AlphaCode
Samples millions of candidate programs per competitive programming problem, filters by example tests, and clusters by synthetic input behavior.
The loop
- Proposes
An encoder-decoder transformer generates a large pool of candidate C++ and Python solutions from the problem description.
- Judged by
A sandboxed execution engine filters out candidates that fail public example tests (>99% pruned) and runs the survivors on generated test inputs to cluster equivalent behaviors.
- Steers
Behavioral clustering selects one representative program from each of the largest equivalence classes for final submission.
Headline result
Placed in the top 54.3% of human participants across 10 recent Codeforces competitions, the first AI system to reach competitive human rank.
What this does not prove
Relies on sampling hundreds of thousands of programs per problem; passing public example tests does not prove correctness on hidden edge cases.
How it generalizes the harness
Large-scale execution filtering and behavioral equivalence clustering select among a massive population of candidate programs.
Classification
| What it proposes | Program |
|---|---|
| Generative model | 41B encoder-decoder code transformer |
| Algorithmic authority | Sandboxed test execution and input-output equivalence clustering |
| Feedback | Test results |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |