The five tiers

  • The Nucleus stores core identity; Shells 1 and 2 describe psychological structure and the life history and cognitive styles shaping its expression.Shell 2 connects a person's Shell 1 profile to their particular behavioral patterns.Source: §4.3 · pp. 11–12
  • Shells 3 and 4 cover behavioral patterns and social context, while cross-shell entries track conflicts, deliberate growth, and alignment between inner and outer selves.The authenticity index can inform whether to match a user's professional register in a private message.Source: §4.3 · pp. 12–13

4.4. The full specification and the reduced instrument

  • The full specification has about 60 fields, but the studies use a reduced 32-field instrument: six fields per tier plus two cross-shell entries.The reduction favored fewer, denser simulated responses per field.Source: §4.4 · p. 13, Table 2 · p. 13

4.5. Two design commitments

  • AUM's intimate inner-shell content is intended to remain on the user's device or in a user-controlled vault; only selected payloads need leave it.Its human-readable fields can be read, edited, or deleted by the user, and eight short strings are few enough to show before sending.Source: §4.5 · p. 13

5. Personality-Aware Retrieval

Personality-aware retrieval treats AUM as an index queried under a context budget rather than injecting the whole profile. The pipeline routes simple requests around the model, selects fields for complex tasks, and returns a bounded payload designed to balance relevance, redundancy, and disclosure.

5.1. The four stages

  • Simple requests bypass AUM, while complex requests proceed through task-specific shell selection, retrieval of k fields, and generation conditioned on that payload.The task classifier labels both task type and complexity; the selected payload is inserted before the user's request.Source: §5.1 · p. 14

5.2. The payload objective

  • A budgeted payload trades query relevance against redundancy because near-duplicate fields waste slots; the resulting subset objective is combinatorial when λ > 0.The redundancy term is computed within the payload, and its form does not generally inherit the classical greedy guarantee.Source: §5.2 · p. 15

5.3. Retrievers

  • The study compares dense top-k, forward greedy, random-k control, and artificial fish swarm retrieval, with exhaustive search where feasible.Forward greedy is the deterministic baseline; random-k tests whether any k user strings help, while the swarm uses a fixed search budget.Source: §5.3 · pp. 15–16

5.4. Why local residence becomes measurable

  • Constraining payload size makes disclosure measurable: if k fields suffice, the remaining n − k fields stay on the device.The paper treats achievable k as both an efficiency and privacy measure.Source: §5.4 · p. 16

6. Simulation Study: Design

The study uses simulation to test whether a budgeted retrieval pipeline over a structured user model runs end to end and produces measurable differences, while recognizing that simulation cannot establish real users’ voice recognition or feelings about model storage.

6.1. Why simulate, and what a simulation can establish

  • The simulation can test end-to-end feasibility and measurable differences from text produced without the model, but not whether real people recognize their own voice or how they feel about model storage.The recognizing judge is a model reading a persona description, and the human question is reserved for a later study.Source: §6.1 · p. 16

6.2. Simulated participants

  • Sixteen code-declared personas span deliberately varied Big Five profiles and include life context, formative-experience seeds, and idiolect markers.The fixed declarations make runs reproducible from a seed and the population auditable.Source: §6.2 · p. 18, Table 4 · p. 18

6.3. Instrument population

  • Each simulated persona completes a 32-field instrument through guided first-person self-report, with missing or empty fields rejected and retried.Self-report remains a description rather than an observation, and self and other observers do not know the same things about a person.Source: §6.3 · pp. 18–19

6.4. Task battery, prompt surfaces, and references

  • The evaluation uses six complex, style-sensitive tasks; four simple factual lookups test classifier routing and are excluded from the main results.For complex tasks, the prompt surface is the only user-derived text seen by retrieval, while a separately generated reference serves as ground truth.Source: §6.4 · p. 19, Table 5 · p. 19

6.5. The preference-notes baseline

  • The preference-note baseline is distilled from four short conversations on held-out topics by a summarizer blind to both persona and user model.These choices prevent task rehearsal and keep the baseline a lossy distillation of observed conversation rather than a partial copy of the treatment.Source: §6.5 · p. 20

