acceptodds
Under review as a conference paper at ICLR 2027

SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-commerce Agents

38 views

Abstract

Large language model (LLM)-based web agents can navigate live storefronts, yet often collapse to a single average buyer policy, failing to capture the heterogeneous, distributional nature of real buyer populations. Existing personalization relies on hand-crafted prompt-based personas that are brittle, hard to scale, and unable to faithfully represent population-level behavior. We introduce SimPersona, a framework that learns discrete buyer types from historical traffic and exposes them to LLM agents as compact persona tokens. From raw clickstreams, a behavior-aware vector-quantized variational autoencoder (VQ-VAE) induces a discrete buyer-type space capturing the statistical structure of real buyer behavior and merchant-specific population distributions. Each buyer type maps to a dedicated persona token, and the agent is fine-tuned with these tokens on real browsing traces. At inference, each synthetic buyer is assigned a token with a single encoder forward pass, without any retraining or store-specific prompting, and population-level simulation samples types from each merchant's empirical codebook distribution. Evaluated on 8.37M buyers across 42 held-out live storefronts, SimPersona achieves 78% conversion-rate alignment with real buyers, exhibits interpretable behavioral variation across buyer types, and outperforms larger-parameter baselines on goal-oriented shopping tasks. We release an open-source data pipeline that converts raw e-commerce event logs into buyer representations and agent-training traces.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, LLM-based web agents, Agentic user modeling, Agent simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fDzXOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a method SimPersona for simulating users with LLM agents in e-commerce applications. The central question is, how to simulate users that realistically mimic the real-world user distribution, which existing text-based persona grounding methods fail in terms of distribution matching and stability. The methodology involves learning user-shop codes with VQ-VAE, and condition the agent & model on the learned user representation codes. The authors propose various algorithmic engineering efforts to fit the framework better to the goal of online applications, such as feature engineering, defining extra behavior axes, etc.

The authors experiment on their own datasets in terms of 4 metrics: how the learned codes maintain user-shop semantic meanings, how the learned codes help mimic the per-shop conversion rate distribution, how the learned codes help shape individual agentic behaviors in terms of purchase/add-to-cart, etc., and how the learned codes help agents accurately complete shopping tasks, where the proposed SimPersona show good potentials in real-world applications and outperform text-based persona methods.

Critical strengths
  1. Real-world problem. The authors tackle a real-world problem: designing a user simulator that not only shows desired behaviors, but also shows the behaviors with a similar distribution as the online users. More importantly, this paper aims to address the problem at the industrial level, aiming for online applications. This makes the problem more complex with considerably more design considerations.

  2. Practical data engineering with application-specific insights. I appreciate the paper not only aiming to tackle a real-world problem, but also tackling it in a rigorous and application-oriented manner. For example, the feature engineering of the scalar ones, the funnel, the contrastive learning objective with careful design of positive and negative samples, the three behavioral axes of engage, explore and purchase, the two-step fine-tuning with different goals, etc. These designs are not fancy but are critical in real-world LLM-based algorithm engineering, and I appreciate these efforts.

  3. Evaluation is done with real-world application insights. First, in terms of evaluating the code, the authors apply metrics related to the funnel and behavior axes, which provide far more semantic meanings than common clustering metrics. The authors also evaluate the trained models in terms of reconstructing actual user behavior distributions, and most importantly, on a sufficiently large and unseen test dataset.

Critical weaknesses
  1. Lacking illustrative examples throughout the paper. This paper is dense in terms of engineering insights but notably missing on illustrative examples (i.e. concrete, or at least very similar in terms of detailedness instead of a template). Therefore, I think this paper may benefit from some illustrative examples to showcase the generated data. For example, for Section 3.4, it would be good to show some concrete tasks; for the goals (intent-neutral and non-neutral ones), some examples would also be appropriate.

  2. Large-scale results in Sec. 3.2 and 3.3 lack more concrete metrics, such as prices. Intuitively, I think the purchase/add-to-cart rates would be more informative if the actual item information is presented in more details, such as prices, product category, etc., but no such information is present throughout the two sections. This is especially concerning, because in the input channel, view/cart/purchase embeddings, as well as scalar dollar values are present, but no evaluations are present in terms of the exact products the agent interacted with.

  3. The metrics about the three behavioral axes in Section 3.1 are somewhat self-referential, mostly because k-means never takes this in mind, and the VQ-VAE is trained with this objective. This is a minor point but can be enhanced with some more discussion.

  4. In the paper (Section 2.3), the authors train a Qwen3-14B model for the agent, but in Table 4, the authors compare with a Qwen3.5-9B model trained with textual persona grounding. This leads to a potentially unfair comparison and should be justified.

  5. The "average buyer" issue may still exist given only 256 codes. Since the authors experiment with at most 8.37M users, but only 256 codes are present, each code may carry on average 30K users and a median of 2K users (reported in the ethics). Therefore, it is very likely that the "average user effect" may still exist given this scale. Therefore, a discussion on the within-token variance (e.g. depending on goals and shops) may be appropriate to demonstrate that the codes won't become anchors to bound the agent's behavior.

Questions for authors

Please see the 5 weaknesses above.

8 comments
AuthorsOpenReview ↗

Answer to Weakness 1

W1- We thank the reviewer for this helpful suggestion. We agree that the paper is dense and benefits from concrete examples of the objects used in the pipeline. Below we provide representative examples of: (i) intent prompts, (ii) deterministic instruction-following tasks, (iii) generated rollouts, and (iv) behavioral-simulation trajectories. Store names, product names, URLs, prices, and other identifying details are anonymized.

