The Gaining Paths to Investment Success: Information-Driven LLM Graph Reasoning for Venture Capital Prediction

TL;DR

MIRAGE-VC tackles off-graph venture-capital prediction by learning information-gain-driven, multi-hop graph paths and fusing heterogeneous evidence with a gated, multi-agent LLM framework. The approach distills large investment networks into high-value chains that support explicit, step-by-step reasoning, while preserving interpretability through per-view rationales and weighted fusion. Empirical results on PitchBook data show significant gains in AP@K, F1, and Precision over strong baselines, with ablations demonstrating the necessity of graph retrieval and multi-view integration. The method generalizes beyond VC to other off-graph tasks like recommendations and risk assessment, offering a principled way to perform reasoning over complex relational evidence. Practical impact includes better identification of high-potential startups under resource constraints, with transparent rationale suitable for VC practice.

Abstract

Most venture capital (VC) investments fail, while a few deliver outsized returns. Accurately predicting startup success requires synthesizing complex relational evidence, including company disclosures, investor track records, and investment network structures, through explicit reasoning to form coherent, interpretable investment theses. Traditional machine learning and graph neural networks both lack this reasoning capability. Large language models (LLMs) offer strong reasoning but face a modality mismatch with graphs. Recent graph-LLM methods target in-graph tasks where answers lie within the graph, whereas VC prediction is off-graph: the target exists outside the network. The core challenge is selecting graph paths that maximize predictor performance on an external objective while enabling step-by-step reasoning. We present MIRAGE-VC, a multi-perspective retrieval-augmented generation framework that addresses two obstacles: path explosion (thousands of candidate paths overwhelm LLM context) and heterogeneous evidence fusion (different startups need different analytical emphasis). Our information-gain-driven path retriever iteratively selects high-value neighbors, distilling investment networks into compact chains for explicit reasoning. A multi-agent architecture integrates three evidence streams via a learnable gating mechanism based on company attributes. Under strict anti-leakage controls, MIRAGE-VC achieves +5.0% F1 and +16.6% PrecisionAt5, and sheds light on other off-graph prediction tasks such as recommendation and risk assessment. Code: https://anonymous.4open.science/r/MIRAGE-VC-323F.

Figures (5)

• Figure 1: Impact of the selected chain on prediction.
• Figure 2: Overall Framework of MIRAGE-VC. It contains four key components: Graph Retrieval for path selecting, Text Retrieval for companies and investor, Multi-agent Analysis for multi-perspective information, Dynamic Weighted Fusion for adaptive information fusion.
• Figure 3: Illustration of how the path selector retrieves the best path from graph
• Figure 4: (a) Text retrieval: effect of the number of similar companies $K$ and resume breadth $N$ on the agents’ F$_1$. (b) Graph retrieval: effect of search depth $d_{\max}$ and path count $P$ on F$_1$.
• Figure 5: Distribution of path retriever maximum hop length by prediction outcome