6.6. Conditions

  • Ten conditions share a base model and prompt; controls remove one component at a time to test random field selection, component selection, redundancy, and the swarm retriever.The controls distinguish retrieval effects from choosing any fields and test whether the swarm earns its cost.Source: §6.6 · p. 20

6.7. Measures

  • Fidelity, forced-choice identification, authorship-embedding style similarity, and prepended-token context measure different aspects of personalization.The judge does not see the AUM; the style measure uses authorship rather than semantic embeddings, and judge and generator must differ.Source: §6.7 · pp. 20–21

6.8. Models, seeds and cost; 6.9. Analysis plan

  • The analysis averages seeds within 96 participant-task cells, uses paired tests and confidence intervals, and labels post hoc supplementary analyses as non-confirmatory.API calls are cached under argument hashes, and reruns are reproducible from the manifest and a seed.Source: §6.9 · p. 21, §6.8 · p. 21

7. Simulation Study: Results

In the simulation, eight retrieved fields matched full-model fidelity using a fraction of the context and outperformed flat preference notes. Results across budgets, tasks, controls, and participants qualify where that advantage appears.

7.2. The budget curve

  • 2.86, 2.91, and 2.98 fidelity at k = 4, 8, and 16 show gradual gains; k = 8 is not saturation.No degradation appeared as fields were added at these budgets, so the case for a small k rests on cost and disclosure rather than quality.Source: §7.2 · pp. 23–24, Figure 6 · p. 25

7.1. Main comparison

7.4. Four controls return null

  • Four controls were null: random versus selected fields, no component filter, relevance-only top-k, and swarm versus greedy showed no downstream advantage.Source: §7.4 · pp. 24–25

7.5. Which shells each task type draws on

  • Communication payloads drew 44.4% from Shell 3 and 42.3% from Shell 4, while explanatory tasks drew 30.0% from Shell 2.The authors characterize this as evidence that retrieval is legible and behaves sensibly, not that legibility improves output quality.Source: §7.5 · p. 26

7.6. Which tasks the effect holds for

  • All six tasks favored AUM-R on average; gains peaked for explaining to a friend (+0.375) and declining invitations (+0.312), while deadline extensions were near zero (+0.021).Source: §7.6 · p. 27

7.3. Style similarity

  • Figure 9’s semantic embedding compresses condition differences because candidates share a task topic, making it a poor measure of style.Source: Figure 9 · p. 28

7.7. Robustness

  • The fidelity advantage remained after adjusting for length agreement (+0.227), and AUM-R favored 12 of 16 participants and all six tasks.Source: §7.7 · p. 28

7.8. Personalisation buys most where the default voice is furthest away

  • Personalization gains correlated negatively with generic-condition fidelity (ρ = −0.61, p = 0.013); the relation was not significant on LUAR (ρ = −0.12, p = 0.242).Participants with lower generic fidelity gained +0.333 versus +0.139 for those with higher generic fidelity, but that group difference did not reach significance.Source: §7.8 · pp. 29–30

8. Scaling Study of the Retriever

The scaling study isolates payload optimisation from language-model evaluation and tests when population search can improve on greedy. Greedy stayed near the best solution; swarm performance depended on its search budget and the redundancy penalty.

8.1. Separating two questions

  • The study separates retriever optimisation from language-model performance, asking when population search beats greedy on objective value and at what evaluation cost.The sweep required no API calls and was deterministic from a seed.Source: §8.1 · p. 31

8.2. Instance model

  • The synthetic instances clustered fields by shell, with queries drawing on two shells; the study used ρ = 0.65, d = 64, and 25 instances per cell.The sweep covered five instrument sizes, four budgets, and four redundancy weights, yielding 68 cells and 1,700 instances.Source: §8.2 · pp. 31–32, §8.3 · p. 33

8.3. Methods and budgets

  • The comparison included dense top-k, forward greedy, random selection, three AFSA variants, and exhaustive search where tractable.The AFSA variants used fixed, scaled, or greedy-warm-started budgets; solution quality was measured as the relative gap to the best solution found.Source: §8.3 · p. 32

