Substepping
Splitting a frame's timestep into several smaller physics steps — more but shorter steps buy stability and accuracy.
One large step lets constraint error grow before the solver can correct it, and fast bodies leap too far between checks. Running the physics n times per frame with a step of dt/n shrinks both problems at once.
The old rule of thumb was to iterate the solver more times per step instead. Macklin et al. flipped that in 2019 (Small Steps in Physics Simulation): many short substeps with few iterations beat few long steps with many iterations, both in stability and in how stiff a spring you can simulate. Unreal's Chaos and PhysX now build on it.