DIALECTIC: A Multi-Agent System for Startup Evaluation

A research team led by Jae Yoon Bae and Simon Malberg has introduced DIALECTIC, a novel multi-agent AI system designed to automate and scale the early-stage screening of startup investment opportunities. The system uses large language models (LLMs) to first gather and structure factual knowledge about a startup into a hierarchical question tree. It then synthesizes this information into natural-language arguments both for and against an investment, initiating a simulated debate where AI agents critique and refine these arguments iteratively. This process surfaces only the most convincing points and culminates in a numeric decision score, allowing venture capitalists to rank and prioritize opportunities efficiently.

The core innovation is the system's ability to replicate a thorough, dialectical due diligence process autonomously. The researchers evaluated DIALECTIC through rigorous backtesting on real investment opportunities aggregated from five venture capital funds. The results, accepted for publication at EACL 2026, show that the AI system's predictions of startup success matched the precision of human VC investors. This demonstrates a significant step toward automating the high-bandwidth task of initial deal screening, potentially freeing up investor time for deeper analysis and relationship building while ensuring no strong opportunity is overlooked due to human capacity limits.