8.4. Results

  • Forward greedy stayed within 0.10% of the best solution across the sweep and found the exact optimum in 77.7% of tractable instances.The sweep removed the language model and tested whether population methods could improve Equation 6 across instrument size, budget, and redundancy weight.Source: §8.1 · p. 31, §8.4 · p. 33
  • At fixed study budget, AFSA’s gap grew from 0.44% at n = 16 to 43.23% at n = 256, reflecting a budget that did not grow with the search space.It found the exact optimum in 46.3% of tractable instances, versus greedy’s 77.7%.Source: §8.4 · p. 33
  • Warm-started swarm stayed within 0.05% of the best solution at every scale and found the exact optimum in 96.3% of tractable instances, but cost roughly seven times greedy’s evaluations.It won or tied against greedy in 100% of instances, with 16.0% strict wins and 84.0% ties.Source: §8.4 · p. 34
  • At n = 64 and k = 8, dense top-k was optimal when λ = 0; its gap rose to 0.32%, 1.51%, and 7.68% at λ = 0.15, 0.30, and 0.60.Greedy also began to lose ground at λ = 0.60, with a 1.73% gap.Source: §8.4 · p. 34

8.5. What this means for deployment

  • For deployment, the authors recommend forward greedy; below roughly λ = 0.3, they suggest dense top-k if the redundancy weight will not be tuned.A warm-started population method is an option when objective value warrants roughly seven times greedy’s evaluations.Source: §8.5 · p. 35

9. Discussion

Eight retrieved fields matched the full model’s fidelity with a smaller, readable payload, which the authors frame as a measurable per-query privacy bound. The discussion also bounds the retrieval claims, describes uneven benefits, and identifies practical design choices, limitations, and next studies.

9.1. A context result, and why it is a privacy result

  • Eight fields matched the full model’s fidelity, making disclosure per query bounded and measurable while the full model remains on the user’s device or in a user-controlled vault.The authors also argue that a short payload can be shown to users before it is sent.Source: §9.1 · pp. 36–37

9.2. What the null results bound, and what they do not

  • The null controls locate the observed style effect in the structured representation, not in which slice was selected or how hard the system searched.At the tested scale, the admissible pool had about seventeen fields; the random-selection control drew only from task-plausible fields.Source: §9.2 · pp. 37–38

9.3. Personalisation is worth most to the people the default serves least

  • Personalisation benefits were largest for participants whose voice the un-personalised assistant reproduced worst (ρ = −0.61, p = 0.013).The authors note that this participant-level correlation did not replicate on the automatic style metric.Source: §9.3 · p. 38

9.4. Implications for the design of personalised assistants

  • The authors recommend storing stable user information, deriving situational preferences at generation time, budgeting and showing payloads, routing simple requests around the model, and evaluating benefit distributions.The task classifier routed factual lookups to bypass with 98.8% accuracy.Source: §9.4 · p. 39

9.5. Limitations

  • The study used simulated participants and authored personas, so its results do not establish effects with human users or support transferring benefit patterns to demographic groups.Other stated limits include moderate judge reliability, unspecified acquisition and updating, text-only English evaluation, and a single generator and judge model.Source: §9.5 · pp. 39–41

9.6. Future work

  • The next study should have participants populate their own AUMs, write references, and blindly rate outputs; further directions include acquisition, staleness detection, and testing selection at full specification scale.The proposed human study would use self-rated fidelity and forced-choice identification as primary measures.Source: §9.6 · pp. 42–43

10. Ethical Considerations

The paper treats an Atomic User Model as a sensitive record requiring architectural and user-control safeguards. It specifies local residence, user authorship, and purpose restriction, while acknowledging that those constraints cannot prevent misuse by an operator who ignores them.

