EpiPersona: Persona Projection and Episode Coupling for Pluralistic Preference Modeling

Research question

Whether a learned discrete representation of preference history, conceptually separated from the current episode and then coupled to it, improves prediction of the response a new user will choose compared to existing textual summaries, retrieval, and personalized reward models.

Method

Two variants with LoRA on Llama: EpiPersona-A encodes each historical comparison, quantizes it into a persona codebook, aggregates the codes, generates via distillation from Qwen-2.5-72B a JSON of preferences, and asks GPT-OSS-120B or Llama-3.3-70B to judge two responses. EpiPersona-B combines the latent representation, the episode, and each response in a Bradley-Terry model with Llama-3.2-3B or Llama-3.1-8B. It evaluates accuracy, permutes the order of candidates in A, and adds ablations and exploratory analyses of similarity, amount of history, and window.

Sample: Arena: 131 users and 1,338 pairs; train 23 users with 202 historical/103 current, validation 19 with 155/80 and test 89 with 514/284. PRISM: 723 users and 16,705 pairs; train 280 with 4,192/2,177, validation 65 with 973/514 and test 378 with 5,816/3,033. In Task A each current pair is judged in both orders, duplicating observations but not independent pairs. Test users are new to training, although their own history is provided to the system.

Findings

• EpiPersona-A reaches 59.38%/59.15% on PRISM and 66.07%/65.01% on Arena with the two judges; it is highest in four columns.
• Compared to the best baseline of each column, the advantage of A is only 1.90, 1.48, 1.18, and 0.62 points; approximately 3-3.6% of the text selectively compares with PersonalLLM.
• EpiPersona-B obtains 59.60%/58.54% on PRISM and 59.57%/57.45% on Arena; PAL reaches 60.56% on Arena-3B and surpasses it in that column.
• Removing VQ or the encoder reduces A between 1.39 and 3.41 points, but the intervals overlap and there is no paired test or distribution across seeds.
• Figure 3A shows delta-y/delta-x slopes of 0.043, 0.178, and 0.144; the text incorrectly calls them absolute drops in performance.
• The published counts exactly match the sizes of the official partitions of the derived datasets on Hugging Face.

Limitations

• The codes are trained to predict choices; they are not validated against traits, values, identity, temporal stability, or external measures.
• There is no test of disentanglement, episode leakage, semantics, utilization, or codebook collapse, invariance, or counterfactuals.
• Although the codebook has no named axes, distillation imposes persona, value, identification, style, and intent, and the prompt orders inferring stability from a single comparison.
• There is no analysis by minority, country, culture, demographics, worst group, or equity; the rhetoric about minorities is not contrasted.
• Predicting a choice among existing responses does not evaluate aligned generation, well-being, safety, or resolution of collective preferences.
• Unseen user means a user not used in training but with their own history at inference; it is not generalization without data from the individual.
• The low Qwen3 similarity is an exploratory post hoc cutoff, not a temporal, causal, or external shift.
• The more/less values reproduce naive binomial intervals that count both permutations and multiple pairs from the same user as independent.
• The tables call the same values standard deviations and 95% confidence intervals, two incompatible concepts.
• There are no paired tests, bootstrap by user, hierarchical model, or uncertainty of the difference between methods.
• The figure reports accuracy slopes with respect to similarity, but the prose presents them as accuracy drops.
• The sparsity and window analyses lack exact denominators, tests, and complete baselines; a window of up to 60 is not explained with mostly short Arena histories.
• The main equations of EpiPersona-B omit the episode e, while the appendix algorithm does include it.
• Missing are optimizer, epochs/early stopping, seeds, number of runs, exact checkpoints, EMA, K-means, window construction, checkpoint selection, and distillation configuration.
• There is no repo, code, checkpoints, outputs, generated profiles, latent codes, environment, or EpiPersona pipeline; release is promised in the future.
• The derived datasets are open and verifiable, but their cards only contain schema/size and do not document license or terms of use.
• The paper acknowledges privacy, sensitive inference, manipulation, and surveillance, but does not evaluate attribute attacks, stereotypes, fairness, erasure, or mitigations.

What the study does not establish

• It does not demonstrate that the latent codes represent stable personality or that they semantically separate persona and episode.
• It does not establish a statistically reliable improvement over baselines: the differences are small and dependence is ignored.
• It does not prove that EpiPersona-B is always superior; PAL wins on Arena with 3B.
• It does not demonstrate benefit for minorities, demographic equity, or collective pluralist alignment.
• It does not prove generalization without the new user's history or under temporal or external distribution.
• It does not evaluate an aligned generative assistant, utility, safety, or well-being in deployment.
• It does not allow reproducing EpiPersona training and results from public artifacts.
• It does not convert the slopes 0.043/0.178/0.144 into absolute accuracy drops; that reading of the text is incorrect.

Traceability

Scope: Full text

Version: arXiv:2603.28197v1

Consulted source: https://arxiv.org/abs/2603.28197v1

Review: Codex 20-page visual full-text, official arXiv metadata, upstream SynthesizeMe code, Hugging Face split, construct, statistical, minority-claim, ethics and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

• EpiPersona-A
• EpiPersona-B
• Llama-3.1-8B
• Llama-3.2-3B
• Llama-3.3-70B-Instruct
• GPT-OSS-120B
• Qwen-2.5-72B
• Qwen3-Embedding

Instruments and metrics

• Predicción de preferencia entre pares
• LLM-as-a-judge con permutación de orden
• Accuracy micro por comparación
• Ablaciones de VQ y encoder
• Similitud episodio-historial con Qwen3-Embedding
• Análisis exploratorio de cantidad de feedback y ventana

Data used

• MichaelR207/chatbot_arena_personalized_0125, derivado de Chatbot Arena para SynthesizeMe
• MichaelR207/prism_personalized_0125, derivado de PRISM para SynthesizeMe

Evidence and location

• Architecture, stable persona assumptions, VQ, and coupling: arXiv v1, pp. 1-5 and Appendix Algorithms 1-2
• Task A/B results, ablation, similarity, sparsity, and window: arXiv v1, pp. 6-8, Tables 1-3, Figure 3; Appendix Figure 6
• Partitions and sizes by user/historical-current: arXiv v1, p. 12, Table 4; official Hugging Face dataset size endpoints inspected 2026-07-17
• Training parameters, codebooks, and Prism-Arena pretraining: arXiv v1, p. 13, Appendix A.5
• Fixed ontology and stable/dynamic inference prompts: arXiv v1, pp. 15-20, all six prompts visually inspected
• Availability of inherited split and absence of EpiPersona implementation: Official SALT-NLP/SynthesizeMe repository, MichaelR207 Hugging Face dataset metadata and official GitHub repository search inspected 2026-07-17
• Comprehensive audit of construct, statistics, minorities, artifacts, and ethics: reports/verification/article-380-epipersona-disentanglement-minority-generalization-statistics-artifacts-and-ethics-audit.json