Two major things here -

• Often I see that only a partial output of _fit_and_score is taken for use. It is wasteful to generate and discard arrays. It would rather be much better to generate only the stuff that is required.
• Now that we have more options, like @jnothman says here and here should we modify the output of cross_val_score (and also _fit_and_score to be a dict or a named tuple similar to the structure of cv_results_? (I think named-tuple is a better choice atleast for _fit_and_score as we stack the result of multiple _fit_and_score operations via Parallel mostly)

If we are changing the output of cross_val_score, this would be an ideal time to do it as we don't have to deprecate anything...

@jnothman @amueller @vene @GaelVaroquaux @agramfort