Using shadow map only, I don’t think that’s possible. We could find the bounding cone of the receivers and rotate the light towards them and reduce the projection to the cone, but that sounds like very involved and would trash the shadow map cache (very bad for performance).
Instead I think it would be easier to have a denoising + sharpen option on the shadow buffer itself. But that requires deferred shadowing. This would also allow sharp shadows from larger lights that generally produce soft shadows.