A. Intent/goal examples. The behavioral-simulation goals are short intent prompts. They specify the product category and session-level intent, but they do not prescribe a sequence of browser actions. The persona token is supplied separately from the goal text:

Goal conditionExample instruction
Buy-oriented“You are here to buy [product].”
Browse-oriented“You are here to browse [product].”
Outcome-neutral“You are interested in [product].”

The buy-oriented and browse-oriented prompts are used in the standard persona-conditioned simulation setting. The outcome-neutral prompt is used in the neutral-intent ablation in Appendix J.1. It keeps the category of interest but removes the explicit conversion direction. Thus, in the neutral case, the agent is told the topic of interest but is not told whether it should browse, add to cart, proceed to checkout, or stop.

This distinction is important: the goal describes the \emph{session intent}, while the persona token describes the \emph{buyer profile}. SimPersona conditions on both.

B. Section 3.4 task and rollout examples: The deterministic tasks in Section 3.4 are different from the behavioral-simulation goals above. They explicitly prescribe product-level browser actions and stopping conditions. They are used to test whether the agent can execute multi-step shopping instructions on live storefronts.

For example, Section 3.4 includes tasks of the following form:

Cart task. From the homepage at [store], use the search bar to search for [product]. View the product page. Add it to the cart, then end the session.

Multi-product checkout-abandonment task. From the homepage at [store], use the search bar to search for [product 1] and view the product page. Then search for [product 2] and view the product page. Then search for [product 3] and view the product page. Add it to the cart, then start checkout but do not complete checkout; abandon checkout.

Product-and-size-specific checkout task. From the homepage at [store], use the search bar to search for [product name]. View the product page for [product name] in [color/variant], size [size]. Add it to the cart. View the product page again. Then start checkout but do not complete checkout; abandon checkout.

These examples show that Section 3.4 evaluates instruction-following ability: the task tells the agent what product to find, what action to perform, and when to stop. This is different from the behavioral-simulation setting, where the agent receives only a persona token and a broad session intent.

A generated rollout for a cart-style task has the following structure. In this example, the agent receives a learned persona token and a deterministic cart task:

Task. From the homepage at [store A], use the search bar to search for [product A]. View the [product A] product page, then add it to the cart, then end the session.

Persona input. <|special_token_51|>

StepGenerated action
1Click the search link to open the search bar.
2Type [product A] into the search bar.
3Click the [product A] result to open the product page.
4Click “Add to cart” to add the product to the cart.
5End the session after completing the task.

To make the prompt format concrete, below is one structurally complete agent-call example from the same episode, with the long accessibility tree abbreviated. This is the fourth decision in the trajectory, after the agent has already searched for the product and opened the product page.

AuthorsOpenReview ↗

Continue Answer Weakness 1

W1- Example agent call

System message

You are simulating the behavior of a potential human buyer in an ecommerce website. You must answer with a JSON object matching the schema with fields `shouldEnd`, `action`, and `reasoning`. The action contains a `nodeSelector`, `method`, `description`, and `arguments`.

User message.

You are simulating the behavior of a human buyer in an ecommerce website.

You have the following main goal:
<MainGoal>
From the homepage at [store], use the search bar to search for [product].
View the [product] product page, then add it to the cart, then end the session.
</MainGoal>

Your browsing and purchase behavior is highly influenced by your buyer profile.

<BuyerProfile>
<|special_token_51|>
</BuyerProfile>

You log your progress and thoughts along the way. Here is what you have logged so far:
<ProgressLog>
- Started by navigating to [store].
- I clicked the search link to open the search bar.
- I typed [product] into the search bar.
- I clicked the [product] result from the search suggestions.
- I landed on the [product] product page.
</ProgressLog>

<current_page_info>
You are currently on [store]/products/[product].
Here is the accessibility tree of the page, identifying each element with an ID:
<accessibility_tree>
[anonymized ]
</accessibility_tree>
</current_page_info>

<task>
1. Reason about the main goal you are trying to achieve.
2. Review the browsing session <ProgressLog>.
3. Review the current URL and accessibility tree.
4. Generate your output according to the JSON schema provided by the system.
</task>

Assistant response.

{
  "shouldEnd": false,
  "action": {
    "nodeSelector": "0-3136",
    "method": "click",
    "description": "Click on the ADD TO CART button to add [product] to the shopping cart.",
    "arguments": []
  },
  "reasoning": "I have successfully reached the product page for [product]. Now, I am adding it to my cart as the next step in my goal before I finish my session."
}

The next progress log records that the click succeeded and that the cart contains one item. The following model call then terminates the session:

{
  "shouldEnd": true,
  "action": null,
  "reasoning": "I have successfully searched for [product], viewed its product page, and added it to my cart. Since I have completed all the steps of my goal, I am now ending the session."
}

The episode contains five agent decisions: four browser actions and one terminal decision. The product page is viewed, one item is added to the cart, and the evaluator marks the task as successful. This example illustrates the format of the generated browser actions used in Section 3.4; the aggregate success rates are reported in the quantitative evaluation.

C. Behavioral simulation example. The examples in Section B show deterministic instruction-following tasks, where the agent is explicitly told what product to search for and what endpoint to reach. Behavioral simulation is different. In this setting, the agent receives only a persona token and a short session goal, not a step-by-step instruction. The agent then decides how to navigate the store.

