Existing evaluations often test persona expression rather than whether persona information persists in extended decision-making. We introduce THEATRON, a controlled benchmark that uses financial replay to evaluate persona persistence in long-horizon sequential decision-making. In these environments, models conditioned on fixed behavioral-finance personas observe shared daily market evidence and trajectory-specific account-state feedback and make sequential decisions over multi-month trajectories. We evaluate Zero-Shot, One-Shot, and Chain-of-Thought persona injection across four LLMs, using cumulative-return and action trajectories to diagnose within-persona stability, cross-persona separability, and action collapse. Persona conditioning produces systematic behavioral shifts beyond a no-persona baseline, and persona-level separation generally exceeds the variation induced by semantically equivalent descriptions. Nevertheless, these shifts do not reliably form coherent persona-specific policies: apparently stable separation can arise from stereotyped action collapse, while diverse policies can exhibit noisy trajectory switching. These findings expose a gap between observable persona-induced behavior and persistent, context-sensitive persona-conditioned decision policies. Code and benchmark data are publicly available at: https://github.com/yuyan12138/theatron
Figures & tables
Figure 1: Overview of the Theatron benchmark pipeline.
Model
Mtd.
Stock
Overconfidence
Loss Aversion
Herding
B/W (R/A) ↑
GPT-5.4-mini
Zero-Shot
AAPL
−12.83±15.40
4.52±9.03
−3.07±6.35
1.21/2.89
TSLA
−4.35±21.95
22.79±7.05
21.48±11.93
1.78/2.40
One-Shot
AAPL
−39.91±10.33
0.09±5.53
−14.28±5.37
3.40/6.11
TSLA
15.91±23.11
14.55±20.42
20.89±7.63
1.48/4.10
CoT
AAPL
−29.10±6.15
16.76±6.23
−1.38±9.84
3.06/2.03
TSLA
2.74±13.82
−4.10±17.17
−8.80±9.54
2.21/2.31
Table 2: Main results. Persona columns report final cumulative log return (%, mean ± SD); B/W reports return/action ratios. Mean-return 95% CIs: Table 10 .
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Mode
Maximum tokens
Seed semantics
GPT-5.4-mini
no explicit reasoning
512
best effort
Gemini-2.5- Flash
thinking disabled
2,048
best effort
Qwen3-8B
non- thinking
512
explicit
Qwen3-14B
thinking
8,192
explicit
Appendix
Table 3: Model generation settings.
Positive-result component
Operational evidence
Reported in
Beyond no-persona behavior
Exact persona–baseline action and return contrasts
Appendix C , F
Within-persona stability
Din , final-return SD, and intervals
Appendix D – H
Cross-persona separation
Return/action Dout and B/W
Appendix E – H
Non-collapse and state sensitivity
Entropy, dominant actions, switches, and responses to price direction
Appendix D , J
Observable intended direction
Overconfidence activity, Herding alignment, and descriptive Loss-Aversion feedback
Appendix C , F
Appendix
Table 4: Operationalization of the joint positive-result criterion.
Model
Stock
Final CR (mean ± SD)
Entropy
BUY
HOLD
SELL
Qwen3-8B
AAPL
10.0±2.8
.173
1.0
96.5
2.5
Qwen3-8B
TSLA
15.3±12.8
.252
1.4
93.9
4.7
Qwen3-14B
AAPL
12.1±15.6
.690
6.5
75.9
17.6
Qwen3-14B
TSLA
22.0±16.9
.816
4.5
58.4
37.1
GPT-5.4-mini
AAPL
9.7±5.9
.393
4.9
90.0
5.1
GPT-5.4-mini
TSLA
18.1±8.1
.658
3.3
75.1
21.6
Appendix
Table 5: No-persona behavior ( K=5 ). Returns and 102-decision action rates are percentages; entropy also uses all 102 decisions.
Model
Stock
Task non-HOLD
Overconf. non-HOLD
Uplift
Task align.
Herding align.
Uplift
Qwen3-8B
AAPL
3.5
100.0
96.5
3.6
64.4
60.8
Qwen3-8B
TSLA
6.1
100.0
93.9
5.9
72.7
66.7
Qwen3-14B
AAPL
24.1
99.8
75.7
16.6
72.3
55.6
Qwen3-14B
TSLA
41.6
100.0
58.4
32.7
85.5
52.9
GPT-5.4-mini
AAPL
10.0
61.4
51.4
9.1
63.8
54.7
GPT-5.4-mini
TSLA
24.9
74.9
50.0
22.2
78.0
55.8
Appendix
Table 7: Zero-Shot directional checks. Activity uses 102 recorded decisions; alignment uses 101 price transitions and means buy after a rise or sell after a fall. Values are percentages; uplift is in percentage points.
Model
Stock
Paired K
ΔΔH
Sign-flip p
BH q
Qwen3-8B
AAPL
5
-.619
.0625
.100
Qwen3-8B
TSLA
4
.331
.3750
.429
Qwen3-14B
AAPL
5
-.223
.0625
.100
Qwen3-14B
TSLA
5
-.277
.0625
.100
GPT-5.4-mini
AAPL
5
-.000
1.0000
1.000
GPT-5.4-mini
TSLA
5
.273
.1250
.167
Appendix
Table 8: Loss-Aversion feedback diagnostic relative to no-persona behavior.
Figure 2: Action entropy across models, injection methods, and personas.
Figure 3: Zero-Shot cumulative-return trajectories on TSLA (mean ± SD over five runs).
