ShopMind: A Mental World Model for Personalized and Realistic E-commerce User Simulation
Authors: Peixuan Hou, Bin Chen, Li He, Jian Xu, Xiuli Ma, Guojie Song
Organizations: State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University · Alibaba Group
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.
Figures & tables
Figure 1: Illustration of the proposed ShopMind . Users may form different mental states from the same shopping information, leading to different behaviors. By modeling this evolving mental world, ShopMind enables more personalized and realistic user simulation than other approaches.
Figure 2: Overview of the ShopMind framework. The model maintains an evolving mental state, performs mental-world rollouts over candidate behaviors, and selects the final action by jointly considering personalization consistency and behavioral realism through a learned Behavior Critic.
Figure 3: Qualitative comparison of ShopMind and AlignUSER ( Bougie et al., 2026b ) , a strong baseline, on 80 randomly sampled episodes from the Taobao real log dataset. (a) ShopMind shows higher similarity to human trajectories. (b) A case illustrates how mental-world modeling helps ShopMind better follow human decision dynamics, while AlignUSER exhibits larger deviations.
OPeRA
Taobao
Method
Method Type
Ind. JS ↓
Spearman ↑
Gap Err. ↓
Δ JS ↑
Ind. JS ↓
Spearman ↑
Gap Err. ↓
Δ JS ↑
Zero-shot
Prompting
31.02
0.334
0.054
−0.44
34.28
0.281
0.061
−0.62
SFT
SFT
26.48
0.330
0.051
−1.41
29.82
0.308
0.051
−1.94
SFT+Rationale
SFT
24.31
0.363
0.039
−2.61
25.16 ‡
0.337
0.042 ‡
−0.91
SimUSER v2
Agent
32.24
0.238
0.055
+0.29
31.82
0.254
0.057
+1.14
DGDPO
Agent
30.22
0.296
0.037
+1.74 ‡
29.74
0.303
0.044
+1.86
Table 1: Performance comparison of individual consistency. ShopMind achieves the best performance across all metrics. ⋆ , † , and ‡ denote the first-, second-, and third-best results, respectively.
OPeRA
ShopCART
Method
Exact Acc. ↑
Non-Term. Acc. ↑
Macro F1 ↑
Top-3 Acc. ↑
Buy F1 ↑
Exact Acc. ↑
Macro F1 ↑
Top-3 Acc. ↑
Buy F1 ↑
Exit F1 ↑
Zero-shot
6.10
4.85
32.50
8.98
67.31 †
6.96
17.60
9.23
36.76
0.62
SFT
29.38
27.50
45.45
42.35
20.69
20.83
40.62
41.28
35.46
33.71
SFT+Rationale
33.04
29.89
42.32
46.23
52.05
15.14
39.09
43.60
59.70 †
14.25
SimUSER v2
5.32
3.55
35.81
12.31
42.67
4.26
27.98
1.23
29.51
1.23
DGDPO
5.99
5.14
34.33
9.20
67.31 †
6.41
20.72
10.85
41.63
3.17
Table 2: Performance comparison of local behavior prediction. ShopMind achieves the best performance on most metrics. ⋆ , † , and ‡ denote the first-, second-, and third-best results, respectively.
OPeRA
ShopCART
Taobao
Method
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Zero-shot
38.06
7.47
31.11
23.53
51.73
2.80
48.18
31.20
44.82
6.95
27.64
25.81
SFT
18.05
4.78
43.33
14.65
23.84
2.64
6.37
12.76
17.83 ‡
4.35
12.77
13.26
SFT+Rationale
14.45 ‡
5.13
15.56
10.84 †
33.05
4.27
1.40 ‡
15.30
20.31
4.88
10.42
13.95
SimUSER v2
32.74
7.21
12.22
25.04
28.07
3.62
19.89
21.78
29.63
6.38
14.21
22.16
DGDPO
39.14
7.11
34.44
22.16
41.36
3.18
31.27
22.43
34.88
6.17
24.93
20.42
Table 3: Performance comparison of long-horizon behavioral realism. ShopMind achieves consistent gains across all metrics. ⋆ , † , and ‡ denote the first-, second-, and third-best results, respectively.
Type
Model
NDCG @1 ↑
NDCG @2 ↑
NDCG @3 ↑
Primary Acc. ↑
Primary Top-2 Acc. ↑
General LLM
Claude-Haiku-4 ( Anthropic, 2025 )
0.7135
0.6721
0.6716
0.5743
0.7434
DeepSeek-V3.2 ( DeepSeek-AI et al., 2025 )
0.7211
0.6967
0.6917
0.5613
0.7565
Gemini-2.5-Flash ( Gemini Team, Google, 2025 )
0.7750
0.7026
0.6746
0.5780
0.7174
Gemini-2.5-Pro ( Gemini Team, Google, 2025 )
0.5353
0.5170
0.5142
0.4442
0.5576
GLM-5 ( GLM-5 Team et al., 2026 )
0.7044
0.6729
0.6594
0.5780
0.7137
GPT-4o ( OpenAI, 2024 )
0.7211
0.6417
0.6345
0.5260
0.6449
Table 4: Consumption value measurement results. Best, second, and third results are marked by ⋆ , † , and ‡ . MSR-reconstructed mental states provide the strongest gains for value measurement.