The following example is a recorded SimPersona behavioral-simulation rollout from the main analysis run. The goal is category-level, rather than a deterministic instruction:

Goal: “You are here to buy [category].”
Persona token: <|special_token_183|>

The agent is not told to search for a specific product or to follow a fixed action sequence. Nevertheless, it generates the following trajectory:

StepGenerated action
1Open the search bar.
2Type [product] into the search bar.
3Open the [product] product page.
4Add the product to cart.
5Click checkout.
6End the session after reaching checkout.

The recorded outcome is: six decisions, five successful browser actions, one cart item, cart value [price], checkout reached, and termination label GOAL_REACHED. Reaching checkout does not imply completed payment; the live-store simulations stop before any real financial transaction.

A representative action from this rollout is:

{
  "shouldEnd": false,
  "action": {
    "nodeSelector": "[node_id]",
    "method": "click",
    "description": "Click on the Add to cart button to add [product] to my shopping cart.",
    "arguments": []
  },
  "reasoning": "I have found the [product] I was looking for, and now I am adding it to my cart so I can proceed to the checkout as planned."
}

This example illustrates the key distinction between instruction following and behavioral simulation. In Section 3.4, the task tells the agent exactly which product to search for and when to stop. In behavioral simulation, the agent receives a persona token and a short intent, then independently chooses the search query, product page, cart action, checkout action, and stopping point. We added these illustrative examples to Appendix L.

AuthorsOpenReview ↗

Answer Weakness 2

W2- We thank the reviewer for this helpful suggestion. We agree that adding product- and price-level diagnostics makes the large-scale behavioral results more concrete, so we performed the requested analysis.

First, we checked the most direct question: when SimPersona adds to cart or reaches checkout, is the product from the intended product family? Our analysis in the table below shows that category agreement is above 98% for both cart and checkout. “Matched” means at least one carted product belongs to the requested product family; “classifiable” means the product category could be identified.

EventMatched / classifiableMatch rateMatched / all event episodes
Cart5,971 / 6,08098.21%90.74%
Checkout3,410 / 3,46898.33%90.93%

Thus, SimPersona is not matching conversion rates by carting arbitrary merchandise; the products are usually in the intended product family.

Second, we asked whether SimPersona carts products from similar price ranges as humans. Because shops differ substantially in currency, product vertical, and catalog scale, we compare prices within each shop rather than using raw global prices. We rank products within each shop by price and divide them into four local price tiers:

Within-shop product-price percentileHuman cart shareSimPersona cart shareAgent − human
0–2522.17%19.58%-2.59 pp
25–5025.75%21.37%-4.38 pp
50–7528.86%33.07%+4.21 pp
75–10023.22%25.98%+2.76 pp

The aggregate differences are small: SimPersona is within roughly 2.6--4.4 percentage points of the human distribution in every price tier. This indicates that SimPersona covers the same broad within-shop price range as human carts. This percentile-based comparison is important because raw dollar values are highly heterogeneous. Across the catalog products used to define the buckets, the standard deviation increases from 148.52 dollars in the lowest quartile to 7,477.72 dollars in the highest quartile. Thus, raw dollar comparisons can be dominated by a small number of very high-priced catalog items, while within-shop percentile tiers provide a more stable comparison across heterogeneous storefronts. Under this metric, the mean human--SimPersona price-tier overlap across shops is 62.17%, where 100% would mean identical price-tier distributions.

Third, we compared product-category distributions beyond the simple category-match check. At the fine-grained assigned catalog-category level, SimPersona achieves 60.46% overlap with human cart-category distributions and 63.44% overlap with human view-category distributions. When categories are grouped into broader families using the first two taxonomy levels, overlap increases to 78.00% for cart and 79.31% for view:

Category definitionActionMean overlap
Fine-grained assigned catalog categoryCart60.46%
Fine-grained assigned catalog categoryView63.44%
Coarse category: first two taxonomy levelsCart78.00%
Coarse category: first two taxonomy levelsView79.31%

This shows that SimPersona's product interactions are close to human interactions at the broader category-family level, while the finer taxonomy naturally produces lower overlap because it distinguishes more specific product types.

Finally, we compared product semantics using archived product embeddings. For each shop, we compute a centroid of products carted by humans and measure how close SimPersona-carted products are to that centroid. SimPersona carted products have high similarity to the human-carted product centroid 0.858.

These analyzes are added to Appendix J.4 of the revised manuscript.

Reviewer_7xtEOpenReview ↗
Rating
4(+2)
Confidence
4(+1)
Informativeness
3(+1)
Presentation
3
Soundness
3(+2)
Read review
Paper summary

This paper introduces SimPersona, a framework for learning discrete buyer personas from e-commerce clickstreams and using them to condition an LLM shopping agent. It constructs 403-dimensional buyer-shop representations from behavioral statistics and product embeddings, then trains a behavior-aware VQ-VAE with reconstruction, contrastive, and auxiliary classification objectives to assign buyers to 256 discrete codes. These codes are added to the vocabulary of a Qwen3-14B agent as persona tokens and grounded through a two-stage fine-tuning procedure. The paper evaluates clustering quality on 39 training shops, conversion-rate alignment against a reference population of 8.37M buyers from 42 held-out shops, behavioral separation between persona bins, and instruction following on 70 live-store tasks.

