Switching to Jev made Think-Reason-Learn up to 17× cheaper
Think-Reason-Learn, an open-source library from Vela Research and the University of Oxford, now uses TypeSafe AI's Jev in place of Google's Gemini 2.5 Flash to answer the questions behind each prediction. On twelve datasets, its Random Rule Forest method was more accurate with Jev than with Gemini on every one, and a run over all twelve cost $2.04 against an estimated $35 on Gemini.
Get Think-Reason-Learn on GitHubThink-Reason-Learn is an open-source Python library for predicting outcomes from text, developed by Vela Research, the research arm of Vela Partners, in collaboration with the University of Oxford.1 On VCBench, the first benchmark for venture capital,2 it predicts from a founder's profile whether their company will reach a $500 million exit or funding round. Each prediction is split into many small questions, and a simple statistical model combines the answers.
This post focuses on two of its methods. Random Rule Forest3 has a large language model such as Google's Gemini4 write yes-or-no questions about the text, and Policy Induction5 has it write short policies of the form "founders like this tend to succeed". These are written once for each dataset, but every new prediction means answering all of them again, so the answering model accounts for most of the running cost. Since the questions and policies stay the same, Think-Reason-Learn can switch to a cheaper or better answering model whenever one appears.
Why switch from Gemini to Jev?
In our earlier runs, Gemini 2.5 Flash was used to answer the questions and policies written by both Think-Reason-Learn methods. Jev, a model from TypeSafe AI,6 is designed for exactly this kind of question. Given a passage and a yes-or-no question, Jev returns the probability that the answer is yes. Jev writes no text, and its probability goes straight into the statistical model.
Jev is about 17× cheaper than Gemini 2.5 Flash for Random Rule Forest, and about 10× cheaper for Policy Induction, where a run costs $4.12 against an estimated $40. The Think-Reason-Learn Ensemble, which combines five of the library's methods, leads the VCBench leaderboard. With Jev answering its questions, it scores 1,000 founders for $0.87, little enough to score every founder in a deal pipeline.
Which datasets did we test on?
We test the methods on twelve datasets in which each passage of text comes with a known outcome to predict (53,550 cases in all, where a case is one passage and its outcome), spanning medicine, finance, film and more. This post shows four of the sources in detail.8 Results for every dataset, with the full methods, will appear in a forthcoming academic paper on Think-Reason-Learn.
The four sources shown
| Dataset and text | Predict | Cases | How often |
|---|---|---|---|
| Clinical trialsTrial registration | The drug passes the trial (Phases I, II and III) | 12,355 | 56%, 50%, 68% |
| Medical fundraisingCrowdfunding campaign page | Reaches its goal within 30 days | 14,961 | 21% |
| MicroloansKiva microlending listing | Expires unfunded | 5,000 | 4% |
| MoviesPlot summary | Award nomination | 2,200 | 19% |
The last column shows how often the outcome happens.
Are Jev's answers as good as Gemini's?
We score each model's answers by how often they pick the success when shown one case that succeeded and one that didn't, where random guessing would be right 50 times in 100. Both models answered exactly the same questions, written by Gemini, so the comparison isolates the model that answers the questions rather than the model that wrote them. On Random Rule Forest questions, Jev's answers were more accurate than Gemini's on all twelve datasets, at 69 on average against 64 (Figure 1). On Policy Induction policies the two are close, at 70 against 68, so Jev gives similar results for about a tenth of the cost.
Scores for the datasets shown
| Dataset | Random Rule Forest, Gemini | Random Rule Forest, Jev | Policy Induction, Gemini | Policy Induction, Jev |
|---|---|---|---|---|
| Microloans | 0.765 | 0.775 | 0.796 | 0.775 |
| Movies | 0.717 | 0.748 | 0.769 | 0.823 * |
| Medical fundraising | 0.722 | 0.733 | 0.666 | 0.668 |
| Clinical III | 0.606 | 0.698 * | 0.724 | 0.713 |
| Clinical II | 0.611 | 0.660 * | 0.673 | 0.675 |
| Clinical I | 0.566 | 0.600 * | 0.570 | 0.603 * |
| Mean, all twelve datasets | 0.639 | 0.686 | 0.683 | 0.695 |
AUC on cases held back from training, which is the share of success and failure pairs ranked correctly (multiply by 100 for the figures above). * marks a Jev lead that is statistically clear, meaning its 95% paired bootstrap interval excludes zero. None of Gemini's leads is statistically clear.
Jev's larger lead on Random Rule Forest may be because those questions are short and check a single idea (Figure 2), which suits what TypeSafe AI calls a "System One" model, after the psychologist Daniel Kahneman's fast, intuitive thinking.9
Has the founder worked as a professor or researcher in a related field?
Is the campaign written by a friend, neighbor, or co-worker on behalf of the beneficiary?
Trials targeting well-defined diseases with established diagnostic criteria, often employing existing validated therapies as comparators or in combination, tend to succeed. Trials with very broad inclusion criteria for "advanced cancer" or "unspecified solid tumors" without clear molecular or histological specificity, especially in early phases, frequently fail.
Could a free model on our own computers replace Jev?
We also tested small models that run on our own computers, which would make each answer free. None matched Jev on either method (Figure 3). The best came close to Gemini only on Random Rule Forest questions, so for now Jev is worth paying for.
