AlphaCode 2
Pairs Gemini Pro code generation with execution filtering, behavioral clustering, and a learned scoring model to rank competitive programming solutions.
The loop
- Proposes
A family of fine-tuned Gemini Pro models samples up to one million diverse C++ candidate programs per problem.
- Judged by
Sandboxed execution prunes about 95% of samples that fail the public example tests or do not compile, then clusters survivors on synthetic inputs.
- Steers
A second Gemini scoring model ranks the top cluster representatives; there is no repair step.
Headline result
Solved 43% of Codeforces Division 1+2 problems and ranked at the 85th percentile of human competitors.
What this does not prove
High inference compute budget per problem; correctness remains bounded by test coverage rather than formal proof.
How it generalizes the harness
Execution-based pruning and dynamic input-output clustering filter a million-candidate population down to 10 submissions.
Classification
| What it proposes | Program |
|---|---|
| Generative model | Fine-tuned Gemini Pro |
| Algorithmic authority | Sandboxed C++ execution filter and behavioral clustering engine |
| Feedback | Test results |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Successor |