SimPersona: Grounded Buyer Personas

• SimPersona is a framework that learns discrete buyer personas from raw clickstream data, enabling realistic simulation of diverse e-commerce shopper behaviors.
• It employs a behavior-aware VQ-VAE to compress 403-dimensional buyer features into 256 controllable and interpretable tokens for efficient agent conditioning.
• The framework outperforms traditional persona methods by achieving higher conversion alignment and robust performance across varied merchant storefronts.

SimPersona is a framework for grounding LLM-based e-commerce agents in empirically learned buyer heterogeneity rather than hand-crafted prompt personas. It learns discrete buyer types directly from historical clickstreams, maps those types to compact persona tokens in the LLM vocabulary, and uses merchant-specific mixtures over those learned types to simulate population-level buyer behavior on live storefronts. The framework is motivated by the claim that standard web agents collapse toward a single “average buyer” policy, while prompt-engineered personas are brittle, difficult to scale, context-inefficient, and weakly faithful to real buyer populations (Foumani et al., 14 May 2026).

1. Conceptual scope and problem formulation

SimPersona addresses grounded e-commerce simulation: agents navigate real storefronts, inspect DOM pages, search, browse, add to cart, and potentially proceed toward checkout. The central problem is not generic task completion but population-faithful behavioral simulation. Real merchants face heterogeneous buyers with different conversion tendencies, engagement depths, and browsing styles, so a single-policy agent is insufficient for modeling store traffic or conversion behavior (Foumani et al., 14 May 2026).

The framework therefore replaces natural-language persona prompting with learned discrete buyer personas. Each buyer is assigned to one codebook index $k \in \{1,\dots,K\}$, and each code is exposed to the LLM through a single special token. The paper emphasizes three reasons for discreteness: controllability, interpretability, and context efficiency. One token corresponds to one buyer type, codebook entries can be inspected behaviorally, and a full persona occupies one context position instead of a long textual description (Foumani et al., 14 May 2026).

This design sits within a broader critique of shallow persona conditioning. SCOPE reports that demographic-only personas are a structural bottleneck and that demographics explain only about $1.5\%$ of the variance in human response similarity, while demographic prompting often exaggerates demographic similarity (Venkit et al., 12 Jan 2026). SimPersona operationalizes a related objection in e-commerce: personas should be learned from behavioral traces rather than authored from coarse templates.

2. Clickstream grounding and buyer representation

SimPersona begins from de-identified raw platform logs containing page views, cart actions, searches, checkout actions, and related low-level events. These logs are enriched by joining them with the product catalog, collection directory, and search-query records. A single enrichment pass produces two outputs: buyer-level representations for persona discovery and executable multi-turn traces for agent training (Foumani et al., 14 May 2026).

For persona discovery, each buyer-shop pair is collapsed into a 403-dimensional vector. This representation combines 16 behavioral scalars with three channels of semantic product preference and 3 binary masks. The behavioral features are grouped into exposure volume, engagement, funnel, intent strength, and dollar value. Product preference is represented by averaging 768-dimensional product embeddings over viewed, carted, and purchased items, compressing each channel to 128 dimensions with PCA, and appending observation masks. The paper states the final structure explicitly as:

• $16$ scalar features
• $3 \times 128$ embedding dimensions
• $3$ masks for a total of $403$ dimensions (Foumani et al., 14 May 2026).

Component	Contents	Dimensionality
Behavioral scalars	Exposure, engagement, funnel, intent, dollar value	16
Product preference embeddings	Viewed, carted, purchased channels after PCA	384
Channel masks	Presence of viewed/carted/purchased channels	3

The intent-strength feature is a hand-defined composite based on add-to-cart, checkout-start, and purchase activity, with weights $w_{\text{atc}}=3$, $w_{\text{co}}=5$, and $w_{\text{pur}}=8$ (Foumani et al., 14 May 2026). Heavy-tailed counts and dollar values are transformed with $\log(1+x)$ and robust $1.5\%$0-scoring; bounded rates are Bayesian-smoothed via

$1.5\%$1

then logit-transformed and robustly standardized (Foumani et al., 14 May 2026).

The paper also defines five funnel strata reused in training and evaluation: purchasers, checkout abandoners, cart builders, window shoppers, and bouncers. For VQ-VAE experiments, the bouncer stratum is removed because it contains little behavioral signal (Foumani et al., 14 May 2026).

3. Behavior-aware VQ-VAE and discrete buyer-type induction

The core representation learner is a behavior-aware VQ-VAE over buyer vectors $1.5\%$2. The encoder maps $1.5\%$3 to a continuous latent $1.5\%$4, which is quantized to the nearest codebook entry: $1.5\%$5 The codebook is discrete, with $1.5\%$6 entries in $1.5\%$7, and gradients are propagated with the straight-through estimator (Foumani et al., 14 May 2026).

