A code model can produce several candidate solutions to the same task, with only some passing the task's checks. Code chooser studies whether a smaller selector can choose a useful candidate without generating a new solution itself.
Comparison
The experiments generate candidate sets, evaluate them in a sandbox, and compare selection strategies: random choice, the generator's likelihood, a larger teacher model, and a trained selection head over a frozen model.
Training targets come from teacher preferences, observed test outcomes, or a mixture. This separates the ability to imitate a teacher from the ability to learn directly from passing and failing programs.
Selection is scored on mixed candidate sets containing both outcomes. That isolates the choice problem, but it is not overall coding success: a selector cannot recover a correct solution when none was generated.
Recorded result
On 300 mixed HumanEval+/MBPP+ tasks, the teacher-distilled selection head chose a passing program 83% of the time, compared with 54% for random selection and 87% for the teacher. Those tasks were not used to train the selection head.
These are conditional selection rates, not overall coding success. The recorded study uses a single seed. HumanEval and MBPP also have possible training-data exposure, so held-out tasks from a separate source matter when interpreting transfer.