DIALECTIC: A Multi-Agent System for Startup Evaluation

Jae Yoon Bae; Simon Malberg; Joyce Galang; Andre Retterath; Georg Groh

cs.MAFeb 19, 2026

DIALECTIC: A Multi-Agent System for Startup Evaluation

Authors: Jae Yoon Bae, Simon Malberg, Joyce Galang, Andre Retterath, Georg Groh

Organizations: Technical University of Munich · Earlybird Venture Capital · UVC Partners

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.

Figures & tables

Figure 1: Overview of the DIALECTIC method. The right side shows the flow of operations. Agents are shown in red, agent inputs/outputs are shown in blue, and loops are shown in green. The left side illustrates the key outputs of the agents.

Figure 2: Results from the hyperparameter optimization, showing AUC-PR, raw argument scores, QA pair count, and argument length for different numbers of arguments

K_{t}

and iterations

T

Figure 3: Precision and recall of DIALECTIC across all possible values of the decision threshold

\tau

in comparison to the human VCs and GPT IO prompting baseline.

Figure 4: Distribution of decision scores.

Aspect	Usage	Availability	Ratio
General	34.40%	34.88%	0.986
Team	20.52%	20.17%	1.017
Product	29.77%	29.43%	1.011
Market	15.31%	15.51%	0.987

Table 1: Utilization of factual evidence in arguments. Aspect refers to the question trees built for the four seed questions. Usage measures the share of all cited facts, availability measures the relative size of question trees, and ratio is the ratio of usage and availability.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Data Field	Description	Value As Of
Watchlist	Name F	The company name.	Some time in 2021
	Domain F	The company web domain.	Some time in 2021
	Date Added P	The date the company was added to the watchlist. Only companies added between January 1, 2021 and December 31, 2021 are considered.	Some time in 2021
	Status P	The last stage reached in the investment process (e.g., Added to Watchlist , Initial Review , or Investment Made ). Used to determine whether the real VC invested in the company.	Some time in 2021
Crunchbase	Funding Rounds P	A list of all funding rounds, including round type ( pre-seed , seed , series A , IPO , etc.), amount, and announcement date of the round.	September 1, 2025
	Current Name P	The company’s current name.	September 1, 2025

Appendix

Table 2: An overview of the data sources and extracted data fields used for the dataset creation. Data fields marked with F are used as predictive features (independent variables), whereas data fields marked with P are only used during preprocessing (e.g., for merging data from different sources) and were not made available to the prediction method.

Preprocessing Stage	Companies	Success Rate (%)
Export from CRM system	3,441	—
Remove duplicates and missed deals	3,404	—
Remove series A investments by VC funds	3,401	—
Match with current funding data	2,192	21.6
Remove companies without cutoff date	1,715	23.4
Remove companies added post–seed announcement	587	18.7

Appendix

Table 3: Dataset size and success rate after successive preprocessing steps. The success rate refers to the share labeled successful at the corresponding stage.

Work	Success Criterion	Accuracy	Recall	Precision	Interpretability
Arroyo et al. (2019)	First event in 3 yrs (AC = acquired, FR = funding round, IPO, CL = closed, NE = no event)	Global 82.2%	FR: 40%, AC: 3%, IPO: very low, NE: 95%	FR: 64%, AC: 33%, IPO: 44%, NE: 85%	Feature-based
Żbikowski and Antosiuk (2021)	AC, IPO, Series B	85%	34%	57%	Feature-based
Retterath (2020)	Follow-on round, trade sale, IPO	80%	80%	–	No mention
Antretter et al. (2019)	5-year survival	76%	86%	80%	Feature-based
Sharchilev et al. (2018)	Series A+ within 1 yr	–	–	62.6%	Feature-based
Gavrilenko et al. (2023)	Raise Series A+ within 1 yr	–	82.7%	74.4%	Feature-based

Appendix

Table 4: Comparison of startup success prediction studies: success criteria, predictive performance of the best model, and interpretability.

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,039 papers

DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage
Oct 23, 2026Ming Ze Tang, David Cutting, Boyd Carson +1Financial Language Model Evaluation LLM Agent Evaluation
Queen’s University Belfast, Belfast, United Kingdom · Sapphire Capital Partners, Belfast, United Kingdom
A Multi-Agent Orchestration Framework for Venture Capital Due Diligence
May 13, 2026Grigorios Alexandrou, Katerina PramatariMulti-Agent LLM Systems LLM Agent Orchestration
Department of Management Science and Technology, Athens University ofMay Economics and Business (AUEB), Athens, Greece
Predicting Startup Exit from Textual Descriptors - A Computational Linguistics Framework
Jul 24, 2026Alberto M. G. Saruggia, Sebastien Germano
Independent Researcher MBA Candidate, Boston University · Independent Researcher MSc Candidate in Data & AI, ECE Paris

Oct 23, 2026J/K move · Enter open · S save

DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage

Ming Ze Tang, David Cutting, Boyd Carson, Vasiliki Carson

Queen’s University Belfast, Belfast, United Kingdom · Sapphire Capital Partners, Belfast, United Kingdom

Financial Language Model Evaluation LLM Agent Evaluation Human-in-the-Loop Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) can read startup pitch decks, but letting them emit investment scores directly invites score inflation, run-to-run variance and rationales that are hard to audit. We describe DealsFlow, a human-in-the-loop agentic harness deployed at a UK early-stage fund. The LLM reads the deck, cites evidence and rates each category; code combines the ratings, caps them and only ever lowers them. Three mechanisms implement this: a stage-aware corroboration ceiling (78, 70 or 60 out of 100) on factual categories without external evidence; a sceptic call that may lower but never raise a score; and a pre-fetch pattern that retrieves and indexes all evidence before a single generation pass. We evaluate on

N=243

SEIS/EIS investments with recorded outcomes. Because recent investments have unresolved outcomes, we use the 142 that are at least three years old. There, the full pipeline's score has Spearman

ρ=0.45

(95% CI 0.32 to 0.57) with the total-value multiple and AUC 0.71 for identifying write-offs, against

ρ=0.38

and AUC 0.68 for an unconstrained LLM. The gain is small and comes mostly from the sceptic call; we did not test whether its lower-only rule matters. Investments that later exceeded

2\times

had 64% of factual categories corroborated, against 38% for write-offs (omnibus Kruskal-Wallis test over five outcome groups,

p=0.031

), and our retrieval was not date-restricted, so these figures may overstate screening skill. In a study of 20 inbound applications, analyst time per application fell from 44.7 to 18.9 minutes including verification and correction, and 84% of 240 cited claims were fully supported by their sources, but 70% of reports needed at least one correction.

Abstract

Multi-Agent LLM Systems

Explore similar work

DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage

A Multi-Agent Orchestration Framework for Venture Capital Due Diligence

Predicting Startup Exit from Textual Descriptors - A Computational Linguistics Framework