Enhancing Team Diversity with Generative AI: A Novel Project Management Framework

1. Problem Statement

The paper addresses a critical bottleneck in academic and scientific research projects: team homogeneity.

• Context: Research teams, particularly in universities, often consist of members with similar educational backgrounds and theoretical knowledge (students and academic staff).
• Limitations: This lack of diversity in social experience, personality, and professional roles (e.g., operators, leaders, engineers) limits the team's ability to solve complex problems and successfully commercialize research outcomes.
• Constraint: Traditional recruitment for these projects is hindered by funding limitations, project instability, and a lack of professional HR management, making it difficult to assemble the diverse, multi-role teams required for success.
• Gap in Current AI: While Generative AI (GenAI) is used in project management for data analysis and risk prediction, existing applications largely overlook the psychological and human elements (personality traits and role diversity) central to team dynamics.

2. Methodology

The authors propose a GenAI-empowered Project Management Framework that integrates a "Team Analysis" module into the traditional project lifecycle (Initiation, Planning, Execution, Monitoring, Closing).

A. The Framework Architecture

1. Team Analysis Module: Inserted before the planning phase, this module analyzes existing team members' personalities and roles.
2. Gap Identification: It compares the current team composition against "ideal" successful entrepreneurial patterns to identify missing roles and personality traits.
3. GenAI Agent Integration: GenAI agents are introduced to simulate the missing roles and personalities, effectively acting as virtual teammates to fill these gaps.

B. Technical Implementation of GenAI Agents

• Base Model: The authors selected gpt-3.5-turbo-1106 as the base model.
• Personality Theory: The system is grounded in the Big Five Personality Model (Agreeableness, Conscientiousness, Extraversion, Openness, Neuroticism).
• Fine-Tuning Strategy:
  • Dataset: The FriendsPersona dataset was used for fine-tuning. This dataset contains dialogue from the TV show Friends annotated with Big Five personality traits.
  • Pre-processing: Raw data was cleaned to remove HTML tags and character names, then converted into JSONLines format (System/User/Assistant roles) suitable for ChatGPT API fine-tuning.
  • Split: The data was split 80% for training and 20% for validation.
• Objective: To move beyond standard "role" prompts (e.g., "act as a manager") and imbue the model with deep, nuanced understanding of specific personality traits.

3. Key Contributions

• Novel Framework: Introduction of a project management framework that explicitly uses GenAI agents to simulate missing human roles and personality traits, shifting AI's role from purely quantitative data processing to qualitative team augmentation.
• Personality-Driven Fine-Tuning: Demonstration of a specific fine-tuning pipeline using the FriendsPersona dataset to equip LLMs with the ability to recognize and simulate complex personality traits, rather than just functional roles.
• Bridging Theory and Practice: The framework connects sociological theories regarding team diversity and entrepreneurial success with practical AI implementation in academic research settings.

4. Results and Analysis

The authors conducted experiments to validate the effectiveness of the fine-tuned model in recognizing personality traits.

• Training Performance: Training on the FriendsPersona dataset showed effective learning with no overfitting, though some fluctuation in accuracy was noted.
• Evaluation Dataset: The Essay Dataset (2,467 user essays annotated with Big Five traits) was used as a held-out test set to evaluate the fine-tuned model against the baseline gpt-3.5-turbo-1106.
• Performance Metrics (Precision, Recall, F1 Score):
  • Overall Improvement: Fine-tuning significantly improved the F1 Score for four out of five personality categories.
  • Specific Gains:
    • Conscientiousness: F1 Score increased from 0.237 (Baseline) to 0.547 (Fine-tuned).
    • Agreeableness: F1 Score increased from 0.285 to 0.542.
    • Openness: F1 Score increased from 0.439 to 0.536.
    • Neuroticism: F1 Score increased from 0.590 to 0.625 (with a notable Recall jump from 0.614 to 0.760).
  • Trade-off: While Extraversion saw a slight decrease in F1 score (0.358 to 0.296), the model generally achieved a better balance between identifying relevant instances (Recall) and reducing false positives (Precision).

5. Significance and Future Work

• Significance: The study validates that GenAI agents can be effectively "trained" to understand and simulate human personality traits. This suggests a paradigm shift where AI can actively participate in team composition to enhance diversity, potentially leading to more innovative and commercially viable research outcomes.
• Future Work:
  • Role Expansion: Extending fine-tuning beyond personality to specific professional roles (e.g., engineers, developers, leaders).
  • Synthetic Data: Employing synthetic datasets to train these professional roles and validating their quality.
  • Real-World Validation: Deploying these agents in actual project settings to measure their impact on project progress, output quality, and commercialization success compared to non-GenAI teams.