Can GPT solve Brazilian university entrance exams?
This project is an implementation of how to use LLMs to solve challenging Brazilian university entrance exams.
We started with o1-preview, and gpt-4o to describe the exam images so that o1-preview could solve them, because it did not have image capabilities yet. Now that reasoning models are processing images as well, we've updated the code to process them directly. Also, we've started solutions with o1 and hope to add other models as well. Results are saved as txt files with LaTeX formatting, and you can optionally convert them to a nice PDF with this package or using some LaTeX editor.
The project begins with the ITA (Instituto Tecnológico de Aeronáutica) 2025 exam, focusing first on the Math essay section. This section, from the recent exam on November 5, 2024, demands deep subject understanding and step-by-step solutions. More details are in the report documentation.
Spoilers: o1-preview scored 90% in the ITA 2025 Math Essay Exam, 95% in Chemistry Essay and only 65% in Physics Essay. o1 scored 95% in the same Math exam.
After the first ITA 2025 exam is fully solved, the project will try to expand to other sections and eventually other exams. Feel free to contribute with ideas and implementations of other exams!
Table of some exams to be solved:
| Exam | Year | Status | Score | Report |
|---|---|---|---|---|
| ITA | 2025 | 🚧 In Progress | - | Report |
| IME | 2025 | 📅 Todo | - | - |
| Escola Naval | 2025 | 📅 Todo | - | - |
| Fuvest/USP | 2025 | 📅 Todo | - | - |
| AFA | 2025 | 📅 Todo | - | - |
| UNICAMP | 2025 | 📅 Todo | - | - |
| ENEM | 2025 | 📅 Todo | - | - |
gpt-resolve is distributed in pypi:
pip install gpt-resolvegpt-resolve provides a simple CLI with two main commands: resolve for solving exam questions and compile-solutions for generating PDFs from the solutions, as you can check below. During development, install locally with pip install -e ..
To generate solutions for an exam:
- save the exam images in the exam folder
exam_path, one question per image file. File names should follow the patternq<question_number>.jpg, e.g.q1.jpg,q2.jpg, etc. - add
OPENAI_API_KEYto your global environment variables or to a.envfile in the current directory
then, run
gpt-resolve resolve -p exam_pathand grab a coffee while it runs.
If you want to test the process without making real API calls, you can use the --dry-run flag. See gpt-resolve resolve --help for more details about solving only a subset of questions or controlling token usage or changing the model.
Once you have the solutions in your exam folder exam_path, you can compile them into a single PDF running:
gpt-resolve compile-solutions -p exam_path --title "Your Exam Title --author 'o1 (OpenAI)'"For that command to work, you'll need a LaTeX distribution in your system. See some guidelines here (MacTeX for MacOS was used to start this project). This project was developed with MacTeX distribution.
Sometimes, it was observed that the output from o1-preview produced invalid LaTeX code when nesting display math environments (such as \[...\] and \begin{align*} ... \end{align*} together). The current prompt for o1-preview adds an instruction to avoid this, which works most of the time. If that happens, you can try to solve the question again by running gpt-resolve resolve -p exam_path -q <question_number>, or making more adjustments to the prompt, or fixing the output LaTeX code manually. Experiments with o1 did not show any similar errors so far.
Reasoning models like o1 are much more expensive than non-reasoning models like gpt-4o, consuming much more tokens to "reason" (see more here), so be mindful about the number of questions you are solving and how many max tokens you're allowing gpt-resolve to use (see gpt-resolve resolve --help to control max-tokens-question-answer, which drives the cost). You can roughly estimate an upper bound for costs of solving an exam by
(number of questions) * (max_tokens_question_answer / 1_000_000) * (price per 1M tokens)
For the current price for o1 of $15/$60 per 1M tokens for input/output tokens, an 10 question exam with 10000 max tokens per question would cost less than $6. For o3, for which the price is $10/$40 per 1M tokens for input/output tokens, the cost would be less than $4.
There are several ways you can contribute to this project:
- Add solutions for new exams or sections
- Improve existing solutions or model prompts
- Add automatic evaluation metrics for answers
- Create documentations for exams
- Report issues or suggest improvements
To contribute, simply fork the repository, create a new branch for your changes, and submit a pull request. Please ensure your PR includes:
- Clear description of the changes
- Updated table entry (if adding new exam solutions)
- Any relevant documentation
Feel free to open an issue first to discuss major changes or new ideas!
This project is proudly sponsored by Buser, Brazil's leading bus travel platform.