Predicting Corporate Culture for M&A: Leveraging Fine-Tuning and Chain-of-Thought Strategies with LLMs

This study explores a novel approach to assessing cultural fit during the early stages of mergers and acquisitions (M&A) by leveraging publicly available employee review data and large language models (LLMs). Recognizing the limitations of traditional due diligence in accessing internal cultural data, the proposed framework utilizes fine-tuned models and chain-of-thought (CoT) reasoning strategies to infer corporate cultural characteristics based on the Denison and Ko [2016] framework. The model evaluates four key traits-Mission, Consistency, Involvement, and Adaptability-across twelve dimensions, using Low-Rank Adaptation (LoRA) for efficient fine-tuning. Experimental results demonstrate that LoRA-tuned models consistently outperform few-shot prompting across both proprietary (e.g., GPT-4o) and open-source (e.g., Llama 3.2-3B) models, with significant improvements in both text summarization and numerical prediction accuracy. Additionally, CoT reasoning-particularly Multi-step and Hybrid strategies-yields substantial performance gains, especially in smaller models, enabling them to approximate the results of large-scale systems at reduced cost. These findings highlight the practical utility of combining PEFT and CoT methods for scalable, objective, and early-stage cultural assessments in M&A decision-making.