Critical strengths

The paper addresses an important and practically relevant gap between offline user representation learning and closed-loop agent simulation. Representing personas as single learned tokens is computationally attractive, and the attempt to model both persona types and their merchant-specific mixture is more scalable than manually written profiles. The shop-disjoint evaluation, broad range of storefronts, detailed pipeline description, codebook-size analysis, and comparison of one- versus two-stage fine-tuning are valuable. The paper is also unusually clear about several observed limitations, including the failure to transfer the exploration dimension and the lack of visual perception.

Critical weaknesses
  1. The main conversion-fidelity evidence is not sufficiently independent on the signals used to construct and condition the model. The VQ-VAE input already includes add-to-cart, checkout, browse-only, and intent-related features, while evaluation supplies an intent derived from the realized funnel stratum: strata A-C receive a buy intent and stratum D receives a browse intent. The correct-versus-mismatch comparison changes persona pairings but does not compare the same trained agent under identical intents with its token removed or shuffled. Appendix J.1 removes tokens only under neutral intents and evaluates crashes and engagement rather than the main conversion-alignment metric. Moreover, Appendix B reconstructs funnel strata and behavioral features from token frequencies using token profiles and centroids computed from “all buyers” assigned to each token. These statistics are not restricted to the training shops and the reconstructed quantities substantially overlap with the representation inputs. Consequently, the reported 0.787 and 0.927 alignment scores do not cleanly demonstrate that the agent learned held-out buyer behavior beyond signals already encoded in the persona assignment, intent, and reconstruction procedure.

    • The downstream policy is not trained on actual human action trajectories. GPT-5 converts human event histories into goals, Gemini 3 Flash generates the action trajectories by replaying those goals, and GPT-5 filters the generated traces. Thus the action supervision reflects another LLM agent’s policy, while the human data supply aggregate features, inferred goals, and outcomes. The paper does not directly evaluate whether held-out human action sequences, navigation choices, abandonment points, or product-view transitions are reproduced. The comparisons also do not isolate the persona representation: Section 3.4 compares Qwen3-14B SimPersona against a text-persona model with a different Qwen version and size as well as different page access, while Appendix J.2 uses Gemini 3 Flash and switches to a coarser shop-level metric. In addition, the purchase task result in Table 4 measures only whether checkout was reached, not whether a purchase was completed. These design differences prevent attributing the reported downstream gains specifically to learned persona tokens or interpreting them as evidence of human-level behavioral fidelity.
Questions for authors
  1. Can you repeat the Section 3.2 evaluation with outcome-neutral goals and the same trained backbone under matched-token, shuffled-token, and no-token conditions, while reporting the raw human and agent ATC/purchase rates and the number of agent episodes in each shop-token-stratum cell?

  2. Were P(stratum∣k)P(\text{stratum}\mid k) and the token centroids in Appendix B estimated exclusively from the 39 training shops? If not, can the authors recompute the analysis using train-only token statistics and evaluate reconstruction of behavioral targets that were not inputs or auxiliary labels of the VQ-VAE?

  3. Can the token and text-persona approaches be compared using the same Qwen3-14B backbone, training trajectories, observation modality, and evaluation protocol, and can the evaluation include actual purchase completion and similarity to held-out human action sequences?

10 comments
AuthorsOpenReview ↗

Answer to Question 1

Q1-We ran the requested neutral-goal analysis, but we emphasize that this setting should not be interpreted as the primary ablation of the persona token. Replacing the stratum-derived goal with a neutral goal and then removing or shuffling the token changes two parts of the policy input at once. SimPersona is trained as a joint conditional policy, p(trajectory∣token,intent,storefront)p(\text{trajectory} \mid \text{token}, \text{intent}, \text{storefront}) where the intent specifies the coarse session objective and the token specifies the buyer profile, behavioral style, and intensity. As stated in Section 2.3, these two signals are complementary: the goal specifies what the session is trying to accomplish, while the token specifies how the buyer tends to behave.

This separation between who the user is and what the user is trying to do in the current episode is not specific to SimPersona; it is common in interactive user/agent simulation. For example, SOTOPIA [1], which is published in the same conference, samples both character profiles and private goals for each episode, while ShopperBench [2] constructs persona-conditioned shopping tasks where agents must satisfy both persona preferences and shopping goals. In all of these settings, persona controls style, preferences, or behavioral tendency, while the goal or intent controls the episode objective.

The same persona can therefore behave differently under different goals: an exploratory buyer may browse and then check out under a buy intent, but browse without entering the funnel under a browse intent. For this reason, the neutral-goal setting is expected to lose part of the conversion calibration: it removes the outcome signal that tells the agent how the persona should be expressed in the current session. The cleaner ablation is to keep the stratum-derived intent fixed and change only the token. We therefore report both: the requested neutral-goal results for completeness, and then the fixed-intent no-token and shuffled-token ablations, which directly isolate the token under the intended SimPersona setup.

We first report the requested neutral-goal results; the matched-token and no-token arms each contain 50,400 attempted episodes across 42 held-out shops, paired by buyer, shop, theme, exact neutral goal, and intended token. For the shuffled-token condition, we replace the intended token X with a different token Y and evaluate the generated behavior against the human cohort associated with X. Before interpreting conversion rates, the table below shows one of the reasons why a direct conversion-rate comparison in this neutral setting is not a clean conversion-calibration comparison. Once the intent is neutralized, removing the token changes both the behavioral prior and the execution stability of the agent. The no-token agent fails much more often, so its observed ATC/checkout rates are computed only over a different surviving subset of episodes. Thus, raw conversion-rate comparisons are confounded by differential failure. This is why Appendix J.1 first reported stability and engagement before interpreting conversion behavior. In the neutral-goal runs, hard errors increase from 5.26% with the matched token to 15.60% with no token, driven mainly by browser-target-closure errors.

