# Think-Reason-Learn (TRL)

> Think-Reason-Learn (TRL) is the decision harness for predictions: an open-source Python library from Vela Research and the University of Oxford that joins machine learning and language models, so every prediction is accurate, explained in plain words and cheap to run.

A language model reads each case and answers plain-language questions; TRL learns from past outcomes which questions matter and by how much. The result is language-model judgment at machine-learning speed, with the reasons behind every prediction.

Key facts:

- #1 on VCBench, a benchmark for predicting which founders succeed, scored on 4,500 held-out founders (F0.5, as of 29 September 2026): TRL Ensemble 37.9, GPT-6 Sol 33.7, best machine learning model 31.5, tier-1 VCs 10.7. That is 12% more accurate than GPT-6 Sol and 20% more accurate than the best machine learning model.
- About $0.00087 per prediction on VCBench, returned in milliseconds: a small classifier answers the learned questions in parallel, so no large model reasons about each case.
- Learns from about 50 examples. Reads text, transcripts, video and documents, which machine learning cannot use without hand-built features.
- Six methods: GPTree, Random Rule Forest, Policy Induction, Reasoned Rule Mining, Verifiable RL and Feature Generation.
- Open source under the MIT licence, Python 3.13 or newer. Install from GitHub: `pip install "git+https://github.com/vela-research/think-reason-learn.git"`. Works with Anthropic, Google, OpenAI, xAI and TypeSafe models.
- Built by a research lab within Vela Partners and the University of Oxford. Methods accepted at IEEE and a NeurIPS 2026 workshop.

## Site

- [Home](https://thinkreasonlearn.com/): the decision harness for predictions, how it compares with machine learning and language models, and the six methods
- [Models](https://thinkreasonlearn.com/models): pre-trained models built with TRL, open and private, and co-building a model
- [FounderBrain](https://thinkreasonlearn.com/founderbrain): a free model for your AI agent that analyses how you came across in a pitch or investor call, compared with 773 public interviews of founders and CEOs; transcripts are not stored
- [Benchmarks](https://thinkreasonlearn.com/benchmarks): VCBench results for every TRL method against frontier language models, machine learning and tier-1 VCs, and TranscriptBench
- [Research](https://thinkreasonlearn.com/research): the papers behind the methods, and posts about the work
- [Think-Reason-Learn tests three new models and keeps Jev](https://thinkreasonlearn.com/posts/decision-models-compared): on twelve datasets OpenAI's GPT-6 Luna Decisions was about as accurate as TypeSafe AI's Jev at almost five times the cost, while Cloudflare's Clef Flash and Microsoft-Decision-1 were cheaper but less accurate
- [Switching to Jev made Think-Reason-Learn up to 17× cheaper](https://thinkreasonlearn.com/posts/jev-vs-gemini): on twelve datasets Random Rule Forest was more accurate with TypeSafe AI's Jev than with Gemini 2.5 Flash on every one, and a run cost $2.04 against an estimated $35
- [Team](https://thinkreasonlearn.com/team): the research lab within Vela Partners and the University of Oxford
- [Privacy](https://thinkreasonlearn.com/privacy): how FounderBrain handles transcripts

## Docs

- [Documentation](https://thinkreasonlearn.com/docs/): start here
- [Installation guide](https://thinkreasonlearn.com/docs/getting_started/installation_guide.html)
- [Quick start](https://thinkreasonlearn.com/docs/getting_started/quick_start.html): fit a small GPTree on six founder profiles
- [API reference](https://thinkreasonlearn.com/docs/modules.html)
- [GPTree](https://thinkreasonlearn.com/docs/think_reason_learn.gptree.html): decision trees whose questions a language model writes
- [Random Rule Forest](https://thinkreasonlearn.com/docs/think_reason_learn.rrf.html): many yes-or-no questions combined into one transparent model
- [Policy Induction](https://thinkreasonlearn.com/docs/think_reason_learn.policy_induction.html): short policies learned from examples
- [Reasoned Rule Mining](https://thinkreasonlearn.com/docs/think_reason_learn.reasoned_rule_mining.html): if-then rules with calibrated confidence
- [Verifiable RL](https://thinkreasonlearn.com/docs/think_reason_learn.verifiable_rl.html): learns what to check next, and stops when it is sure
- [Feature Generation](https://thinkreasonlearn.com/docs/think_reason_learn.features.html): rules written as small Python functions that run without a model

## Optional

- [Changelog](https://thinkreasonlearn.com/docs/changelog.html)
- [Source code on GitHub](https://github.com/vela-research/think-reason-learn)
- [VCBench leaderboard](https://vcbench.com)