The training objective augments reconstruction with commitment, contrastive, and auxiliary behavioral supervision: $1.5\%$8 The commitment term is

$1.5\%$9

Reconstruction is group-aware: scalar features use MSE, while semantic product-preference blocks use cosine distance, and missing embedding channels are masked out (Foumani et al., 14 May 2026).

The contrastive loss is InfoNCE over encoder outputs: $16$0 Positive pairs are chosen through a three-stage gate: same highest funnel level, then top-$16$1 peers by product-embedding similarity, then top-$16$2 by Euclidean distance over exploration and engagement features; the reported settings are $16$3, $16$4, and $16$5 (Foumani et al., 14 May 2026).

Three auxiliary heads predict coarse engagement depth, exploration breadth, and purchase intensity from $16$6 using weighted cross-entropy. These heads are what make the VQ-VAE “behavior-aware”: the latent partition is regularized to preserve behavioral distinctions rather than only geometric reconstruction (Foumani et al., 14 May 2026).

Merchant-specific populations are represented as empirical mixtures over codebook entries. For store $16$7,

$16$8

This histogram is the sampling distribution used for population-level simulation. The paper also defines store-level reconstruction of funnel composition and aggregate behavior from token mixtures, making merchant heterogeneity an explicit object of simulation rather than an incidental by-product (Foumani et al., 14 May 2026).

4. Persona tokens and LLM-agent conditioning

After learning the discrete codebook, SimPersona extends the tokenizer of Qwen3-14B-Base with $16$9 special tokens: $3 \times 128$0 Each token corresponds one-to-one with a learned buyer type (Foumani et al., 14 May 2026).

The agent is trained with a two-stage supervised fine-tuning procedure. In the first stage, the Qwen3-14B backbone is frozen and only the 256 new persona-token embeddings are updated. Goals are made intent-neutral, such as “You are interested in skirts,” so that the token itself must carry the behavioral distinction. In the second stage, the full model is unfrozen and trained on richer traces with explicit session intent. Higher-funnel strata receive buy-oriented goals; window shoppers receive browse-oriented goals. The paper presents these signals as complementary: the persona token encodes the buyer’s general behavioral profile, while the goal text specifies the intent of a particular session (Foumani et al., 14 May 2026).

Training traces are reconstructed from enriched sessions. For each session, an LLM rewrites the event sequence into a natural-language shopping goal, another agent replays that goal on the live storefront, and successful trajectories are filtered by an independent LLM judge. Each trace includes a system prompt, a user prompt containing the inferred session goal and the learned persona token, a progress log acting as memory, and the current page DOM snapshot; the assistant target is a structured JSON action plus a reasoning trace (Foumani et al., 14 May 2026).

At inference time, assigning a buyer to a persona is cheap: one encoder forward pass produces $3 \times 128$1, nearest-code quantization yields a codebook index, and the corresponding token is inserted into the agent prompt. For merchant-level simulation, SimPersona samples persona indices from the store’s empirical distribution $3 \times 128$2, instantiates token-conditioned agents, and runs them on the merchant’s live storefront without retraining or store-specific prompt engineering (Foumani et al., 14 May 2026).

5. Empirical performance and behavioral findings

The paper reports evaluation on $3 \times 128$3M buyers across $3 \times 128$4 held-out live storefronts and presents three layers of evidence: clustering quality, population-distribution realism, and agent-level behavioral transfer (Foumani et al., 14 May 2026).

For buyer-type induction, the behavior-aware VQ-VAE outperforms MiniBatch $3 \times 128$5-means trained on the same buyer vectors. Reported gains include:

• stratum purity: $3 \times 128$6 vs. $3 \times 128$7
• incompatible mixing: $3 \times 128$8 codes vs. $3 \times 128$9
• engagement incoherence: $3$0 vs. $3$1
• pairwise cosine similarity: $3$2 vs. $3$3
• Calinski–Harabasz index: $3$4 vs. $3$5 (Foumani et al., 14 May 2026)

For merchant-level realism, the store-level stratum distribution reconstruction achieves mean Jensen–Shannon divergence $3$6 and median $3$7 across the $3$8 evaluation stores, with $3$9 of $403$0 stores below $403$1. Aggregate continuous features reconstructed from token mixtures reach $403$2 values of $403$3 for intent strength, $403$4 for add-to-cart rate, $403$5 for checkout rate, and $403$6 for browse-only rate (Foumani et al., 14 May 2026).

The main simulation metric is stratified Action Rate Alignment. The paper defines: $403$7

$403$8

$403$9

The headline result is a stratified ARA of $w_{\text{atc}}=3$0, reported in the abstract as $w_{\text{atc}}=3$1 conversion-rate alignment with real buyers. Correct token-buyer matches substantially outperform mismatch baselines: $w_{\text{atc}}=3$2 for correct pairings versus $w_{\text{atc}}=3$3 for all-mismatch and $w_{\text{atc}}=3$4 for random mismatch (Foumani et al., 14 May 2026).