Figure 3, small models on our own computers
What does this mean for VCBench?
Of the ways of answering the questions that we tested, Jev gave the best predictions per dollar, and we now use it across Think-Reason-Learn, including on VCBench. There, the leading Think-Reason-Learn entry has an F0.5 of 37.9%, as our VCBench leaderboard update describes. F0.5 is a score that balances precision and recall while weighting precision more heavily. The best frontier LLM, OpenAI's GPT-6 Sol, scores 33.7%.
We also compared Jev and Gemini 2.5 Flash on VCBench itself, using the 4,500 founders in the benchmark's public half. Jev's answers gave more accurate predictions on the Random Rule Forest questions, and the two were about as accurate on the Policy Induction policies. On both methods, Jev cost around a tenth as much as Gemini.
We are also testing newer models of this kind from OpenAI, Cloudflare and Microsoft, and will report the results in a follow-up post. Researchers and VC teams can try other answering models with Think-Reason-Learn and email their results and costs to benchmark@vela.partners.
Frequently asked questions
What is Think-Reason-Learn?
An open-source Python library for predicting outcomes from unstructured text with readable methods, including Random Rule Forest, Policy Induction, Reasoned Rule Mining, GPTree and Verifiable-RL. It is developed by Vela Research, the research arm of Vela Partners, in collaboration with the University of Oxford.
What is Jev?
A model from TypeSafe AI that answers a question about a passage of text with a probability instead of generated text. It is priced on input only, at $0.042 per million tokens.
Does Jev beat Gemini for Random Rule Forest?
On twelve datasets tested in September 2026, Jev's answers were more accurate than Gemini 2.5 Flash's on every dataset, scoring 69 out of 100 on average against 64. For Policy Induction the two were close, at 70 against 68.
Why not ask Gemini to predict the outcome directly?
Asked directly, Gemini 3.6 Flash scored higher on these datasets (74 times in 100 on average), but memory tests show it remembers outcomes on some of them, such as which films were nominated for awards, so part of that score is memory. A single verdict cannot be checked for memory. Readable questions can be, for example by hiding names in the text, and on new cases, such as a founder raising today, there is no outcome to remember. Splitting the prediction up does not remove memory entirely, since on one medical dataset the answers to some questions were partly linked to what the model remembered.
Why not train a small model instead?
Small models trained for one task often match much larger ones on text classification.10 The Think-Reason-Learn questions have no labelled answers of their own, so a small model can only be trained to copy another model's answers, and in our tests none of the small models we tried matched Jev.
How were the Gemini costs estimated?
The original Gemini runs were not tracked in full. We took the number of tokens Jev measured for each run, added a short instruction and short answers, and priced every question about a case in a single call at Gemini 2.5 Flash list prices ($0.30 and $2.50 per million input and output tokens). Our earlier runs asked one question per call, which would cost about $240 for Random Rule Forest. Figures for other models use the same token counts at their list prices in September 2026. On that basis the same Random Rule Forest run would cost roughly $65 on Google's Gemini 3.6 Flash and $350 to $900 on Anthropic's Claude Opus 5.5 or OpenAI's GPT-6 Astra, before counting the text these models write while reasoning, although we have not compared their answers with Jev's.
How was the comparison kept fair?
Jev and Gemini answered exactly the same questions and policies, which Gemini 2.5 Flash wrote, and the statistical model was refitted to each model's answers in the same way. If anything, questions written by Gemini should suit Gemini's answers. Vela Research has no investment in or partnership with TypeSafe AI.
References and notes
- Think-Reason-Learn: github.com/Vela-Research/think-reason-learn and thinkreasonlearn.com.
- Chen, Ternasky, Kwesi, Griffin, Yin, Salifu, Amoaba, Mu, Alican and Ihlamur. VCBench: Benchmarking LLMs in Venture Capital. In Intelligent Computing (Computing Conference 2026), Lecture Notes in Networks and Systems, vol. 1949, pp. 167–187. Springer, 2026. Preprint arXiv:2509.14448. The live leaderboard is at vcbench.com.
- Random Rule Forest (RRF): Interpretable and Manageable Ensembles of LLM-Generated Questions for Predicting Success from Unstructured Data. arXiv:2505.24622.
- Gemini 2.5 Flash, a large language model from Google.
- Policy Induction: Predicting Startup Success via Explainable Memory-Augmented In-Context Learning. IEEE.
- TypeSafe AI, makers of Jev.
- Jev costs are measured from usage at $0.042 per million input tokens. Gemini and other model costs are estimates; see "How were the Gemini costs estimated?" above.
- Sources for the datasets shown: clinical-trial outcomes from Fu et al., HINT: Hierarchical Interaction Network for Clinical-Trial-Outcome Predictions (Patterns, 2022); medical fundraising from Xu, Li and Zhou, MDCC: A Multimodal Dynamic Dataset for Donation-based Crowdfunding Campaigns (CIKM 2023); loan listings from Kiva; and films from the Movie-O-Label dataset.
- Daniel Kahneman. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011.
- Bucher and Martini. Fine-Tuned 'Small' LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models in Text Classification. arXiv:2406.08660, 2024.