Behaves like the identity near zero and like a logarithm far from it, so a single network can be trained on targets whose scale is not known in advance — rewards that are 0.01 in one environment and 10000 in another land in the same range without per-task normalisation. That is what it is for in DreamerV3.

The inverse is symexp, and the pair is exact: symexp of symlog returns the input. Implemented through log1p rather than log(1 + ·) so small inputs keep their precision.