acceptodds
Under review as a conference paper at ICLR 2027

ShopMind: A Mental World Model for Personalized and Realistic E-commerce User Simulation

Abstract

Large language models are widely used in e-commerce user simulation, providing a foundation for understanding users and predicting human decisions. However, existing simulators still struggle to achieve both personalization and behavioral realism. Most methods predict the next action directly from interaction histories and the current context, without accounting for the evolving mental world that underlies user behavior. Theory of Mind suggests that behavior is driven by mental states, including beliefs, uncertainties, and intentions. Users may interpret the same information in distinct ways, leading to divergent behaviors, while an individual’s mental state evolves through interaction and shapes subsequent actions. Ignoring this process prevents simulators from faithfully capturing individual differences and realistic behavior patterns, and undermines its effectiveness in downstream applications. To address this, we propose ShopMind, a Mental World Model for e-commerce user simulation that captures how consumers interpret shopping information and how their mental states evolve, better reflecting individual users and real-world behavior patterns. We further use behavior–mental state pairs synthesized by ShopMind to train a Mental State Reconstruction (MSR) model, which infers consumers’ latent mental states from observable behavior. Experiments show that ShopMind improves personalization by 18.0% and behavioral realism by 19.6% over the strongest alternatives. Moreover, MSR improves performance on consumption value measurement and CTR prediction, showing the utility of ShopMind for understanding and predicting real user behavior.

Primary area:
generative models
Keywords:
LLM, User Simulation, Theory of Mind, World Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xCMXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ShopMind uses a large language model to simulate users' shopping behavior. It records simulated users' beliefs, uncertainties, and purchase intentions, simulates several possible next actions, and then selects one. The paper also trains an MSR model on shopping behavior and mental-state descriptions generated by the large language model. MSR infers users' mental states from shopping records and is used for consumption value measurement and CTR prediction.

Critical strengths
  1. The method considers differences between users, as well as changes in the same user's thoughts during shopping.
  2. The ablation experiments examine the roles of four main components.
  3. The trained MSR model improves scores on consumption value measurement and CTR prediction. It also performs better than directly using the same base model, without MSR training, to generate mental-state descriptions.
Critical weaknesses
  1. There is still a lack of direct validation of whether the generated mental-state descriptions reflect real users' thoughts.

MSR learns from mental-state descriptions generated by ShopMind itself. The paper uses results on consumption value measurement and CTR prediction to show that these descriptions are useful. However, it does not sufficiently show whether they reflect what real users were thinking at the time.

  1. The paper does not clearly state which models each method uses or how many model calls it makes.

The main text says that ShopMind uses DeepSeek-V4-Pro, while the appendix says that the comparisons mainly use Qwen2.5-7B-Instruct. These statements seem inconsistent. In addition, the paper does not compare the total number of model calls or the running costs of the methods.

  1. The gains on some metrics are small.

For example, on ShopCART, ShopMind's Macro-F1 is 49.14, compared with 48.40 for AlignUSER, a difference of only 0.74 points.

  1. The code and model files needed to reproduce the experiments are incomplete.

The public repository does not explain how to obtain the Behavior Critic's model weights or provide instructions for training it. It also lacks the MSR training code, making it difficult to reproduce the main results with the available materials.

Questions for authors
  1. Beyond improved prediction scores, what evidence shows that the mental-state descriptions generated by MSR reflect what real users were thinking at the time?

  2. Which models does each method actually use, and how many model calls does each method make?

  3. Why are the gains on some metrics small? Are these gains stable across multiple runs?

  4. Will you provide the Behavior Critic's model weights and code for MSR training and related steps?

3 comments
AuthorsOpenReview ↗

Response to Reviewer xCMX–Section1

We sincerely thank the reviewer for the constructive feedback and for recognizing ShopMind's personalized modeling, ablation studies, and the downstream utility of MSR. We address each concern below.

Weakness 1 & Question 1: Validation of Mental States

First, consumers' mental states are inherently latent and difficult to observe. Real-world e-commerce logs typically record behavioral interactions without continuous self-reports of beliefs, uncertainties, or intentions, making it difficult to obtain ground-truth mental states. This motivates our use of ShopMind to generate behavior–mental-state pairs and train MSR to reconstruct latent mental states from observed behavior.

Second, we evaluate mental states through behavioral consistency and downstream utility. During data construction, multi-stage quality control and human review assess whether mental states are supported by observed evidence, consistent with user behavior, and coherent over time. On real user data, MSR-reconstructed mental states improve Primary Accuracy from 0.6518 to 0.7868 in consumption value measurement and AUC from 0.6441 to 0.7156 in CTR prediction with DeepSeek-V4-Pro.

