That's an entire research/engineering field. I guess you can read/view SIGGRAPH papers and presentations to get an idea what they're doing to achieve realtime performance.
Common things that make a scene too difficult to handle for brute force sampling approaches: caustics, difficult to reach light sources (small or behind openings), many light sources, camera or motion blur, volumetric materials, etc.
Algorithms approximating the behavior of real materials also is a complex topic, e.g. skin or hair are handled separately from dielectric or metallic surface models.
Some optimizations can solve one problem well in isolation but fail when encountering a combination of them.
Common things that make a scene too difficult to handle for brute force sampling approaches: caustics, difficult to reach light sources (small or behind openings), many light sources, camera or motion blur, volumetric materials, etc. Algorithms approximating the behavior of real materials also is a complex topic, e.g. skin or hair are handled separately from dielectric or metallic surface models. Some optimizations can solve one problem well in isolation but fail when encountering a combination of them.
Here's one that tries to solve caustics and small light sources at the same time: https://cgg.mff.cuni.cz/~jaroslav/papers/2012-vcm/2012-vcm-p...