The key to deriving that formula is seeing how the various paths overlap and remembering that any number XOR itself equals 0. In particular, Q(x, y) XOR Q(x, y) will also equal 0.
The Quick Derivation
F(u) is the path from the root node to u. F(v) is the path from the root node to v. Because any number XOR itself equals 0, F(u) XOR F(v) will "cancel out" all the nodes above the lowest common ancestor of u and v, LCA(u, v). So F(u) XOR F(v) = Q(LCA(u, v), u) XOR Q(LCA(u, v), v). This happens to be the XOR sum of all nodes on the unique path from u to v, except for LCA(u, v) itself, since it's currently counted twice. So if we simply count it a third time, then we get the sum of all the nodes in Q(u, v). Thus Q(u, v) = F(u) XOR F(v) XOR LCA(u, v).
I suspect this "proof" leaves out too many details for you or contains a few steps you're not already convinced of, so here's a much longer version:
The Thorough Derivation
We're assuming that our graph is a valid tree, and there is a unique single path between u and v. This ensures u and v have a single "lowest common ancestor", LCA(u, v), which I will refer to as a. I will call the root node r. Keep in mind that a may be the same node as u, v, a, or even all three. The argument I'm about to show should still work even in those degenerate cases.
So there are unique paths (possibly of length zero) between these four nodes, like this:
Following the "editorial", let Q(x, y) denote the XOR sum of the nodes on the unique path from x to y.
We're interested in Q(u, v). Clearly, the unique path from u to v passes through a. However, it would be wrong to say that Q(u, v) = Q(u, a) XOR Q(a, v) because that would count the value of a twice. Fortunately, because XOR'ing a value by itself results in zero, counting a three times is the same as counting it only once. So the correct statement is:
Q(u, v) = Q(u, a) XOR Q(a, v) XOR A[a]
Next, F(i) is defined as Q(r, i). So we can follow the same logic as above and say:
F(u) = Q(r, u) = Q(r, a) XOR Q(a, u) XOR A[a]
F(v) = Q(r, v) = Q(r, a) XOR Q(a, v) XOR A[a]
Now, what is F(u) XOR F(v)? Because the XOR operation is commutative and associative, we can safely rearrange a series of XORs like we would a series of +s, which lets us do this:
F(u) XOR F(v) = (Q(r, a) XOR Q(a, u) XOR A[a]) XOR (Q(r, a) XOR Q(a, v) XOR A[a])
= (Q(r, a) XOR Q(r, a)) XOR (Q(a, u) XOR Q(a, v)) XOR (A[a] XOR A[a])
= 0 XOR (Q(a, u) XOR Q(a, v)) XOR 0
= Q(a, u) XOR Q(a, v)
More intuitively: Whenever you XOR the XOR sums of two overlapping paths, the parts that overlap simply disappear, and we're left only with the XOR sums of the non-overlapping parts of the paths.
Finally, recall that Q(u, v) = Q(u, a) XOR Q(a, v) XOR A[a]. We now know that Q(a, u) XOR Q(a, v) is equal to F(u) XOR F(v), which means:
Q(u, v) = Q(u, a) XOR Q(a, v) XOR A[a]
= F(u) XOR F(v) XOR A[a]
And since I defined a as LCA(u, v), we're done.