Moreover, our Zero-shot Mental State baseline shows that directly inferring mental states with a general-purpose LLM provides smaller gains or even degrades performance. This suggests that the improvements are not simply due to adding mental-state descriptions, but benefit from ShopMind-generated supervision.

We acknowledge that these results demonstrate behavioral consistency and predictive utility, rather than direct agreement with users' subjective experiences. Such validation requires collecting self-reported mental states during real shopping sessions. We will clarify this limitation in the revised manuscript and consider human-grounded validation in future work.

Our contribution is not to claim precise recovery of consumers' actual thoughts, but to explicitly model evolving mental states and demonstrate their value for realistic behavior simulation and downstream prediction.

Weakness 2 & Question 2: Model Configurations and Inference Cost

The main text and appendix refer to different model configurations. ShopMind uses DeepSeek-V4-Pro for simulation, Qwen3-4B-Instruct-2507 for the Behavior Critic, and Qwen3-8B for MSR. The Qwen2.5-7B-Instruct in Appendix D.2 is used for baseline comparisons. We acknowledge that Appendix D.3 is unclear and will clarify the backbone and training settings for each method. To examine the impact of backbone differences, we additionally evaluate ShopMind with Qwen2.5-7B-Instruct on Taobao, using the same backbone as the baselines.

MethodTrans. JS ↓Len. Err. ↓Buy Err. ↓Action JS ↓
Zero-shot44.826.9527.6425.81
SFT17.834.3512.7713.26
SFT+Rationale20.314.8810.4213.95
SimUSER v229.636.3814.2122.16
DGDPO34.886.1724.9320.42
ContextSim18.945.229.1814.03
IToM40.266.8429.7723.65
AlignUSER16.724.125.3612.47
Shop-R127.144.563.9211.84
Customer-R131.773.917.1314.28
ShopMind (Qwen2.5-7B)15.383.844.2610.96

Under the same backbone, ShopMind achieves the best performance on three of the four metrics, reducing Trans. JS, Len. Err., and Action JS by 8.0%, 1.8%, and 7.4%, respectively, over the strongest baselines. It also achieves the second-best Buy Err.

Regarding inference cost, ShopMind's multi-stage decision process introduces additional overhead. On OPeRA and ShopCART, Mental-World Rollout is disabled, avoiding the cost of future trajectory simulation. When enabled, rollout considers K = 4 candidates, each expanded for up to L = 2 steps, yielding at most eight simulated rollout steps per decision. Actual model calls depend on state updates, early termination, and behavior evaluation. Appendix B further shows that increasing L from 0 to 2 reduces Transition JS from 22.63 to 12.18 on Taobao, while longer rollouts provide no further gains. We will clarify these configurations and discuss the trade-off between performance and inference cost in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer xCMX–Section2

Weakness 3 & Question 3: Performance Gains

We agree that some individual improvements are modest, such as the 0.74-point Macro-F1 gain over AlignUSER on ShopCART. However, ShopMind aims to improve both individual consistency and behavioral realism, so its performance should be assessed across multiple dimensions. On ShopCART, ShopMind also improves Exact Accuracy by 3.27 percentage points, while reducing Transition JS and Action JS by 14.6% and 15.2%, respectively, over the strongest baselines. Similar gains appear on other datasets: OPeRA Top-3 Accuracy increases from 47.89 to 60.85, and Taobao Transition JS decreases from 16.72 to 12.18. These results show that ShopMind's improvements extend beyond individual metrics, particularly in long-horizon behavioral realism across datasets.

Weakness 4 & Question 4: Reproducibility

We have already made the core implementation and detailed experimental configurations available in our anonymous repository to support reproducibility. While some materials are not yet included, we will release the remaining resources, together with the necessary training instructions, upon acceptance of the paper to further support reproducibility.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable feedback. We have provided detailed responses and additional experimental results to address your concerns. We would greatly appreciate it if you could review our rebuttal at your convenience and let us know whether any concerns remain. We sincerely hope that our responses and additional results may help alleviate your concerns and provide a basis for reconsidering your initial assessment. If you have any further questions or suggestions, we would be more than happy to provide additional clarifications. Thank you for your time and consideration.

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

This paper presents ShopMind, a method for tracking users' mental states while browsing e-commerce websites to improve personal behavior simulation. The authors state that existing e-commerce simulators directly predict the next action based on observations. ShopMind adds an explicit mental state layer to support next-action prediction by explicitly modeling how users comprehend the current information.

Critical strengths

The overall method makes sense, and the results look good.

Critical weaknesses

My main concern is the unclear description of the study. As far as I know, neither the Opera dataset nor the Shopcart dataset provides an executable environment.

