Reduced bispectrum, Q(θ₁₂), as a function of the subtended angle, θ₁₂, for four different triangular configurations: (i) k₁ = 0.1,  k = 0.2 h Mpc⁻¹ (top left), (ii) k₁ = 0.2,  k = 0.2 h Mpc⁻¹ (top right), (iii) k₁ = 0.3,  k = 0.5 h Mpc⁻¹ (bottom left), and (iv) k₁ = 0.5,  k = 0.5 h Mpc⁻¹ (bottom right). The plot displays the bispectrum of the reference simulation (solid red line), compared to the predictions from the standard FGPA (dashed yellow line), the FGPA with velocity bias (dotted purple line), the FGPA with velocity and threshold bias (dash-dotted black line), the nonlocal FGPA (dashed green line), and the nonlocal FGPA with stochasticity (dash-dotted blue line). The latter (i.e., our preferred model) reproduces the target bispectrum with reasonable accuracy for all the probed triangular configurations, while the others start to feature significant deviations at k ≳ 0.2 h Mpc⁻¹.