  • A model containing fears, difficult episodes, and core values is highly sensitive, and opt-in status alone does not adequately address the risks of building one.The paper identifies misuse by operators seeking to influence users as a risk the architecture cannot prevent.Source: §10 · p. 43
  • The authors bind AUM to local residence, user authorship, and purpose restriction: fields are inspectable, editable, and deletable, and the model is not for profiling or persuasion against the user’s interest.Only the retrieved payload leaves the device or user-controlled vault.Source: §10 · p. 43
  • Budgeted retrieval bounds per-query disclosure; in the study, eight fields and 211 tokens matched the fidelity of the 32-field, 915-token model.A simple request sends nothing, while the complex-request payload can be shown before sending.Source: §10 · p. 43, Figure 22 · p. 45
  • The study’s benefit pattern concerns fictional, authored personas, so it should not be used to infer which demographic groups benefit most.The human study would require informed consent, withdrawal with deletion, ethics approval, and optional, reviewable inner-shell fields.Source: §10 · p. 44

11. Conclusion

The paper proposes storing a structured model of the person and retrieving task-relevant fields rather than treating preferences as a proxy for personality. In a simulation, eight fields matched the full model’s style fidelity with much less context; the authors conclude that the benefit was largest for participants least well served by the default voice.

  • The paper names personality seepage as the prompt’s personality fingerprint that an assistant mirrors without reading, and proposes AUM as a human-readable model with five tiers.Its retrieval pipeline uses AUM as an index rather than a prompt prefix.Source: §11 · p. 45
  • In a simulation with 16 participants, six tasks, and three seeds, eight retrieved fields matched full-model style fidelity using 23% of the context and improved on preference notes by 0.24 points.Forced-choice identification rose from 14.9% under preference notes to 42.7%, against 25% chance.Source: §11 · p. 45
  • The authors conclude that personalisation benefited most those participants whose voice the default assistant served worst.They frame a small, readable, task-conditioned payload as something a person can keep on-device and review before sending.Source: §11 · p. 45

CRediT authorship contribution statement

B. Sankar held the broadest set of listed roles, while the other contributors’ roles included methods, software, validation, and writing.

CRediT authorship contribution statement

Declaration of generative AI in the writing process

The authors used a large language model to assist with drafting and editing the manuscript’s prose, then reviewed and edited its content and accepted responsibility for it.

Declaration of generative AI in the writing process

Funding declaration

The authors report that the research received no specific grant from funding agencies.

Funding declaration

Appendix A. The Full Atomic User Model Specification

The appendix documents the Atomic User Model specification and the prompts, simulated participants, and evaluation details used in the studies. It distinguishes the reduced instrument tested in both studies from the broader specification and reports supporting diagnostics.

Appendix A. The Full Atomic User Model Specification

  • The full AUM specification has 60 fields across five tiers plus cross-shell integration, but only its 32 marked fields were used in both studies and have reported evidence.Fixed shell-name pairs support structural comparison; tier-level update periods support refresh policies, while cross-shell relations capture conflicts, alignment, confidence, and provenance.Source: Table A.19 · p. 53, Appendix A · p. 53

Appendix B. Prompts

  • Appendix B reproduces the simulation prompts verbatim, with placeholders substituted at runtime.Source: Appendix B · p. 53

Appendix B.3. Prompt surface

  • The retriever sees only user-derived text generated using a persona block and completed instrument; the prompt asks participants to write an assistant request in their own voice under 60 words.Source: Appendix B.3 · p. 54

{prompt_surface}

  • The classifier prompt requests JSON labels for simple or complex requests and one of seven task types.Source: {prompt_surface} · p. 55

Appendix C. The Simulated Participants

  • The study used 16 personas declared in code, each with Big Five z-scores, life context, formative-experience seed, and idiolect markers.The appendix provides their life contexts and formative backgrounds in full.Source: Appendix C · p. 56, Table C.20 · p. 57

Appendix D.1. Judge reliability across seeds

  • Across 960 targets, agreement across the three fidelity-judge seeds was ICC(2,1) = 0.573; the slot-choice distribution test gave χ2 = 7.37, 3 d.f., p = 0.061.The rating and identification distributions are reported over 2,880 trials.Source: Table D.21 · p. 58, Table D.22 · p. 58

Appendix D.3. Payload composition by task type

  • Figure 11 and Table D.23 report the share of each retrieved payload drawn from each shell, broken down by task type.Table D.23’s rows sum to 100% up to rounding.Source: Table D.23 · p. 59

Appendix D.4. Optimality of each retriever where exhaustive search is tractable

  • For the retriever scaling sweep, exhaustive search was computed when the search space was at most 60,000, using 25 instances per cell.Source: Appendix D.4 · p. 59
Loading 2609.12086v2…