SIGNAL
The expensive part of any search is what you spend money on to find out if a guess was right. The cheap part is everything you do before that moment: reading, simulating, ranking. A new class of AI agents is being built to put almost all of the work on the cheap side, and to spend the expensive evaluations only on candidates the cheap model already believes are worth the bet.
The WMLLM preprint on arXiv, a self-evolving optimization agent built around what its authors call predict-then-act world modeling, is the cleanest recent example. The agent first estimates which directions in the search space are likely to improve on the goal, and only then generates candidates along those directions. The world model and the search policy refine together as real evaluations come back.
The pattern generalizes. Any domain where each real test is costly (a chemistry assay, a clinical trial arm, a chip tapeout, a live recommendation) is a candidate for the same split: build a cheap predictor, point the expensive evaluator only at the bets the predictor rates highly, then update the predictor with what comes back. The WMLLM preprint is one bounded demonstration on a multi-objective molecular benchmark under a limited evaluation budget, not a paradigm claim. But the move is portable, and other labs are already sketching the same architecture under different names.