OutcomeMatched tokenNo token
Attempted episodes50,40050,400
ERROR2,650 (5.26%)7,862 (15.60%)
Browser target closed1,430 (2.84%)6,579 (13.05%)

Thus, the neutral no-token condition is not only a different behavioral condition; it is also a less stable execution condition. We report the requested raw rates below, but their role is diagnostic: they show what happens when the outcome intent is removed, not whether the token alone should reproduce the full Section 3.2 calibration.

ATC table:

StratumHumanMatched tokenShuffled proxy ATCNo token
A69.35%44.94%38.16%40.10%
B71.72%44.46%38.22%44.13%
C74.40%37.02%38.40%32.34%
D1.12%36.54%38.45%33.39%
Stratified A–D54.15%40.74%38.31%37.49%

Checkout table

StratumHumanMatched tokenShuffled proxyNo token
A63.95%22.84%18.85%20.54%
B52.59%22.28%18.89%23.51%
C23.62%17.23%19.03%15.17%
D0.82%18.10%19.04%16.42%
Stratified A–D35.24%20.11%18.96%18.91%

The above raw rates show the expected pattern under neutral goals: once the buy/browse signal is removed, none of the token conditions recovers the full human conversion profile.

[1] Zhou, X., et al. Sotopia: Interactive evaluation for social intelligence in language agents. In International Conference on Learning Representations.

[2] Ling, Y., et al. ShopperBench: A benchmark for personalized shopping with persona-guided simulation. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics

AuthorsOpenReview ↗

Continue, Response to Question 1

Q1- The reviewer also asked for the number of agent episodes supporting each shop-token-stratum comparison. The full table has 444 (shop, token, stratum) cells; we summarize counts by stratum below. “Target cells” is the number of evaluated groups, “paired attempts/arm” is the number of attempted simulations per condition, and the last two columns show successful episodes.

StratumTarget cellsPaired attempts/armMatched successesNo-token successes
A501,7291,3301,278
B441,8661,6031,484
C542,7162,0741,910
D29643,23434,54333,103
Total44449,54539,55037,775

The table also illustrates the stability gap: the no-token arm has fewer successful episodes under the same attempted counts. The takeaway from the neutral-goal analysis is not that the token is irrelevant; rather, it shows that the token is not meant to replace session intent. Under a generic product-interest goal, the model lacks the coarse outcome direction needed to decide whether the same persona should browse, cart, or check out, so conversion ARA is close across matched-token, shuffled-token, and no-token conditions. However, the token still affects navigation: matched-token is directionally strongest in raw ATC/checkout rates, and removing the token substantially increases hard failures. Thus, under neutral goals the token acts as a navigation and engagement prior, while full conversion calibration requires token + session intent.

This is consistent with the intended role of persona tokens: they are learned from clickstreams and should preserve buyer tendencies such as browsing depth, cart propensity, and checkout propensity. The key question is whether this compressed representation improves population behavior when intent and storefront context are fixed. Therefore, the cleaner token-isolation test is to keep the intended conditioning setup intact and change only the persona signal. We ran this fixed-intent ablation and added it to Appendix J.3. When the token is removed while intent is held fixed, stratified ARA drops from 78.65% to 70.94%:

StratumSimPersonaNo token
A: Purchasers72.4%72.4%
B: Checkout abandoners74.8%77.7%
C: Cart builders72.1%68.9%
D: Window shoppers95.4%64.6%
Stratified average78.65%70.94%

This shows that intent alone is useful but too coarse. For low-intent buyers, especially window shoppers, removing the token causes severe over-conversion: with browse intent, 46.14% of no-token agents add to cart and 24.01% start checkout, while SimPersona reduces these to 0.90% and 0.33%.

We also keep the same intent but assign an incorrect token. In this setting, ARA falls to the mismatch range:

Stratum of XShuffled token: Y → XAll mismatch1-R
A: Purchasers45.63%65.51%65.45%
B: Checkout abandoners48.92%66.14%66.20%
C: Cart builders57.27%69.01%69.06%
D: Window shoppers74.81%66.76%66.75%
Stratified average56.66%66.85%66.86%

The correct SimPersona pairing reaches 78.65% ARA, while assigning the wrong token Y to target population X drops ARA to 56.66%, close to the original mismatch baselines. This is informative because the intent is unchanged; only the behavioral code is wrong. Once token Y drives the agent, the trajectory no longer matches target population X and instead moves toward other token populations. The other two columns are a sanity check: after shuffling, trajectories are often closer to mismatched/random token groups than to the original target group. Thus, calibration is not produced by intent alone; the correct alignment between token, intent, and target population is necessary.

