Perhaps things like casing conventions? That would be covered by a linter too.

Or does it mean style-of-semantics (if that's even a thing), for example in C#, preferring to define explicit constructors instead of a parameterless ctor and trusting the user to set properties in the post-ctor initializer correctly?

Or use of higher-order/FP techniques (like map/filter/reduce/fold) over explicit for-loops?

The summary says this (note the last remark):

> The results demonstrate that good programmers may be identified using supervised machine learning models, despite that no particular style groups could be attributed as a good style.

Obviously they're not saying, for example, that one particular bracing style is better than another (1TBS tho!), nor that co-incidentally "good" programmers prefer one style over another; but despite knowing what they're not saying, I don't know what they are saying with that...