It is therefore unclear to me how the authors performed rollouts without executable environments. How can the authors obtain the observation for the next step if the agent's action is not present in the dataset (for example, if the agent performs another search with a different keyword)?

I also think an important baseline experiment is missing: a mental-model-aware prompting-based baseline. Similar to the zero-shot prompting baseline, this approach would first prompt the model to explicitly write out its mental state and then predict the next action.

Finally, a minor issue is that the authors should discuss other relevant work on evaluating user simulator agents in e-commerce scenarios, including but not limited to:

https://arxiv.org/pdf/2608.20707

https://dl.acm.org/doi/full/10.1145/3706599.3719729

https://arxiv.org/pdf/2604.08362

https://dl.acm.org/doi/full/10.1145/3772363.3799039

Questions for authors

How did the authors perform rollouts given that neither dataset provides an executable environment? Specifically, how are next-step observations obtained when the agent takes an action that is not present in the dataset?

3 comments
AuthorsOpenReview ↗

Response to Reviewer AYeF

We sincerely thank the reviewer for the thoughtful and constructive feedback, as well as for recognizing the merits of our method and experimental results. We address the concerns about the evaluation environment, prompting baseline, and related work below.

Weakness 1 & Question 1: Mental-World Rollout and Evaluation Environment

Thank you for raising this important concern. We agree that OPeRA and ShopCART do not provide fully interactive environments supporting arbitrary action execution. Our original manuscript did not clearly distinguish the configurations and evaluation protocols used across datasets, particularly regarding Mental-World Rollout. We apologize for this lack of clarity and provide the following clarification.

For these two datasets, we use a ShopMind configuration without future rollouts, retaining mental-state modeling and transition, along with candidate action selection based on the current context. Thus, the model does not execute candidate actions to obtain future observations, nor does it treat subsequent recorded pages as action-dependent rollout branches. Consequently, the results on these datasets reflect the configuration without rollout and should not be interpreted as demonstrating the contribution of Mental-World Rollout itself. The ablation study of this component in Table 5 is conducted on Taobao.

Weakness 2: Mental-Model-Aware Prompting Baseline

Thank you for this valuable suggestion. We agree that a mental-model-aware prompting baseline would help determine whether ShopMind's improvements arise simply from explicitly describing mental states before predicting actions. Following this suggestion, we introduce a Mental-Model-Aware Prompting baseline. Given the user profile, shopping task, current observation, and interaction history, the model first generates the user's beliefs, uncertainties, and intentions, and then predicts the next action:Context → Beliefs / Uncertainties / Intentions → Next Action . For a consistent comparison, Zero-shot, Mental-Model-Aware Prompting, and ShopMind all use Qwen2.5-7B-Instruct. The following table reports long-horizon behavioral realism on Taobao, where lower values indicate better performance.

MethodTrans. JS ↓Len. Err. ↓Buy Err. ↓Action JS ↓
Zero-shot44.826.9527.6425.81
Mental-Model-Aware Prompting29.765.8317.6218.94
ShopMind (Qwen2.5-7B)15.383.844.2610.96

Mental-Model-Aware Prompting improves all four metrics over Zero-shot, suggesting that explicitly inferring mental states before action prediction is beneficial. However, it still falls considerably behind ShopMind. The result show that independently inferring mental states at each step may lead to inconsistencies across the generated sequence, even when individual predictions appear reasonable. Moreover, direct next-action prediction cannot account for how different plausible actions may lead to different subsequent shopping trajectories. In contrast, ShopMind maintains consumers' beliefs, uncertainties, and intentions across steps, updating them only when new information warrants a change. Mental-World Rollout further explores how candidate actions may affect subsequent observations, mental states, and behaviors. Personalized Consistency evaluates whether candidate trajectories align with the user's characteristics and current mental state, while the Behavior Critic selects actions based on patterns learned from real user behavior. Together, these mechanisms help maintain coherent mental-state transitions and realistic behavioral trajectories, which explains ShopMind's stronger sequential simulation performance. We will include this additional baseline in the revised manuscript.

Weakness 3: Discussion of Related Work

We thank the reviewer for suggesting these relevant studies. RecVerse improves shopping behavior simulation through hierarchical memory and trajectory-level reinforcement learning, while OmniBehavior evaluates long-horizon user simulation and highlights limitations such as excessive activity and behavioral homogenization. UXAgent and Agent A/B further demonstrate the applications of simulated users in usability testing and A/B experimentation.

