Modeling Strong and Human-Like Gameplay with KL-Regularized Search - we read this one on the transhumanists in vr discord server to figure out what they were testing and what results they got. key takeaways according to me, note that I could be quite wrong about the paper's implications:
Multi-agent game dynamics change significantly as you add more coherent search and it becomes harder to do linear learning to approximate the search. (no surprise, really.)
it still takes a lot of search.
guiding the search is not hopeless in the presence of noise!
shallow, no-planning equilibrium in imperfect-information games is not improved in emulating a reference policy as much? this seems to make sense as an additional confirmation of the basic hypothesis "search helps model searching beings".