An issue with hash tree signatures is the problem of multi-target attacks. Suppose that a forger generates many instances their own WOTS+ public key/private key pairs and tries to switch this public key for one a WOTS+ key validated by the upper tree. If the leaves of the upper tree all use the same hash function, then the forger will try and use that hash function to match any leaf. If she finds any one, then she can choose to swap out her WOTS+ key with that particular leaf. If there are $N$ leaves, then the forger's chance of success for a 256-bit hash is $N2^{-256}$ rather than just $2^{-256}$. Worse, a forger could target multiple legitimate signers all of whom are using the same hash function.
To avoid this detect-one-of-many (DOOM) situation, rather than use a single hash function SPHINCS+ uses a family of very similar hash functions where the choice of hash function from the family is known as the tweak. By allowing the tweak to depend on a user specific parameter we can sidestep attacks on multiple users, by allowing the tweak to depend on input size, they block length extension attacks and by allowing the tweak to depend on leaf index (or deriving the leaf index from the output) they block multi-targetting on a given tree.
In terms of how WOTS+ signatures are used in previous proposals, the collection of $\ell$ $256$-bit hash values that make up a WOTS+ public key were combined using a binary tree (with some missing nodes) to a single 256-bit value requiring $\ell$ hash evaluations. In SPHINCS+ the $\ell$ public hash value (say, $\mathrm{pub}_1,\mathrm{pub}_2,\ldots,\mathrm{pub}_\ell$) are just combined in a single (tweaked) hash function: $\mathrm{Th}(\mathrm{pub}_1,\mathrm{pub}_2,\ldots,\mathrm{pub}_\ell)$ which removes significant overhead.