Model
Stock
Return B/W
Return q
Action B/W
Action q
GPT-5.4-mini
AAPL
3.401
.000114
6.111
.000114
GPT-5.4-mini
TSLA
1.476
.000200
4.098
.000114
Gemini-2.5-Flash
AAPL
3.760
.000114
11.147
.000114
Gemini-2.5-Flash
TSLA
6.745
.000114
6.422
.000114
Qwen3-8B
AAPL
4.026
.000114
26.479
.000114
Qwen3-8B
TSLA
2.703
.000114
14.881
.000114
Appendix
Table 11: Primary One-Shot separation under the shared mixed-evidence demonstration.
Model
Scenario
Stock
Return B/W
Return q
Action B/W
Action q
GPT-5.4-mini
Positive
AAPL
2.264
.000133
3.989
.000133
Positive
TSLA
2.055
.000133
3.515
.000133
Negative
AAPL
1.212
.035410
4.242
.000133
Negative
TSLA
2.099
.000686
2.916
.000133
Gemini-2.5-Flash
Positive
AAPL
2.623
.000474
5.155
.000133
Positive
TSLA
4.413
.000133
6.715
.000133
Appendix
Table 12: One-Shot demonstration-scenario ablations. The positive scenario uses BUY for every persona.
Scenario
Overconfidence
Loss Aversion
Herding
Mixed evidence
65.5–100.0
50.9–88.5
42.0–55.1
Positive evidence
67.4–92.7
0.4–5.2
28.5–40.8
Negative evidence
15.2–41.3
63.5–89.3
33.5–50.0
Appendix
Table 13: Demonstration-action match-rate ranges across models (percent).
Model
Condition
Final CR (mean ± SD)
BUY
HOLD
SELL
Entropy
Qwen3-8B
No-persona
−1.05±0.11
1.8
98.2
0.0
.089
Overconfidence
−14.68±2.29
99.4
0.6
0.0
.031
Loss Aversion
4.67±1.58
0.0
95.8
4.2
.170
Herding
−14.05±5.40
33.9
63.0
3.2
.730
Qwen3-14B
No-persona
.96±2.84
5.3
94.5
0.2
.217
Overconfidence
−14.75±4.45
93.9
5.9
0.2
.235
Appendix
Table 14: BMRN condition results. Returns and action rates are percentages; entropy uses the 101 consequential actions.
Figure 4: Zero-Shot persona separation across AAPL, TSLA, and BMRN. The dashed line marks B/W =1 .
Overconfidence non-HOLD uplift
Herding alignment uplift
Model
AAPL
TSLA
BMRN
AAPL
TSLA
BMRN
Qwen3-8B
.964
.939
.976
.608
.667
.267
Qwen3-14B
.754
.586
.885
.556
.529
.935
GPT-5.4-mini
.515
.499
.503
.547
.558
.709
Gemini-2.5-Flash
.414
.226
.911
.442
.323
.913
Appendix
Table 15: Baseline-relative directional effects across assets.
Model
Persona
Return ΔD
Return q
Action ΔD
Action q
Qwen3-8B
Overconfidence
.0974
.00794
.9901
.00794
Loss Aversion
.0254
.00794
.0327
.00794
Herding
.0354
.00794
.8139
.00794
Qwen3-14B
Overconfidence
.0946
.00794
.7817
.00794
Loss Aversion
.0386
.00794
.6209
.00794
Herding
.0141
.00794
.8492
.00794
Appendix
Table 17: AAPL R2 exact persona–baseline tests. ΔD discounts within-condition dispersion as defined in Appendix C ; BH correction is separate for the 12 return and 12 action tests.
Model
Stock
Metric
W
P
B
B/P
95% CI for B−P
GPT-5.4-mini
AAPL
Return
.0652
.0687
.0887
1.29
[.0079,.0270]
AAPL
Action
.1631
.2313
.4771
2.06
[.2261,.2628]
TSLA
Return
.1055
.1223
.1488
1.22
[.0075,.0432]
TSLA
Action
.1793
.2400
.4278
1.78
[.1750,.1978]
Gemini-2.5-Flash
AAPL
Return
.0436
.0757
.1811
2.39
[.0910,.1162]
AAPL
Action
.1122
.2163
.6673
3.09
[.4439,.4602]
Appendix
Table 18: Persona separation ( B ) relative to wording variation ( P ) and repeated-run variation ( W ).
Figure 5: Ratio of persona separation to paraphrase variation. The outlined cell is the sole B/P<1 result.
Return B/W
Entropy
Method
AAPL
TSLA
Overconf.
Loss Av.
Herding
Zero-Shot
2.61
12.93
.26
.45
.77
One-Shot
4.03
2.70
.11
.36
.67
CoT
2.25
2.60
.32
.49
.79
SFT
4.44
3.40
.58
.00
.86
Appendix
Table 19: Qwen3-8B SFT ablation and prompt-based conditions.
Date
Price
Action
Cumulative return
2025-01-02
379.28
buy
0.00%
2025-01-03
410.44
buy
7.90%
2025-01-06
411.05
buy
8.04%
2025-01-07
394.36
buy
3.90%
2025-01-08
394.94
buy
4.05%
2025-01-10
394.74
buy
4.00%
Appendix
Table 20: Representative all- buy trajectory on TSLA.
Figure 6: Five GPT-5.4-mini Overconfidence Zero-Shot trajectories on AAPL.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
Faculty of Computer and Information Science, University of Haifa, Israel · School of Therapy, Counseling and Human Development, University of Haifa, Israel · The PhD Program of Hermeneutics and Culture, Interdisciplinary Studies Unit, Bar-Ilan University, Israel +1