Leela without search supposedly plays around expert level, but I thought the no-search Leela approach ran out of gas around there. Without search there means evaluating 1 board position per move. The engine in the paper (per the abstract) use a big LLM instead of a Leela style DCNN.
