CrunchLLM: Multitask LLMs for Structured Business Reasoning and Outcome Prediction.

Area of Science:

• Entrepreneurship and Innovation Research
• Computational Social Science
• Machine Learning Applications

Background:

• Startup success prediction is vital for venture capital and policy.
• Leveraging heterogeneous data (structured and unstructured) from sources like Crunchbase remains a challenge.
• Traditional machine learning models and standard large language models (LLMs) have limitations in predicting entrepreneurial outcomes.

Purpose of the Study:

• To develop a domain-adapted LLM framework (CrunchLLM) for predicting startup success.
• To effectively integrate structured business attributes with unstructured textual data for enhanced prediction.
• To improve the reliability, interpretability, and accuracy of startup success predictions.

Main Methods:

• Developed CrunchLLM, a domain-adapted and backbone-agnostic LLM framework.
• Integrated structured company data with unstructured text using parameter-efficient fine-tuning and prompt optimization.
• Introduced a self-verifiable multitask objective with justification loss and hierarchically ordered input encoding.

Main Results:

• Achieved 89% accuracy on the Crunchbase startup success prediction task.
• Significantly outperformed traditional classifiers and baseline LLMs.
• Generated interpretable reasoning traces, enhancing prediction transparency.

Conclusions:

• Domain-aware LLM adaptation and structured-unstructured data fusion advance predictive modeling of entrepreneurial outcomes.
• CrunchLLM provides a practical tool for data-driven decision-making in venture capital and innovation policy.
• The framework enhances transparency and trustworthiness for financial and policy decision-makers.