Policy Induction: Teaching an LLM an Investment Policy in Plain English

Policy Induction teaches a language model an investment policy written in plain English, then uses that policy to predict which early-stage founders will succeed. Every decision can be read, audited and edited by a person.

Why a readable policy

Early-stage investing runs on scarce data and uncertain outcomes. Traditional machine learning needs large labeled datasets, and the models it produces are hard for investors to question or improve. Policy Induction takes a different route. It keeps the decision logic outside the model, in a structured set of plain-text heuristics that sits in the prompt. The language model applies the policy to each founder and answers with a single word, True or False.

Because the policy is text, an investor can read exactly why the model backs a founder, change a rule they disagree with, and see the effect on the next run. That makes Policy Induction a natural fit for AI-native venture capital, where models make venture capital predictions on every founder before a partner ever takes a meeting.

How Policy Induction works

The method was introduced by Xianling Mu (University of Oxford) with Joseph Ternasky, Fuat Alican and Yigit Ihlamur of Vela Research, the research arm of Vela Partners. It has three steps.

1. Draft an initial policy. The model reads 20 successful and 20 unsuccessful founders and drafts a first set of rules, which an expert then edits.
2. Refine it by in-context learning. The model is shown the current policy and a new founder with a known outcome, and asked to refine and expand the policy. Updates run either one example at a time, keeping whichever policy scores better, or in parallel, keeping the top 10% of candidate policies and merging them. The policy itself acts as the model's memory from one round to the next.
3. Add reflections. The model writes a one-sentence reflection on why representative founders succeeded or failed, and those lessons are folded back into the policy. Experts can reorder or rewrite rules at any point.

Candidate policies are scored by their precision on the training set. The best policy in the paper was trained on 120 successes and 120 failures over four rounds, using GPT-4o mini and a few dollars of compute. Its rules cover themes such as verified patents, audited outcomes, investor validation and whether a founder's record is consistent across sources.

What the paper found

The preprint (arXiv:2505.21427, May 2025) evaluated policies on founders from US companies founded in 2010 or later, where success means an IPO or acquisition above $500M, or more than $500M raised.

• Without a policy, GPT-4o mini reached 13.7% precision on a test set with a 9% base rate. With the best learned policy it reached 64.5%.
• On test sets with a realistic 1.9% base rate, a policy generated by o3 and applied by GPT-4o mini averaged 40.5% precision. That is more than 20× random selection and 7.1× the 5.6% precision of tier-1 venture firms.

A later version, published at IEEE CSCloud 2025 in New York with Rick Chen (University of Oxford) as an additional author, weights and combines several policies into a policy set using the model's log-probability outputs, and evaluates on VCBench. It reports precision more than 7.4× better than chance and 3× tier-1 VCs, with the highest F_0.5 of all baselines.

The authors are open about the limits. Training is nondeterministic and sensitive to prompt wording, results vary across data segments, and even anonymized data may be partly familiar to a large model.

Where Policy Induction stands on VCBench today

Policy Induction is part of the open-source Think-Reason-Learn library from Vela Research and the University of Oxford, alongside Random Rule Forest and Reasoned Rule Mining.

Read the paper

• Policy Induction: Predicting Startup Success via Explainable Memory-Augmented In-Context Learning (arXiv preprint, May 2025)
• IEEE CSCloud 2025 version (November 2025)
• Policy Induction on Vela Research