Before it
Code models produced one attempt per problem.
The problem
Competitive programming problems are too hard to solve first try.
What it introduced
Generate a million candidate programs, filter by running the examples, and cluster to pick submissions.
Why it matters
An early argument that spending more compute at inference buys accuracy, which is the test-time compute thesis.
What modern models inherit
Sampling widely and filtering, rather than trusting a single generation.
What came next
Self-consistency, verifier-guided search and eventually reasoning models.