Do Persona-Infused LLMs Affect Performance in a Strategic Reasoning Game?

Abstract: Although persona prompting in LLMs appears to trigger different styles of generated text, it is unclear whether these translate into measurable behavioral differences, much less whether they affect decision-making in an adversarial strategic environment that we provide as open-source. We investigate the impact of persona prompting on strategic performance in PERIL, a world-domination board game. Specifically, we compare the effectiveness of persona-derived heuristic strategies to those chosen manually. Our findings reveal that certain personas associated with strategic thinking improve game performance, but only when a mediator is used to translate personas into heuristic values. We introduce this mediator as a structured translation process, inspired by exploratory factor analysis, that maps LLM-generated inventory responses into heuristics. Results indicate our method enhances heuristic reliability and face validity compared to directly inferred heuristics, allowing us to better study the effect of persona types on decision making. These insights advance our understanding of how persona prompting influences LLM-based decision-making and propose a heuristic generation method that applies psychometric principles to LLMs.

• The paper demonstrates that integrating a personality inventory mapping significantly improves performance, with top TrueSkill scores exceeding 27 compared to controls near 18.
• It employs the PERIL platform to simulate complex strategic gameplay by comparing direct heuristic prompting with structured psychometric-based persona mapping.
• The study finds that PI-based agents exhibit strong correlations (ρ ≈ 0.4-0.6, p < 0.005) between persona features and strategic decisions, indicating improved interpretability and behavioral diversity.

Persona Prompting and Strategic Reasoning in LLMs: An Analysis via Strategic Gameplay

Experimental Design and PERIL Platform

This study examines whether persona-based prompting in LLMs induces measurable behavioral variability and impacts decision-making in adversarial, strategic contexts. The authors introduce PERIL, a bespoke implementation inspired by canonical strategic world conquest board games, designed for controlled evaluation of LLM-guided agents under persona-driven prompts.

Figure 1: PERIL, our implementation inspired by a popular strategic world conquest board game, was used to study the effects of persona-based prompting in LLMs.

PERIL is engineered to support a large action space, stochastic dynamics, and high-pressure decision cycles, making it an adverse environment for straightforward reasoning by LLMs. The platform operationalizes decision-making via explicit heuristics that govern agent actions across multiple game phases (initialization, attack, redeployment). Crucially, these heuristics are selected by LLMs at the start of each game—providing a proxy for precommitment and constraining cognitive variability during play.

Persona Derivation and Heuristic Mapping

The methodology centers on two distinct pipelines for mapping persona descriptions to agent heuristics:

1. Direct Heuristic (DH) Method: The LLM receives a persona and is prompted directly to assign numerical priorities to a slate of heuristics. This method is susceptible to high variability and poor alignment between persona and strategic decision parameters.
2. Personality Inventory (PI) Method: Inspired by psychometric instrumentation (notably the PID-5), a structured inventory questionnaire is constructed. Each persona is filtered through this inventory, with LLMs providing responses to each item. Inventory responses are algorithmically mapped to heuristic weights via a transformation that ensures proportionality and default baselining.

   Figure 2: Pipeline from persona descriptions to heuristic agents. Note that LLMs generate heuristic proxies (DH or PI), but do not play PERIL directly.

The PI method is underpinned by exploratory factor analysis principles, intended to mediate LLM interpretation and enforce more globally consistent persona-to-heuristic translation.

Evaluation: Model Families, Persona Selection, and Tournaments

Experiments leverage four LLMs (GPT-4o, Llama 3, Llama 4, and Mistral) to generate both DH and PI agents, using a backdrop of 50 personas selected from the Persona Hub corpus via greedy maximization of feature-variance across strategicThinker, domainExpert, perilSpecific, riskTaker, and doOrBe dimensions. Persona annotations are assigned via zeroshot GPT-4o mini inference.

Tournaments pit agents representing distinct personas against each other in thousands of two-player matches, with skill evaluated via the TrueSkill system. This regime allows identification of relationships between inferred persona features, selected heuristics, and empirical win rates.

