The miner binaries in both those links are the same code. The first link is the minerd binaries only, while the second link is a collection of batch-file wrappers plus the minerd binaries get downloaded from the first link during Setup. All the miner binaries are compiled using the Wolf0 updated code design with joe's optimizations available here:-
https://github.com/magi-project/wolf-m7m-cpuminer-V2which is the "official" source code. I use a custom-built MSYS/Mingw64 compiler suite with a customized hybrid version of the GMP math library, plus a fair bit of know-how and jiggery-pokery to get the fastest possible miner for Windows. I also have a very sturdy birch twig which doubles-up as a magic wand from time to time
I don't think that I can get the 64 bit miner any faster, but I will be revisiting the 32 bit version to see if I can squeeze a bit more out of that - there is room for some improvement.
The minerd.exe in each of the sub-folders avx, avx2 and generic have the same date and the same file size because they are all the same code. Earlier versions of the miner ran faster when using avx or avx2 instruction sets (on CPUs that support avx or avx2) but this latest miner seems to run everso slightly slower when avx instructions are used, so I just copied the generic compilation to the other two sub-folders to maintain compatibility with the existing software suite.
You will get better hashrate using the batch-file wrappers because they are designed to allow the miner(s) to run with the greatest level of efficiency on any given machine. Running multiple instances of the miner with a single thread and deliberately load-balancing these across the available CPU cores is more efficient than running a single instance of the miner with multiple threads and just allowing the system to get on with it (badly), so you will get a slightly higher total hashrate as a result. Not by very much - say about 0.6 percent extra hashrate - but measurable with careful benchmarking.