This observation is further supported by Appendix J.2. Prompt-based baselines over-convert, especially at add-to-cart, while SimPersona remains close to human funnel rates. To ensure this is not due to the different SimGym model, we added a Qwen3-14B text-persona comparator. It still over-carts, while SimPersona remains calibrated:

MethodATC: agent / humanCheckout: agent / humanARA
Qwen3-14B text-persona0.570 / 0.1280.063 / 0.0780.728
SimPersona0.128 / 0.1280.072 / 0.0780.927
AuthorsOpenReview ↗

Answer to Question 2

Q2- We thank the reviewer for this comment as it adds a useful stricter generalization test to Appendix B. Appendix B is designed as a population-summarization test. After buyers in a shop are assigned to persona tokens, the question is whether the resulting token mixture summarizes that shop's real buyer population. In other words, if millions of buyers are compressed into a distribution over learned persona tokens, do we still recover the shop's funnel composition and aggregate behavioral profile? This is why Appendix B defines each shop's empirical token distribution p^s(k)\hat{p}_s(k) and uses it to reconstruct the shop's stratum mixture and aggregate behavioral features. Figure 6 evaluates recovery of the stratum distribution, and Figure 7 evaluates reconstruction of aggregate behavioral features. For this purpose, the natural comparison is within the same shop population: compute the real buyer mix of a shop, compute the tokens assigned to buyers in that shop, and ask whether the token mixture reconstructs the real mix.

However, the reviewer’s question adds a stricter analysis to our evaluation. It asks whether the meaning of each token can be estimated only from the training shops, frozen, and then used to reconstruct the 42 evaluation shops. We agree that this is a useful generalization test, so we recomputed Appendix B under this train-only protocol. The result is essentially unchanged. For stratum recovery, using train-only token statistics gives mean JS 0.052 and median JS 0.045, compared with 0.054 and 0.045 in the Appendix B. The number of stores below JS 0.10 is also nearly identical: 37/42 with train-only statistics versus 36/42 in the submitted version:

MetricAppendix BTrain-only token statistics
Mean JS0.0540.052
Median JS0.0450.045
Stores below 0.10 JS36/4237/42
Stores >90% stratum D1414

This shows that the Appendix B result does not depend on estimating token profiles from the same population being reconstructed. Even when token meanings are estimated only from training shops, the token mixture still reconstructs the evaluation-shop funnel composition with the same accuracy. The feature-reconstruction results are also stable. Using train-only centroids, the token mixture continues to recover the main conversion-related behavioral axes:

FeatureAppendix B R²Train-only R²
Intent strength0.9580.949
Checkout rate0.8970.947
ATC rate0.9500.892
Browse-only rate0.8730.764

These results show that the Appendix~B conclusion does not depend on estimating token statistics from the evaluation shops: train-only token profiles recover nearly the same funnel and conversion-related structure. Thus, the learned codes have stable behavioral meaning across shops.

Regarding the request for targets that were not inputs or auxiliary labels, a representation model cannot be expected to reconstruct arbitrary variables that are neither observed nor predictable from its clickstream inputs. The meaningful test is whether the learned token mixture preserves behavioral structure relevant to simulation, and whether those tokens then affect generated agent behavior. However, if the reviewer has specific features in mind, we will add those analysis.

Reviewer_K1BhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies persona conditioning for LLM shopping agents and proposes SimPersona. A behavior-aware VQ-VAE maps each buyer’s clickstream history to one of 256 discrete codes, and each code becomes a special token in Qwen3-14B. A two-stage SFT grounds the tokens in agent behavior, and each store’s token frequencies serve as its buyer-type distribution.

Critical strengths
  1. The problem is practical and under-explored. Prompt-based personas are brittle and say nothing about how many buyers of each type a store has.
  2. Evaluation runs on 42 live production storefronts (8.37M buyers) that share no shops with training, including unseen verticals and theme changes.
  3. Conditioning is cheap: one encoder forward pass assigns each buyer a single token. The data pipeline is open source.
Critical weaknesses
  1. The persona token’s contribution is not isolated. At evaluation, each agent also gets a buy or browse intent derived from the real buyer’s funnel stratum. The mismatch baselines change only the reference group; the agent’s input stays fixed. Under neutral intents, token-conditioned agents reach checkout in 15.5% of runs versus 7.8% for real buyers (App. J.1, Table 11), so calibration seems to come mainly from the intent.
  2. The persona-discovery evidence is partly circular. The contrastive funnel gate and the purchase head effectively supervise the codebook with the strata, so recovering store-level stratum mixes (JS 0.054) is expected. Table 1 also scores k-means on the VQ-VAE’s own training targets, and there is no downstream comparison between VQ-VAE tokens and k-means tokens.
  3. The agent baselines are confounded. The Table 4 tasks spell out every action step by step, so persona information cannot affect success. The Qwen3.5-9B baseline also differs from SimPersona in size, version, and modality. GPT-OSS-120B activates only about 5.1B parameters per token, which makes the “8× larger” claim misleading.
Questions for authors
  1. What ARA do agents reach with the same stratum-derived intents but no token or a shuffled token? Does the gap over the mismatch baselines remain?
  2. Is the window used to assign tokens and strata separate from the window in which real add-to-cart and purchase rates are measured? If the two overlap, the evaluation only reconstructs rates the token already encodes.
  3. Synthetic A/B testing needs agents that respond to storefront changes the way humans do. When a store changes its theme, does agent conversion shift in the same direction, and by a similar amount, as real conversion?
9 comments
AuthorsOpenReview ↗

Answer to Question 1

Q1- Thank you for the suggestion. We ran the requested controlled ablations to measure the contribution of the persona token while holding the stratum-derived intent fixed. First, we remove the token entirely:

StratumSimPersonaNo token
A: Purchasers72.4%72.4%
B: Checkout abandoners74.8%77.7%
C: Cart builders72.1%68.9%
D: Window shoppers95.4%64.6%
Stratified average78.65%70.94%

The no-token result shows that the stratum intent is useful but too coarse. For high-intent cohorts, especially A and B, it already explains much of the behavior: purchasers remain unchanged, and checkout abandoners even improve slightly. This is expected because A–C receive purchase-oriented intents, and stratum B contains users who naturally proceed far into the funnel. However, as true shopping tendency decreases, the no-token agent becomes less aligned: ARA drops for cart builders and collapses for window shoppers. This indicates that without the token, agents tend to over-act on the coarse intent rather than learning when to simply browse. For example, among window shoppers with a browsing intent, 46.14% of no-token agents added to cart and 24.01% started checkout; with SimPersona, these rates drop to 0.90% and 0.33%, respectively.