Numerical Results: Impact of PI-Based Prompting

The first salient finding is that strategically-oriented personas (e.g. "a geopolitical strategist," "a retired intelligence officer") robustly outperform control personas (e.g. "a young child," "an uninterested student") when mapped through PI-based prompts. The PI approach yields TrueSkill scores for top personas exceeding 27, while lower-performing personas cluster near 18.

Statistical correlation analysis further emphasizes the strength of the PI method: Spearman coefficients between PI heuristic weights and features like strategicThinker and domainExpert are consistently significant ($\rho \approx 0.4-0.6$, $p < 0.005$) for GPT-4o and Llama 4. In contrast, the DH method yields nonsignificant or erratic correlations, indicating poor mapping fidelity between personas and strategic behavior.

Figure 3: Heuristic-feature correlations achieved by PI1 on GPT-4, demonstrating strong alignment between persona features and heuristic policy selection.

Replications with Mistral and Llama 3 confirm reduced effect size in smaller models, but the trend remains: PI outperforms DH in aligning strategic play with persona intent. Notably, opposite-value heuristic consistency (e.g., preference for attacking the strongest vs. weakest opponent) is far higher in PI agents developed via larger LLMs.

Stability and Interpretability

Repeated tournament runs demonstrate that PI-based persona rankings are less stable across trials than DH rankings—likely due to higher behavioral variability imparted by more differentiated heuristic choices. Nonetheless, the reproducibility of feature-performance correlations remains robust in PI, indicating that instability arises mostly from the competitive dynamics and not from failure to extract persona signal.

Figure 4: Heuristic Correlations - PI1 - LLaMA 4; PI-based agents show significantly greater dispersion and interpretability across heuristic axes relative to their DH counterparts.

Moreover, analysis of heuristic correlation matrices reveals that PI agents exhibit policy structures more congruent with human intuition: for instance, high strategicThinker scores are associated with preferences for heuristics targeting high-value objectives, risk-averse maneuvers, and enemy modeling responsive to board state contingencies.

Methodological Implications

The results demonstrate that direct persona prompting in LLMs does not reliably translate into differentiated, interpretable, or effective decision strategies in complex, adversarial settings. The introduction of a personality inventory as a translation mediator—by integrating psychometric paradigms—substantially boosts the reliability, face validity, and practical interpretability of persona-driven agents.

The contrast is especially pronounced in the fact that, with direct heuristics, persona-to-behavior mapping is largely a stylistic artifact with limited behavioral diversity; while with PI pipelines, personas instantiated with relevant strategic characteristics consistently yield higher game-theoretic efficacy.

Broader Implications and Future Directions

The incorporation of psychometric theory to operationalize persona effects in LLMs opens several new lines of inquiry in AI and cognitive modeling. In multi-agent simulation domains, inventory-mediated persona assignment could enable generative models of human-like variability and team dynamics, with applications in wargaming, training, and human-autonomy teaming. Additionally, the approach offers a template for integrating structured, theoretically-motivated behavioral priors into LLM-driven decision-making, transcending the limitations of raw prompt engineering.

The authors also note the release of both PERIL and their heuristic translation codebases, promoting replicability and extensibility for future studies in strategic reasoning and behaviorally rich simulation environments.

Figure 2: The transformation pipeline from persona description to actionable heuristic agent analyzed in the paper.

Limitations and Theoretical Constraints

While the findings support enhanced face validity and behavioral diversity via PI, ultimate claims about psychometric alignment remain limited by the lack of ground-truth human behavioral data as reference. The platform utilizes fixed heuristics and does not enable agents to adapt heuristics mid-game, restricting adaptive strategic reasoning. Additionally, the generalizability of these findings to environments with more nuanced reward structures or less well-defined action taxonomies remains open.

Conclusion

This paper provides strong evidence that persona-based prompting, when coupled with a personality inventory heuristic mapping, produces measurable, strategically relevant behavioral diversity in LLM-driven agents within adversarial environments. The absence of such a mediation layer results in largely undifferentiated behavior, undermining the promise of persona-based control via direct prompting alone. The fusion of psychometric techniques with prompt engineering thus constitutes a promising methodology for future development of adaptive, human-like AI agents in high-variance, high-stakes applications.

(2512.06867)