Not a proper solution to the question asked, but per @Oray's request, perhaps a reasonably elegant benchmark?
Label the balls A-H, and start ABCD in the left pan, EFGH in the right. Add a dot to each ball in the heavier pan. For each subsequent weighing, the left pan should contain the lexicographically first combination of four balls among those which distribute the dots as evenly as possible between the two pans, excluding those which have already been tried. For example, the second weighing will be ABEF<>CDGH, since each pan will have two dots regardless of which initial set was heavier.
To take this for a test drive, I took two sets of balls which summed to one, and partitioned each into distinct fractions with coprime denominators. To start, (1,2,3,4)/10 and (1,2,3,5)/11. This does give two 2-subsets that add to 5/10, but as I mentioned above, I don't think this is actually relevant, and you can always add/subtract an epsilon if it bothers you without affecting the results. Then I assigned all 40320 permutations of A-H and counted weighings up to the one that balanced, with the following results.
1 1152
2 1152
3 1264
4 1388
5 1296
6 1396
7 1404
8 1517
9 1530
10 1532
11 1674
12 1782
13 1871
14 1777
15 1645
16 1832
17 1945
18 2028
19 2109
20 2199
21 2146
22 2183
23 1722
24 1033
25 504
26 171
27 38
28 29
29 1
for an average-case of ~13.68...an improvement over brute-force case-checking (17.5), but not as much as I might have hoped. There is something obscenely inefficient going on in my code, as it only checks about 20 cases/second, otherwise would have many more cases checked. That lone straggler, by the way, is:
[A,B,C,D,E,F,G,H]=[44, 10, 50, 11, 30, 20, 22, 33]
and the resulting sequence of marks and weightings:
1 [0, 0, 0, 0, 0, 0, 0, 0] ABCD 0 0 115 105
2 [1, 1, 1, 1, 0, 0, 0, 0] ABEF 2 2 104 116
3 [1, 1, 2, 2, 0, 0, 1, 1] ABCE 4 4 134 86
4 [2, 2, 3, 2, 1, 0, 1, 1] ABDF 6 6 85 135
5 [2, 2, 4, 2, 2, 0, 2, 2] ABCF 8 8 124 96
6 [3, 3, 5, 2, 2, 1, 2, 2] ABDE 10 10 95 125
7 [3, 3, 6, 2, 2, 2, 3, 3] ABGH 12 12 109 111
8 [3, 3, 7, 3, 3, 3, 3, 3] ABCG 16 12 126 94
9 [4, 4, 8, 3, 3, 3, 4, 3] ABDG 15 17 87 133
10 [4, 4, 9, 3, 4, 4, 4, 4] ABEG 16 20 106 114
11 [4, 4, 10, 4, 4, 5, 4, 5] ABFH 18 22 107 113
12 [4, 4, 11, 5, 5, 5, 5, 5] ABCH 24 20 137 83
13 [5, 5, 12, 5, 5, 5, 5, 6] ABDH 21 27 98 122
14 [5, 5, 13, 5, 6, 6, 6, 6] ACDE 29 23 135 85
15 [6, 5, 14, 6, 7, 6, 6, 6] ADEF 25 31 105 115
16 [6, 6, 15, 6, 7, 6, 7, 7] ACDF 33 27 125 95
17 [7, 6, 16, 7, 7, 7, 7, 7] ADEG 28 36 107 113
18 [7, 7, 17, 7, 7, 8, 7, 8] ACDG 38 30 127 93
19 [8, 7, 18, 8, 7, 8, 8, 8] ADFG 32 40 97 123
20 [8, 8, 19, 8, 8, 8, 8, 9] ABEH 33 43 117 103
21 [9, 9, 19, 8, 9, 8, 8, 10] ACFG 44 36 136 84
22 [10, 9, 20, 8, 9, 9, 9, 10] AEFH 38 46 127 93
23 [11, 9, 20, 8, 10, 10, 9, 11] AEGH 41 47 129 91
24 [12, 9, 20, 8, 11, 10, 10, 12] AFGH 44 48 119 101
25 [13, 9, 20, 8, 11, 11, 11, 13] AEFG 46 50 116 104
26 [14, 9, 20, 8, 12, 12, 12, 13] ABFG 47 53 96 124
27 [14, 9, 21, 9, 13, 12, 12, 14] ADEH 50 54 118 102
28 [15, 9, 21, 10, 14, 12, 12, 15] ADFH 52 56 108 112
29 [15, 10, 22, 10, 15, 12, 13, 15] ADGH 53 59 110 110
A long way from the lower bound of 6, but still better than exhaustively testing all 35. The other angle I was considering was to start with 7 weighings such that every triplet of balls shares a pan exactly once, but I suspect pathological cases will stop that from being sufficient, and this would be easier to execute from memory/by hand in a pinch.
UPDATE: Some combinations of 13ths and 14ths got up to a worst-case of 32, suggesting that additional heuristics like @Ben J's are needed to avoid particularly unfortunate orderings (with the same inputs his worst-case is 30, so they do help). I have a couple of refinements in mind myself, but perhaps a more binary-tree oriented approach would be better, if we could map out the feasible sum-orderings (which look to be a MUCH smaller set than I'd imagined, but tricky to make sure I've caught them all...)
...or maybe, per Dave B's suggestion, [1,2,3,25,26,27,28,100] renders the whole exercise moot. Well, at least it behaves interestingly for more uniform distributions. ^_^b