Let $X$ be a projective complex manifold of dimension $n$. Are torsion cohomology classes in $H^{2n-2}(X,\mathbb{Z})$ algebraic? (We may assume, without loss of generality, that $n=3$, because of the Lefschetz hyperplane theorem.) I know that torsion classes (of even codimension) aren't always algebraic; the first counterexamples were found by Atiyah and Hirzebruch in 1960s. But I do not know any counterexample in this codimension. Note that by Poincare duality $H^{2n-2}(X,\mathbb{Z})\cong H_2(X)$, so the equivalent question is whether torsion classes in $H_2(X)$ are generated by algebraic curves in $X$. This looks like a "dual" to the well known fact that torsion classes in $H^{2}(X,\mathbb{Z})$ are generated by divisors (it is a torsion part of the Neron-Severi group).