Implicit inference augmented with gradients: quality of the cosmological posterior approximation as a function of the number of simulations used. We compare three methods: (1) ∂NLE with the gradients of the simulator ∇_θlog p(x,z|θ) (yellow); (2) ∂NLE with marginal gradients ∇_θlog p(x|θ) (blue); and (3) the classical NLE method (black). The dashed line indicates the C2ST threshold of 0.6, marking the point at which the posterior is considered equal to the true distribution (see Sect. 4). We show that the gradients provided by the simulator (yellow curve) do not help reduce the number of simulations as they are too noisy (see Fig. 7).