Technical Summary: Persuasive and Compliant Tendencies Predict Group Decision-Making in Humans and Language Models

Problem Statement

Large Language Models (LLMs) are increasingly deployed in group decision-making scenarios alongside other LLMs and humans. However, the mechanisms driving their influence remain unclear. Specifically, it is unknown whether an LLM's impact is primarily driven by persuasion-oriented expression (active argumentation and assertiveness) or compliance-oriented accommodation (listening and adapting to others). While social network research offers two perspectives on influence—those with broad outgoing communication channels versus those who aggregate incoming information—it is not established which behavioral orientation yields better group outcomes in LLM systems. Furthermore, there is a need to determine if these intrinsic behavioral tendencies are measurable and if they correlate with task performance, particularly in settings involving asymmetric information and hidden objectives.

Methodology

The authors propose a two-stage framework to isolate and evaluate these tendencies:

1. DecisionQE: Measuring Intrinsic Tendencies

To separate individual behavioral tendencies from the confounding factors of interactive group dynamics (such as role assignment), the authors developed DecisionQE, a questionnaire-based framework.

• Structure: The framework assesses models across six decision-related domains: public communication, everyday decision-making, marketing and persuasion, presentation and expression, interpersonal communication, and negotiation and strategic interaction.
• Scoring: It utilizes 70 items derived from psychological literature. Responses are scored on a continuum from 0 (compliant) to 100 (persuasive).
• Stability: Models were evaluated over five repeated runs to ensure the stability of their tendency profiles.

2. Werewolf Game: Interactive Testbed

The authors utilized the Werewolf game as an interactive testbed to validate how these measured tendencies affect social influence and group outcomes under asymmetric information.

• Setup: Games involved eight players (two werewolves, four villagers, one witch, one seer). Werewolves knew their teammates and aimed to eliminate villagers; villagers aimed to identify and eliminate werewolves.
• Variables: The study manipulated role assignments to create specific configurations:
  • Random Allocation: To assess general performance correlation.
  • Controlled Allocation: Six configurations (Reverse, High, Low) assigned informative roles (Werewolf, Seer, Witch) to models at different positions on the persuasive-compliant continuum to test role-dependent effects.
• Human-LLM Integration: To verify behavioral consistency, the authors conducted experiments where a human participant replaced an LLM in either a "good" (villager/seer/witch) or "wolf" role, interacting with seven LLMs.

Key Contributions

1. Quantification of Behavioral Tendencies: The paper introduces a standardized method (DecisionQE) to measure and categorize LLMs along a persuasive-compliant spectrum, demonstrating that different models occupy distinct, stable positions on this continuum.
2. Decoupling Tendency from Performance: The study establishes that higher persuasive scores do not automatically translate to better group outcomes. Instead, compliant-oriented models often demonstrate superior performance in cooperative tasks.
3. Dual-Effect of Compliance: The research reveals a "dual-sided" nature of compliance: it supports cooperation when models are in honest roles (villagers) but enhances concealment and success when models are in adversarial roles (werewolves).
4. Human-LLM Consistency: The findings show that the role-dependent patterns observed in LLM-only groups persist in mixed human-LLM interactions, suggesting these tendencies are a shared behavioral axis for analyzing both artificial and human-mediated decision-making.

Results

Tendency Profiles

• Variation: Evaluated models showed substantial variation in their tendency scores. For instance, Kimi K2 and Gemini 3.5 scored highest (most persuasive), while Doubao 1.5 and GLM 4 scored lowest (most compliant).
• Stability: Within-model standard deviations were low (0.6–2.1 points), indicating stable behavioral profiles across repeated evaluations.

Game Performance (Random Allocation)

• Win Rates: Models with stronger persuasive tendencies achieved lower overall win rates (40.28%–51.39%). Conversely, moderately compliant and compliant models achieved higher win rates (54.17%–56.94%).
• Survival and Identification: Compliant models survived longer (3.10–3.24 rounds vs. 2.94–3.15) and achieved higher role identification accuracy (43.75%–52.78% vs. 35.42%–41.67%).

Role-Dependent Outcomes

• Good Team Success: Configurations where persuasive models were assigned to werewolf roles (forcing them to be aggressive) and compliant models to good roles (H1, R1) resulted in the highest good-team win rates (90.00%).
• Wolf Team Success: Configurations where compliant models were assigned to werewolf roles (R2, L1) resulted in the highest werewolf-team win rates (45.00%).
• Interpretation: Persuasive behavior increases visibility and initiative, which aids in mobilizing a group but attracts scrutiny in hidden-role games. Compliant behavior allows models to blend in, facilitating coordination in honest roles but improving concealment in adversarial roles.

Human-LLM Experiments

• Behavioral Alignment: Human participants generally scored closer to the compliant end of the spectrum than the evaluated LLMs, though with higher variance.
• Outcome Consistency: When humans played good roles, the team win rate (70%) and survival days closely matched the H1 (LLM-only) condition. When humans played wolf roles, the wolf team win rate (80%) aligned with the R2 condition, confirming that the role-dependent patterns hold in mixed groups.

Significance and Claims

The paper claims that LLM group interactions serve as a controllable observational setting for sociological research, revealing measurable patterns of intrinsic behavioral tendencies that extend beyond simple task reasoning.

• Safety Implications: The authors argue that safety evaluations must incorporate intrinsic behavioral tendencies. While aggressive or manipulative outputs are easily flagged, compliant-oriented models may pose a unique risk in adversarial settings. Their low-salience, cooperative style can mask unsafe objectives, making them harder to detect during screening.
• Evaluation Shift: The findings suggest that trustworthy LLM systems require "tendency-aware" safety evaluations. Models should be assessed not only on the safety of isolated outputs but on how their intrinsic tendencies (persuasive vs. compliant) affect the visibility and detectability of risks across interactive, multi-turn settings.
• Sociological Lens: The study posits that LLMs can serve as a lens for observing language-mediated interaction, demonstrating that "influential listeners" (compliant agents) can shape group outcomes as effectively as, or more effectively than, active persuaders, depending on the context and objectives.