We present a mitigation strategy to reduce the impact of non-linear galaxy bias on the joint `$3 \times 2 $pt' cosmological analysis of weak lensing and galaxy surveys. The $\Psi$-statistics that we adopt are based on Complete Orthogonal Sets of E/B Integrals (COSEBIs). As such they are designed to minimise the contributions to the observable from the smallest physical scales where models are highly uncertain. We demonstrate that $\Psi$-statistics carry the same constraining power as the standard two-point galaxy clustering and galaxy-galaxy lensing statistics, but are significantly less sensitive to scale-dependent galaxy bias. Using two galaxy bias models, motivated by halo-model fits to data and simulations, we quantify the error in a standard $3 \times 2$pt analysis where constant galaxy bias is assumed. Even when adopting conservative angular scale cuts, that degrade the overall cosmological parameter constraints, we find of order $1 \sigma$ biases for Stage III surveys on the cosmological parameter $S_8 = \sigma_8(\Omega_{\rm m}/0.3)^{\alpha}$. This arises from a leakage of the smallest physical scales to all angular scales in the standard two-point correlation functions. In contrast, when analysing $\Psi$-statistics under the same approximation of constant galaxy bias, we show that the bias on the recovered value for $S_8$ can be decreased by a factor of $\sim 2$, with less conservative scale cuts. Given the challenges in determining accurate galaxy bias models in the highly non-linear regime, we argue that $3 \times 2$pt analyses should move towards new statistics that are less sensitive to the smallest physical scales.