A better approach is to re-use arrays instead of creating intermediate ones, which is what gl-matrix[1] advocates. Each tween might still need its own matrix instance for start and end points, but at least these are not created per tick.
A similar approach is to use a "pool" of arrays[2] which allows you to re-use them without having to manually keep track of the temporary arrays.
Similarly, PhysicsJS employs a "scratchpad" of pre-initialized vectors for fast intermediate computations. That way you aren't constructing new objects, but grabbing a free one, zeroing it, and using it as you would before. Seems quite performant.
Out of curiosity, can you show an example and how it should be done properly?