Meta Hacker Cup (formerly known as Facebook Hacker Cup) is an annual international programming competition hosted and administered by Meta. A lot of diverse participants take part in the competition from different countries worldwide. This year however they did allow the usage of AI within the competition, designed to assess the capabilities of Generative AI in performing autonomous code generation tasks.

The competition featured an AI vs human battle i.e Gen AI or AGIs could be utilised within the coding sphere to see how they rank alongside humans.

We developed a model designed for fully autonomous coding. It works by combining outputs from multiple LLMs to solve coding problems without human input. The model breaks tasks into smaller parts, assigns them to the best-suited LLM, and refines the outputs. It evaluates solutions through automated testing and improves by learning from feedback. The system uses a voting mechanism to choose the best output and runs checks for errors and performance issues. Test cases provided by the problems in hacker cup are also generated based on the problem statement and are made to run against the generated code to check for the accuracy.

This model was a top-10 finalist in Meta's Hacker Cup AI track. It demonstrates a new way of using LLMs to perform end-to-end coding tasks. We will share insights from building and testing the model, including its strengths and challenges. We believe that this can be highly useful for students who don't have the privilege of learning competitive coding by paying high fees to institutions or online learning bootcamps but want an environment and a teacher (in form of AI) which enables them to think more on the logic behind the problems fostering more creativity and less sticking and adhering to guidelines which is a norm still practiced in academia.