Creating an Atomic User Model for Personality-Aware Large Language Model Interaction
B. Sankar, Deepthika S, Pawni Yadav, Amogh A S
cs.HC · cs.AI · cs.CL
Quick verdict
AUM study: personality profile + budgeted field retrieval; gains are simulation-only (LLM personas/judge); the retrieval machinery adds nothing measurable.
Summary
The paper argues that LLM-assistant personalisation should store a person's stable personality structure rather than accumulated task-level preferences. It names 'personality seepage' — prompt wording carries personality traces the model mirrors without knowing the person — and proposes the Atomic User Model (AUM): a human-readable schema with a stable identity Nucleus, four shells (psychological, cognitive/experiential, behavioural, social) and cross-shell conflict/authenticity entries. AUM is treated as a retrieval index: a task classifier, a shell-selection function σ(t), and a budgeted retriever (relevance minus redundancy, λ=0.15) inject only k fields at generation time. In a simulation with 16 LLM-simulated participants, 6 style-sensitive tasks and 3 seeds (generator gpt-5.6-luna, judge gpt-5.6-terra), retrieving 8 fields (211 tokens) matched the fidelity of injecting all 32 fields (915 tokens), beat flat preference notes by 0.24 points on a 5-point scale (dz=0.50), and raised forced-choice voice identification from 14.9% to 42.7% against 25% chance. Four controls (random-k, no σ(t), dense top-k, fish-swarm) were null, and a synthetic scaling study shows greedy is within 0.1% of the exhaustive optimum at this pool size, locating the effect in the representation, not the search. Main limitations: entirely LLM-simulated participants with a same-family judge/generator; personas carry author-planted idiolect markers; judge reliability is moderate (ICC 0.57); acquisition and update mechanisms are unspecified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A prompt's wording carries a personality fingerprint ('personality seepage') that an assistant mirrors without access to the personality that produced it (Section 3).
- The Atomic User Model — a stable identity Nucleus plus four shells and cross-shell entries — is a better unit of storage for personalisation than accumulated preferences (Section 4).
- Retrieving 8 AUM fields (211 tokens) matches the style fidelity of injecting all 32 fields (915 tokens) and beats flat preference notes by 0.24 points on a 5-point scale, raising forced-choice voice identification from 14.9% to 42.7% against 25% chance (Sections 6-7).
- At the tested scale the selection machinery adds nothing measurable — random-k, no-σ(t), dense top-k and swarm controls are all null, and greedy is within 0.1% of the exhaustive optimum — so the effect lives in the structured representation (Sections 7.4, 8).
- Personalisation benefit concentrates on participants whose voice the un-personalised assistant reproduces worst (ρ=−0.61), an exploratory design finding (Section 7.8).
Strengths
Unusually honest null reporting
Four pre-registered controls reported in full (random-k, no σ(t), dense top-k, AFSA vs greedy; Section 7.4, Table 10-11), plus a synthetic scaling study (Section 8, Table 16) that explains the nulls mechanistically: greedy is within 0.1% of the exhaustive optimum at pool ~17, so the effect is localised to the representation, not the search.
Statistical discipline
96 (participant, task) cells with seeds averaged before testing (explicitly avoiding 288-observation inflation), Holm correction within metric, Wilcoxon primary for ordinal fidelity, 10k-resample bootstrap CIs, dz and Cliff's δ, exact binomial with Wilson intervals for identification (Section 6.9, Table 10).
Confound checking
Reproducibility detail
Privacy argument made measurable
Weaknesses
Entire evidence base is LLM-simulated with same-family models
The key null undercuts the retrieval contribution
Personas carry author-planted style markers
Preference-notes baseline is synthetic and anomalous
Headline correlation is fragile
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets LLM-assistant personalisation that stores summarised preferences and reinjects them, arguing this inverts the natural inference order: preferences are the task-dependent surface of a comparatively stable personality (citing Roberts & DelVecchio 2000 for rank-order stability). Four claims: (1) a phenomenon, *personality seepage* — the prompt surface carries a personality fingerprint the assistant mirrors (Section 3, Table 1); (2) the Atomic User Model, a scrutable schema with a stable Nucleus, four shells and cross-shell conflict/authenticity entries encoding differential stability and observability (Section 4, Table 2: 32-field reduced instrument); (3) personality-aware retrieval — task classifier, shell-selection σ(t), budgeted retriever maximising mean relevance minus λ·mean pairwise redundancy, λ=0.15 (Section 5, Eq. 6); (4) evidence that 8 retrieved fields match full-model injection at 23% of context, beat preference notes by +0.24 on a 5-point scale, and raise forced-choice voice identification from 14.9% to 42.7% (Sections 6–7). All evidence is LLM-simulated; the paper says so in the abstract and again in Section 6.1.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Atomic User Model (AUM)?
AUM is a structured, human-readable user profile proposed in this paper: a stable identity core (Nucleus) surrounded by four layers covering psychological, cognitive/experiential, behavioural and social content, plus cross-layer entries for internal conflict and authenticity. The full specification has about 60 fields; the studies use a reduced 32-field version. Values are short natural-language sentences a user can read and edit, and each field carries a nominal update period (years at the core, weeks at the outer layers).
Was this study run with real human users?
No. All 16 participants are LLM-simulated personas, the texts are generated by an LLM (gpt-5.6-luna), and a second LLM (gpt-5.6-terra) does the judging. The authors state explicitly that no result should be transferred to human users, and propose a human study as the necessary next step.
How much does AUM improve personalisation compared with simple preference notes?
In the simulation, retrieving 8 AUM fields improved judge-rated style fidelity by 0.24 points on a 5-point scale over flat preference notes, and raised forced-choice identification of the participant's own voice from 14.9% (not above chance) to 42.7% (chance is 25%), while using 211 tokens versus 915 for the full profile.
Does the paper's fancy retrieval algorithm matter for the results?
No, not at the tested scale. Randomly chosen fields from the same task-relevant pool performed the same as optimised selection, the fish-swarm optimiser did not beat simple greedy search, and a separate scaling study shows greedy is already within 0.1% of the exhaustive optimum at this pool size. The benefit comes from injecting a structured, task-plausible slice of the profile, not from how the slice is chosen.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt (head and tail kept, three marked '[...sampling gap...]' regions); mid-paper content — parts of Section 6 design detail, part of Section 8, and the Appendix A field table — was not visible, so judgments about those parts rely only on what is shown.
- No public code or data release URL appears in the provided text; 'codeAvailable: false' records that absence, not a verified lack of release, which a sampling gap could hide.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.