We note that ARA is still a conversion-alignment metric so it can be influenced by the intent itself. This is why we also examine if the learned behavioral dimensions are preserved in the agent trajectories (not just checking if the final conversion rates match). The no-token comparison shows that intent alone mainly induces purchase behavior, while SimPersona preserves broader behavioral structure. Specifically, we show in the Table below that the no-token model still shows significant separation for purchase, which is expected because purchase intensity is directly encouraged by the explicit intent.

ModelHeadLowMediumHighCohen's dt-test pPermutation p
No tokenPurchase3.254.737.151.873.63e-130.000100
SimPersonaPurchase0.062.947.252.805.5e-22< 0.0001
No tokenEngagement2.182.472.520.380.14860.08349
SimPersonaEngagement1.552.002.410.764.1e-30.003

However, this separation is weaker than SimPersona: Cohen’s d drops from 2.80 to 1.87. The difference is clearer for engagement. Without the token, engagement separation is not statistically significant under either the t-test or permutation test, while SimPersona shows significant separation with a larger effect size. Thus, without the persona token, the agent mostly follows the coarse shopping instruction; with the token, it learns behavioral aspects of the buyer, including how strongly to engage with the storefront and when not to move down the purchase funnel.

We also tested a shuffled-token condition. Here, for a target buyer group/token X, we assign the agent a different token Y while keeping the same stratum-derived intent, then evaluate whether the generated behavior still matches the intended group X:

Stratum of XShuffled token: Y → XAll mismatch1-R
A: Purchasers45.63%65.51%65.45%
B: Checkout abandoners48.92%66.14%66.20%
C: Cart builders57.27%69.01%69.06%
D: Window shoppers74.81%66.76%66.75%
Stratified56.66%66.85%66.86%

Focusing on the first column we observe that when the wrong token Y is assigned to the intended group X, ARA drops to 56.66%. This is far below the correct SimPersona pairing, 78.65%, and is in the same range as the mismatch baselines reported in the paper, where stratified ARA is 58.9% for all mismatches and 57.6% for one-random mismatch. Thus, the gap over the mismatch baselines remains: the correct pairing is about 20–22 percentage points higher than the shuffled/mismatched-token conditions. The other two columns serve as a sanity check on where the shuffled trajectory moves. Once token Y is assigned, the generated behavior is no longer expected to resemble the original target population X; it can become closer to populations associated with other tokens. This is why the mismatch/random columns are higher than the Y -> X score in this table.

The new ablations are added to Appendix J.3 of the revised manuscript.

AuthorsOpenReview ↗

Answer to Question 2

Q2- SimPersona is designed to reconstruct and simulate a merchant's current shopper population. The evaluation pipeline has two steps. First, we pass real buyer clickstreams through the learned, frozen VQ-VAE encoder to assign persona tokens and estimate the merchant's shopper mix, i.e., what fraction of the population belongs to each learned behavioral persona. Second, we sample agents according to that token mixture and condition them on the persona token, session intent, and storefront context to generate shopping trajectories. The evaluation then asks whether this simulated population reproduces the aggregate behavior of the corresponding real shopper population.

In our evaluation, the token-assignment window and the reference-behavior window do overlap: both are computed from the same evaluation population and period. This is intentional because SimPersona is evaluated as current-population behavioral reconstruction. A merchant wants to know whether agents instantiated from its observed shopper mix behave like that shopper population. The learned codebook and agent are reusable; for a new merchant or a later period, we do not need to retrain the model, but re-encode the available clickstreams with the learned encoder to estimate the current persona mixture.

We also emphasize that persona tokens are expected to preserve behavioral information from clickstreams. That is the purpose of persona learning. The evaluation is not trying to prove that tokens are independent of ATC or checkout tendencies. Rather, it tests whether this compressed behavioral representation, when used to condition an agent, can reproduce population-level shopping behavior. The agent is never given the aggregate ATC or checkout rates used for scoring; it receives the persona token, session intent, and storefront/page context, and must interact with the storefront to produce simulated shopping outcomes.

Importantly, a token is not a fixed shop-level conversion-rate label. The same token can appear across many shops and realize different ATC/checkout rates depending on the storefront and context. For example, as shown in the below table, token 56 appears in 20 shops and has substantial ATC and checkout variance, while token 64 appears in 23 shops and also spans a wide range of checkout behavior. Thus, the token should be understood as a behavioral prior, not as a lookup key for a single conversion rate.

TokenShopsATC meanATC varianceCheckout meanCheckout variance
642391.81%21.9457.77%332.32
562068.15%115.4925.52%146.83
801067.93%99.1632.99%261.50
66290.42%0.160.25%0.08

The window-shopper ablation shown for question 1, further illustrates how the learned behavioral representation improves simulation fidelity. Under the same browse intent, removing persona conditioning makes agents over-convert: 46.14% add to cart and 24.01% start checkout. With SimPersona, these rates fall to 0.90% and 0.33%, and window-shopper ARA improves from 64.66% to 95.4%. This is especially important because many merchant populations are dominated by low-intent browsing traffic. Thus, the learned token is not used as a stored shop-level rate; it is a behavioral prior that helps the agent translate the merchant's reconstructed shopper mix into realistic population-level trajectories.

