AlphaCode
A code generation system that solves competitive programming problems.
AlphaCode is a code generation system built by DeepMind researchers to write novel solutions to competitive programming problems needing real reasoning, not just translation of instructions. Described in a 2022 paper, it reached an average ranking in the top 54.3% of large Codeforces contests in simulated evaluations.
- Reached an average ranking of top 54.3% in large Codeforces contests
- Evaluated on competitions with more than 5,000 participants
- Built to solve problems needing algorithmic and language reasoning
- Relies on large-scale sampling plus filtering by program behaviour
- Described in a February 2022 paper by DeepMind researchers
What it is
AlphaCode is a system that writes programs for hard, previously unseen competitive programming tasks. It was created by researchers at DeepMind and presented in a paper posted in February 2022. The work targets problems that need understanding of algorithms and of complicated natural language, which earlier language models handled poorly. In simulated runs on recent contests held on the Codeforces platform, AlphaCode placed on average within the top 54.3% of competitions that drew more than 5,000 participants. The paper treats it as a research result rather than a released product.
How it came about
The system is described in the paper "Competition-Level Code Generation with AlphaCode", first posted to arXiv on 8 February 2022 and later linked to a Science DOI. Yujia Li led a 26-person author list that also includes Nando de Freitas, Pushmeet Kohli, Koray Kavukcuoglu and Oriol Vinyals. The team set out to close the gap between models that merely turn instructions into code and systems that can reason about unfamiliar algorithmic problems.
Inside the model
The authors say three ingredients drove reliable results. First, a large and carefully cleaned dataset of competitive programming tasks, used for both training and evaluation. Second, transformer architectures that are large yet cheap to sample from, meaning many candidate programs can be drawn quickly. Third, heavy sampling: the system explores a wide space of possible solutions, then filters them by how the programs actually behave, keeping only a small set of submissions. The paper does not state parameter counts, context length or training compute.
What it can do
In simulated evaluations on recent Codeforces competitions, AlphaCode achieved an average ranking of top 54.3% in events with over 5,000 participants. The authors frame this as evidence the system can produce genuinely novel solutions requiring deeper reasoning, where earlier models performed poorly. They note that competitive problems demand both algorithmic skill and comprehension of complex natural language. No other benchmark figures are given.
How to use it
The work is presented as a research paper on arXiv, with links to code, data and media services listed alongside it. No downloadable weights, licence terms, app or API are described.
Why it matters
The paper positions AlphaCode as a step beyond code completion, showing that large-scale sampling plus behavioural filtering can reach human competitive programming levels on hard, unseen tasks. Its three-part recipe, clean competition data, efficient-to-sample transformers and heavy sampling with filtering, is presented as the key to that progress.
Questions people ask about AlphaCode
What is AlphaCode?
A code generation system from DeepMind that writes novel solutions to competitive programming problems. Unlike simple code completion tools, it is aimed at hard, unseen tasks that require understanding algorithms and complicated natural language descriptions.
How well did it perform?
In simulated evaluations on recent Codeforces competitions, it achieved an average ranking of top 54.3% in contests with more than 5,000 participants. Those are the only performance figures given in the paper abstract.
How does it work?
It trains on a large, cleaned competitive programming dataset using transformer architectures that are cheap to sample from. It then draws many candidate programs and filters them based on how the programs behave, keeping a small set of submissions.
Can I use AlphaCode?
The work appears as a research paper on arXiv, with links to related code and data services. No public weights, licence, app or API are described in the available material.
Good reading and sources
There is no English Wikipedia article on AlphaCode yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2203.07814. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.