Figure 4: CTR prediction performance with and without MSR-reconstructed mental states. Adding mental states consistently improves AUC, GAUC, and PR-AUC across both models.
Local Fidelity
Sequential Realism
Individual Consistency
Variant
Exact Acc. ↑
Macro F1 ↑
Top-3 Acc. ↑
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Indiv. JS ↓
Spearman ρ↑
Δ JS ↑
ShopMind
53.58
59.02
72.4
12.18
3.62
2.84
9.41
18.63
0.421
+4.36
w/o Mental World
13.27
38.61
21.9
30.84
6.37
11.26
16.92
28.41
0.301
+0.84
w/o Rollout
50.84
56.47
68.9
22.63
5.41
7.38
15.27
20.12
0.397
+3.72
w/o Individual Consistency
50.96
55.83
68.2
13.86
3.91
3.31
10.24
29.16
0.318
+1.43
w/o Behavior Critic
44.72
49.86
62.5
18.41
4.67
6.14
14.38
20.37
0.389
+3.58
Table 5: Ablation study of ShopMind on Taobao logs across behavioral fidelity and personalization.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Local Fidelity
Sequential Realism
Individual Consistency
Candidates K
Exact Acc. ↑
Macro F1 ↑
Top-3 Acc. ↑
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Indiv. JS ↓
Spearman ρ↑
Δ JS ↑
K=2
51.02
56.74
69.5
15.74
4.23
4.12
11.53
19.74
0.404
+3.91
K=4
53.58
59.02
72.4
12.18
3.62
2.84
9.41
18.63
0.421
+4.36
K=6
53.49
58.93
72.3
12.24
3.65
2.89
9.46
18.69
0.420
+4.33
K=8
53.31
58.79
72.1
12.31
3.69
2.95
9.53
18.77
0.418
+4.28
Appendix
Table 6: Sensitivity of ShopMind to the number of candidate behaviors K on Taobao logs. The rollout horizon is fixed at L=2 .
Local Fidelity
Sequential Realism
Individual Consistency
Horizon L
Exact Acc. ↑
Macro F1 ↑
Top-3 Acc. ↑
Trans. JS ↓
Len. Err. ↓
Buy Err. ↓
Action JS ↓
Indiv. JS ↓
Spearman ρ↑
Δ JS ↑
L=0
50.84
56.47
68.9
22.63
5.41
7.38
15.27
20.12
0.397
+3.72
L=1
52.31
57.78
70.7
16.37
4.28
4.63
11.76
19.37
0.409
+4.08
L=2
53.58
59.02
72.4
12.18
3.62
2.84
9.41
18.63
0.421
+4.36
L=3
53.44
58.89
72.2
12.36
3.68
2.93
9.54
18.72
0.419
+4.31
L=5
53.12
58.61
71.9
12.81
3.81
3.18
9.87
18.94
0.415
+4.21
Appendix
Table 7: Sensitivity of ShopMind to the rollout horizon L on Taobao logs. The number of candidate behaviors is fixed at K=4 .
Stage
Input
Accepted
Acceptance
Generated purchase trajectories
7,200
7,200
100.0%
Deterministic checks
7,200
6,550
91.0%
LLM quality evaluation
6,550
5,100
77.9%
Human review and final filtering
5,100
5,000
98.0%
Final MSR training set
–
5,000
69.4% †
Appendix
Table 8: Quality-control statistics for MSR training data. Acceptance at each stage is computed relative to its input.
Dataset
Sessions
Train
Test
Eval. Steps
OPeRA
527
437
90
902
ShopCART
31,865
31,508
357
1,149
Taobao
10,000
8,000
2,000
17,436
Appendix
Table 9: Dataset statistics used in the behavior simulation experiments.
Backbone
Input
AUC ↑
GAUC ↑
PR-AUC ↑
DeepSeek-V4-Pro
Raw
0.6441
0.6456
0.1653
Raw + Zero-shot Mental State
0.6081
0.6090
0.1597
Raw + MSR Mental State
0.7156
0.7057
0.1924
Qwen2.5-72B
Raw
0.6317
0.6332
0.1628
Raw + Zero-shot Mental State
0.5964
0.5981
0.1491
Raw + MSR Mental State
0.6968
0.6895
0.1857
Appendix
Table 10: CTR prediction results on real-world Taobao interaction logs. Best results within each backbone are highlighted in bold .