[Paper Review] DIALECTIC: A Multi-Agent System for Startup Evaluation

Jae Yoon Bae, Simon Malberg+3 more|arXiv (Cornell University)|Feb 19, 2026

Private Equity and Venture Capital0 citations

TL;DR

DIALECTIC is an LLM-based multi-agent system that collects startup facts, builds argument-based pro/contra debates, iteratively refines arguments, and outputs decision scores to rank early-stage startup opportunities. It achieves predictive performance comparable to human VCs in backtested data while providing a ranked decision frontier.

ABSTRACT

Venture capital (VC) investors face a large number of investment opportunities but only invest in few of these, with even fewer ending up successful. Early-stage screening of opportunities is often limited by investor bandwidth, demanding tradeoffs between evaluation diligence and number of opportunities assessed. To ease this tradeoff, we introduce DIALECTIC, an LLM-based multi-agent system for startup evaluation. DIALECTIC first gathers factual knowledge about a startup and organizes these facts into a hierarchical question tree. It then synthesizes the facts into natural-language arguments for and against an investment and iteratively critiques and refines these arguments through a simulated debate, which surfaces only the most convincing arguments. Our system also produces numeric decision scores that allow investors to rank and thus efficiently prioritize opportunities. We evaluate DIALECTIC through backtesting on real investment opportunities aggregated from five VC funds, showing that DIALECTIC matches the precision of human VCs in predicting startup success.

Motivation & Objective

• Motivate early-stage venture screening under bandwidth constraints by introducing an iterative, argument-based evaluation framework.
• Develop a structured LLM-driven pipeline that collects facts, generates pro/contra arguments, critiques and refines them, and outputs investment scores.
• Enable interpretable reasoning and ranking to match or exceed human VC precision while increasing screening throughput.
• Demonstrate predictive performance on real-world VC watchlist data and analyze how evidence sources contribute to evaluation.

Proposed method

• Seeds seed questions about general company, team, product, and market and decomposes them into industry-specific sub-questions to build a hierarchical fact tree.
• An answer agent uses company features and web search to generate answers for the question tree, producing a rich fact base F.
• In the reasoning phase, generators produce K arguments per stance, critics craft critiques, evaluators assign quality scores via a 14-criteria scheme, and refiners iteratively improve arguments.
• Arguments survive through a survival-of-the-fittest process across T refinement iterations to form a positive and a negative set with equal size.
• A decision score is computed as the sum of surviving pro- and contra argument quality scores minus a threshold, and investment occurs if the score is positive.
• Hyperparameters include per-iteration survivors K_t, total iterations T, and decision threshold tau; implementation uses GPT-5-mini with tailored temperature settings.

Experimental results

Research questions

• RQ1 Can an iterative, argument-based LLM system match human VC precision in predicting startup success on historical data?
• RQ2 Does structured fact collection and dialectical reasoning improve early-stage startup screening efficiency and interpretability?
• RQ3 How do different evidence sources (general, team, product, market) contribute to the argumentation and decision quality?
• RQ4 What are optimal hyperparameters (K_T, T, tau) for balancing predictive performance and argument quality in a backtesting setting?

Key findings

• The best performing configuration uses T=2 iterations with K_T=4 arguments per side.
• DIALECTIC achieves an AUC-PR of 0.2422 on the test set, with precision comparable to human investors and GPT-IO prompting baselines.
• Prediction performance on validation data outperformed real VCs, while the test set performance was comparable to baselines.
• The method outputs a full ranked frontier rather than a single operating point, enabling threshold customization for screening capacity.
• Evidence usage shows general company and product information dominating references, with team information slightly overrepresented relative to availability.