Opening — Why this matters now
Venture capital has always been a strange mix of narrative craft and network math. Partners talk about vision, conviction, and pattern recognition, but behind the scenes, outcomes are brutally skewed: most startups fail quietly, a few dominate returns, and almost everything depends on who backs whom, and in what order.
As LLMs enter financial decision-making, the obvious temptation has been to throw graphs, documents, and pitch decks into a giant black box and hope for better predictions. That approach mostly fails. Graph neural networks see structure but cannot explain it. LLMs explain well but choke on graph structure. The paper behind MIRAGE-VC asks a sharper question: What if we taught LLMs to reason over venture networks the way experienced VCs already do—selectively, narratively, and with intent?
Background — Context and prior art
VC prediction has traditionally lived in three camps:
| Approach | What it does well | Where it breaks |
|---|---|---|
| Firm-level ML | Scales cleanly | Ignores relational context |
| Graph Neural Networks | Capture investor–startup structure | Opaque, non-explanatory |
| Text + RAG with LLMs | Coherent narratives | Treats graphs as flat text |
Recent Graph-RAG hybrids improve retrieval but still assume in-graph objectives: answering questions whose answers already exist as nodes or edges. VC success prediction is fundamentally off-graph. The outcome—future funding or exit—does not exist in the network. The graph is evidence, not the answer.
This distinction turns retrieval into the real problem. If you retrieve everything, the LLM drowns. If you retrieve nothing, it hallucinates. MIRAGE-VC’s core contribution is realizing that path selection is the bottleneck, not model size.
Analysis — What the paper actually does
MIRAGE-VC reframes graph traversal as a decision problem: Which next node reduces uncertainty about startup success the most?
1. Information-gain-driven path retrieval
Instead of enumerating thousands of possible 3–5 hop paths, MIRAGE-VC expands the graph greedily, hop by hop. At each step, it evaluates candidate neighbors by task-specific information gain:
- Does adding this investor or company improve the LLM’s prediction?
- Does it increase confidence in the correct direction?
This mirrors decision-tree logic applied to graphs. Crucially, the expensive LLM-based gain signals are computed offline, and a lightweight selector model learns to approximate them at inference time. The result is compact, high-signal investment chains that fit comfortably into an LLM context window.
2. Multi-agent evidence decomposition
VCs do not reason from one lens. Neither does MIRAGE-VC. It splits analysis into three specialist agents:
- Peer-Company Analyst — looks at historically similar startups
- Investor Profile Analyst — evaluates lead investor track record
- Investment Chain Analyst — reasons over the retrieved graph paths
Each agent produces an independent verdict and rationale, using the same frozen LLM. Differences in output therefore reflect evidence, not model variance.
3. Adaptive gating instead of naïve fusion
Not all evidence matters equally for every startup. A deep-tech spinout leans on investor credibility; a consumer app leans on peer trajectories. MIRAGE-VC learns this explicitly using a gating network that assigns instance-specific weights to each agent’s reasoning.
This step is subtle but important: rather than averaging signals, the model learns when to trust each perspective.
Findings — Results that actually matter
The performance gains are concentrated where real investors care: the top of the ranking.
| Model | AP@5 | Precision | F1 |
|---|---|---|---|
| Best GNN baseline | ~29 | ~22 | ~34 |
| Best LLM-RAG baseline | ~30 | ~23 | ~35 |
| MIRAGE-VC | 34.3 | 24.3 | 36.5 |
Two patterns stand out:
- Precision rises as K shrinks — the model is better at surfacing the few deals that matter.
- Deeper retrieved paths correlate with correctness — but only when those paths are filtered by information gain.
In other words, MIRAGE-VC does not win by seeing more of the graph. It wins by seeing the right parts.
Implications — Why this extends beyond venture capital
The architectural lesson generalizes cleanly:
- In recommendation systems, the target is future preference—not a node.
- In credit risk, the target is default—not a transaction.
- In compliance and assurance, the target is violation—not a rule edge.
All are off-graph prediction problems. MIRAGE-VC suggests a reusable pattern:
Use graphs as hypothesis generators, not feature dumps.
Select evidence by marginal utility, reason explicitly, and fuse perspectives adaptively.
For businesses deploying agentic AI systems, this is a warning against brute-force retrieval and a blueprint for scalable interpretability.
Conclusion — Calm intelligence beats maximal context
MIRAGE-VC does not make LLMs smarter by feeding them more data. It makes them selective. By combining information-gain-driven graph traversal, role-specialized agents, and adaptive fusion, the system behaves less like a classifier and more like a disciplined investment committee.
That is the quiet insight here: the future of AI decision-making is not bigger context windows, but better judgment about what deserves attention.
Cognaptus: Automate the Present, Incubate the Future.