A Nature study assessed repeated code generation and testing for scientific forecasting across six fields. Of 55 combinations of existing methods, 24 outperformed both methods from which they originated.
On June 24, 2026, a publication described a Nature study on repeatedly generating and evaluating code for scientific forecasting tasks across six fields. The researchers examined 55 combinations of existing methods; 24 performed better than both methods from which they originated.
The work indicates that automated cycles of code creation and evaluation can accelerate experiments with forecasting software. To confirm the scope and results, consult the original Nature publication and check how it defines the fields, combinations, and comparison with the original methods.