8.5.0 has further improvements, including fully working sparse model support, empirically optimized exploration algorithms, a new cost-sensitive active learning algorithm (https://arxiv.org/abs/1703.01014), and baseline prediction support.
8.4.0
There aren’t many new features since 8.3.0 (better JSON language support, better offline eval, sparse parameter support), so this version simply addresses a number of issues that have shown up over time
8.2.0
Many things added including more sophisticated contextual bandits, the recall tree, and OjaNewton.
7.10
No significant changes from 7.9---primarily bugfixes.
7.9
Primarily updates to learning reductions modularity.
7.8
<a href="http://hunch.net/?p=222592">A post</a> with some notes