AuthorsOpenReview ↗

Answer to Question 3

Q3- We thank the reviewer for this suggestion. Offline A/B testing is an important downstream use case for user simulation, and it provides another way to evaluate whether SimPersona captures meaningful buyer behavior beyond static conversion alignment. We therefore ran a theme-change diagnostic that asks whether simulated buyers move in the same direction as real buyers when the storefront changes. Each of our 42 evaluation shops straddles a live theme change. For each shop, we compute the pre/post change in human ATC rate and the pre/post change in simulated ATC rate. SimPersona matches the direction of the human ATC change in 28/42 shops, i.e., 66.7% of theme changes.

We also report Spearman rank correlation between the signed human and simulated ATC changes. We use Spearman rather than Pearson here because theme-change magnitudes are heterogeneous across shops and can be affected by a few large deltas; the rank correlation asks the more stable question of whether shops with larger human shifts also tend to have larger simulated shifts among the directionally correct cases. Under this metric, SimPersona obtains Spearman ρ=0.691\rho=0.691, with a mean absolute discrepancy of 3.79 percentage points.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary: Contributions, Evidence, and Clarifications

Dear Area Chair,

As the discussion period comes to a close, we thank you and the reviewers for your time and for the opportunity to strengthen our work.

We would like to highlight two main contributions of SimPersona:

1- To the best of our knowledge, SimPersona is the first approach for LLM-based shopping simulation that represents a buyer’s clickstream history with a single learned persona token. The single token keeps the conditioning context short, and assigning it takes one encoder pass per buyer rather than generating a text profile.

2- SimPersona estimates each merchant’s buyer-type distribution from the frequencies of tokens assigned to its buyers. Sampling from this distribution connects the learned behavioral representation to population-level simulation, allowing the simulated shopper mix to reflect the merchant’s observed population.

We also provide an open-source pipeline for converting e-commerce event logs into buyer representations and agent-training traces. Our evaluation compares agent simulations with a human reference population of 8.37 million buyers across 42 storefronts held out from training.

Reviewer 7xtE raised their overall rating to weak acceptance after considering our responses and additional results. Reviewer fDzX confirmed support and wrote that they raised the presentation score from 2 to 3, finding the price-level analysis particularly informative. Reviewer K1Bh has not yet posted a follow-up response.

We highlight the following evidence and clarifications relevant to the reviewers’ concerns:

Token contribution (K1Bh, 7xtE; Appendix J.3). For the same trained agent, holding stratum-derived intent fixed, stratified ARA is 78.65% with the correct token, 70.94% without a token, and 56.66% with an incorrect token. The largest benefit occurs for window shoppers: under the same browse intent, add-to-cart rates are 46.14% without the token and 0.90% with SimPersona.

Matched-backbone comparison (all three reviewers; Section 3.4 and Appendix J.2). We added a Qwen3-14B text-persona comparator with the same training trajectories, text-only observations, and evaluation protocols. On instruction-following tasks, SimPersona reaches the required action in 91.4% versus 69.3% of cart tasks and 76.9% versus 64.1% of checkout tasks. In the separate shop-level simulation evaluation, ARA is 92.7% versus 72.8%.

Population reconstruction (7xtE; Appendix B). Estimating token profiles and centroids exclusively from training shops largely preserves reconstruction quality on held-out shops: mean Jensen–Shannon divergence for stratum recovery is 0.052, compared with 0.054 originally. This addresses whether reconstruction depends on estimating token profiles from the evaluation population.

Clustering evaluation (K1Bh, fDzX; Section 3.1 and discussion responses). The coherence metrics assess the behavioral structure encouraged by the training objective and are interpreted as representation diagnostics. The fixed-intent ablations above separately test whether token assignment affects the trained agent’s behavior.

Product and price behavior (fDzX; Appendix J.4). Over 98% of classifiable cart and checkout episodes contain a product from the requested family. At the broader taxonomy level, mean human–agent category-distribution overlap is 78.0% for carted products and 79.3% for viewed products. In each within-shop price tier, aggregate human and SimPersona cart shares differ by at most 4.4 percentage points.

Within-token variation (fDzX; Appendix J.5). With buy intent fixed, token 64’s simulated checkout rate ranges from 10.0% to 90.9% across 13 storefronts. Episodes sharing a token also vary in interaction depth. These analyses show that sharing a token still permits variation across contexts and trajectories.

Preliminary storefront-change diagnostic (K1Bh, Q3 discussion response). In an observational pre/post analysis, simulated add-to-cart changes match the direction of human changes in 28 of 42 shops (66.7%). Controlled validation of A/B-test effects remains future work.

As clarified in our responses, the evaluation measures current-population reconstruction: token assignment and reference behavior use overlapping periods, and session intents are derived from observed strata. The agent receives a persona token, session intent, and storefront context, then generates shopping actions without being given the aggregate reference rates. Simulated purchase-flow outcomes denote checkout reached. Concrete prompts, tasks, and recorded rollouts are provided in Appendix L.

We believe these additions strengthen the evidence that learned persona tokens provide behavioral control beyond session intent and support population-level simulation across held-out storefronts. We would be grateful for your consideration in the final assessment.

Thank you again for your time.

The SimPersona Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.