VCBench is the first benchmark for predicting founder success in venture capital, providing a standardized dataset of 9,000 anonymized founder profiles to address the industry's sparse and uncertain data.

It demonstrates that top LLMs, such as DeepSeek-V3 and GPT-4o, achieve far higher precision than human and market baselines, with most models outperforming expert investors. The dataset and evaluation framework minimize identity leakage and establish a reproducible, privacy-preserving standard for advancing AGI in early-stage venture forecasting.