The learned personas also transfer interpretable behavioral variation into the agent. Tokens binned high on purchase intensity produce much more conversion than low-purchase tokens under the same broad buy-oriented intent. Mean purchase score per simulated session rises from $w_{\text{atc}}=3$5 to $w_{\text{atc}}=3$6 to $w_{\text{atc}}=3$7 across low, medium, and high purchase bins. Funnel progression by purchase bin is likewise strongly ordered:

• add-to-cart rate: $w_{\text{atc}}=3$8 / $w_{\text{atc}}=3$9 / $w_{\text{co}}=5$0
• checkout rate: $w_{\text{co}}=5$1 / $w_{\text{co}}=5$2 / $w_{\text{co}}=5$3 (Foumani et al., 14 May 2026)

On goal-oriented shopping tasks, SimPersona also outperforms GPT-OSS-120B, a baseline with about $w_{\text{co}}=5$4 more parameters. Reported results include:

• cart tasks $w_{\text{co}}=5$5 added to cart: $w_{\text{co}}=5$6 vs. $w_{\text{co}}=5$7
• checkout tasks $w_{\text{co}}=5$8 reached checkout: $w_{\text{co}}=5$9 vs. $w_{\text{pur}}=8$0
• purchase tasks $w_{\text{pur}}=8$1 reached checkout: $w_{\text{pur}}=8$2 vs. $w_{\text{pur}}=8$3 (Foumani et al., 14 May 2026)

Ablations are integral to the paper’s argument. Two-stage training does not primarily raise mean goal-reached rate, but it sharply improves tail robustness across stores: maximum per-shop error rate falls from $w_{\text{pur}}=8$4 under single-stage training to $w_{\text{pur}}=8$5, and stores with error above $w_{\text{pur}}=8$6 drop from $w_{\text{pur}}=8$7 to $w_{\text{pur}}=8$8. Removing persona tokens triples simulation crashes, from $w_{\text{pur}}=8$9 with tokens to $\log(1+x)$0 without them, and increases browser-level StagehandTargetClosed errors by a factor of $\log(1+x)$1 (Foumani et al., 14 May 2026).

6. Position within persona simulation and open issues

SimPersona differs sharply from narrative persona frameworks. Systems such as Human Simulacra build life-story-grounded simulacra from synthetic biographies and multi-agent cognitive prompting (Xie et al., 2024), while SimsChat emphasizes customizable characters built from career, aspiration, traits, and skills and then expanded through personal and social profiles (Yang et al., 2024). SimPersona instead grounds personas in raw clickstream regularities and product-interest patterns, then compresses them into discrete control tokens (Foumani et al., 14 May 2026).

It also differs from socially grounded orchestration frameworks such as SOTOPIA-S4, which provides infrastructure for persona-driven, role-conditioned, multi-turn social simulation with characters, scenarios, relationships, and evaluation metrics (Zhou et al., 19 Apr 2025). SimPersona is narrower in domain but more tightly coupled to observed behavioral data and merchant-specific population mixtures.

Outside e-commerce, several contemporaneous lines of work address problems adjacent to SimPersona’s. SCOPE argues that persona quality depends on sociopsychological structure rather than demographic templates (Venkit et al., 12 Jan 2026). “Population-Aligned Persona Generation for LLM-based Social Simulation” aligns persona sets to target psychometric distributions using importance sampling and optimal transport (Hu et al., 12 Sep 2025). SemaPop conditions synthetic population generation on LLM-derived semantic personas while regularizing toward target marginals (Qin et al., 12 Feb 2026). Eval4Sim proposes human-reference evaluation of persona simulation through adherence, consistency, and naturalness rather than scalar LLM-judge scores (Bao et al., 3 Mar 2026). These works indicate that persona learning, calibration, and evaluation are increasingly treated as distinct technical problems rather than prompt-writing exercises.

The main limitations of SimPersona are explicit. Exploration behavior transfers weakly: exploration scores are nearly flat, with Cohen’s $\log(1+x)$2 and non-significant tests (Foumani et al., 14 May 2026). The downstream agent only sees DOM structure, so visual cues that affect shopping behavior are absent. Newly opened stores with no clickstream history are not supported, since the framework depends on historical buyer behavior. The paper also notes that the current downstream pipeline uses only the discrete token assignment, not the learned continuous VQ-VAE embeddings, and does not go deeply into privacy, fairness, or temporal drift (Foumani et al., 14 May 2026).

Taken together, SimPersona defines a specific technical answer to persona simulation in commerce: learn buyer heterogeneity from behavioral traces, discretize it into interpretable types, encode those types as efficient LLM control tokens, and preserve merchant-specific mixtures during simulation. Its broader significance lies in showing that persona conditioning can be moved from descriptive prompting to data-driven latent induction, with explicit population reconstruction and measurable alignment to real-world behavior (Foumani et al., 14 May 2026).