In contrast, ShopMind focuses on explicitly modeling consumers' evolving mental states, including beliefs, uncertainties, and intentions, to improve personalized and realistic shopping behavior simulation. Beyond simulation, ShopMind further generates behavior–mental-state pairs to train MSR, whose utility is evaluated on real-world consumption value measurement and CTR prediction tasks. We will incorporate these studies into the revised Related Work section.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable feedback. We have provided detailed responses and additional experimental results to address your concerns. We would greatly appreciate it if you could review our rebuttal at your convenience and let us know whether any concerns remain. We sincerely hope that our responses and additional results may help alleviate your concerns and provide a basis for reconsidering your initial assessment. If you have any further questions or suggestions, we would be more than happy to provide additional clarifications. Thank you for your time and consideration.

thanks, i've adjusted my score accordingly.

Reviewer_hBfoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ShopMind, a mental world model for e-commerce user simulation. Rather than predicting the next action directly from observable context, ShopMind maintains a structured mental state (beliefs, uncertainties, intentions), updates it with a gated transition, evaluates candidate actions via multi-step mental-world rollouts, and selects among them with a Behavior Critic trained on real trajectories. The authors additionally use ShopMind to synthesize 5K behavior–mental-state pairs and train an MSR model that inverts the mapping, and evaluate the transfer of MSR-reconstructed mental states on consumption value measurement and CTR prediction. Experiments cover OPeRA, ShopCART, and Taobao against ten baselines spanning prompting, SFT, RL, and agent paradigms, reporting +18.0% personalization and +19.6% behavioral realism.

Critical strengths
  1. The modeling shift is well motivated and non-trivial. Moving from "context → action" to "context → mental state → action" is grounded in Theory of Mind and in a concrete failure mode of existing simulators: the same observation induces different mental states across users, so observable-signal-only policies cannot be both personalized and realistic at once. The paper states this cleanly and Figure 1 makes the point immediately.
  2. The architecture is coherent as a whole. The three modules are not independent add-ons: the gated transition produces the state, rollouts use that state to score candidate futures, and the Critic arbitrates with real data. The ablation in Table 5 supports this decomposition, with removing the mental world causing by far the largest degradation
Critical weaknesses
  1. The main experiments are confounded by the choice of backbone. ShopMind is built on DeepSeek-V4-Pro, which is substantially stronger than the backbones used by all compared baselines. The paper provides neither a comparison in which the baselines are re-run on the same backbone, nor an ablation of ShopMind across different backbones. The reported gains therefore cannot be separated from backbone capability.
Questions for authors
  1. Can you report ShopMind's performance on other backbones (e.g., Qwen2.5-7B-Instruct, GPT-4o-mini)? This would determine whether the gains come from the architecture rather than the backbone.
2 comments
AuthorsOpenReview ↗

Response to Reviewer hBfo

Response to Reviewer hBfo

We sincerely thank the reviewer for the constructive feedback and for recognizing the motivation and coherence of our framework. We address the concern about backbone differences below.

Weakness 1 & Question 1: Backbone-Controlled Evaluation

Thank you for highlighting this important concern. We agree that the original comparison does not fully separate the contribution of ShopMind's architecture from differences in backbone capability. To address this concern, we additionally evaluate ShopMind using Qwen2.5-7B-Instruct, matching the backbone used in the baseline comparison, while maintaining the same data splits and evaluation protocol. The following table reports long-horizon behavioral realism on Taobao. All methods in this comparison use the same backbone, and lower values indicate better performance.

MethodTrans. JS ↓Len. Err. ↓Buy Err. ↓Action JS ↓
Zero-shot44.826.9527.6425.81
SFT17.834.3512.7713.26
SFT+Rationale20.314.8810.4213.95
SimUSER v229.636.3814.2122.16
DGDPO34.886.1724.9320.42
ContextSim18.945.229.1814.03
IToM40.266.8429.7723.65
AlignUSER16.724.125.3612.47
Shop-R127.144.563.9211.84
Customer-R131.773.917.1314.28
ShopMind (Qwen2.5-7B)15.383.844.2610.96

Under the same backbone, ShopMind achieves the best performance on three of the four metrics, with relative reductions of 8.0%, 1.8%, and 7.4% in Trans. JS, Len. Err., and Action JS, respectively, compared with the strongest baseline on each metric. ShopMind also achieves the second-best Buy Err. These results show that ShopMind retains its advantages over existing methods when using the same backbone. The improvements therefore cannot be attributed solely to a stronger backbone. We will incorporate these additional results and clarify the backbone configurations in the revised manuscript.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable feedback. We have provided detailed responses and additional experimental results to address your concerns. We would greatly appreciate it if you could review our rebuttal at your convenience and let us know whether any concerns remain. We sincerely hope that our responses and additional results may help alleviate your concerns and provide a basis for reconsidering your initial assessment. If you have any further questions or suggestions, we would be more than happy to provide additional clarifications. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.