CFS lifts retrieval rank with overlap penalties
CFS is a retrieval-ranking heuristic that selects candidates by subtracting regions already covered by prior picks, rather than only adding more score mass. In the reported benchmark, it trails the best fusion variant only slightly in absolute terms but improves over mem0’s additive fusion and also adds lift on top of cosine + BM25 RRF, reaching NDCG@10 0.5311 and Recall@10 0.7168.
Hot take: this looks like a sensible diversification trick that is cheap enough to try in real retrieval stacks, and the reported gains are directionally credible. The main caveat is that the evidence is still narrow, so I’d treat it as an interesting benchmark win rather than a broadly validated method.
- –Strongest signal is against additive fusion, where CFS materially improves both ranking quality and recall.
- –The delta over `rrf(cosine, BM25)` is modest but consistent, which suggests it is contributing complementary diversity rather than just reweighting existing signals.
- –The benchmark is limited: no dataset details, confidence intervals, or ablation on when CFS helps versus hurts.
- –The method sounds easiest to apply where near-duplicate or region-overlap suppression matters, especially multi-stage retrieval or memory search.
DISCOVERED
7h ago
2026-05-08
PUBLISHED
10h ago
2026-05-08
RELEVANCE
AUTHOR
mauro8342