Google has moved the cost of "going out to search" from serving time to training time. Other work, meanwhile, deliberately keeps that search on the inference side.
Google Research published Retrieve-for-Train. Rather than returning the single best match, it returns a coherent slate of results that complement one another. Ask for camping gear and it comes back with a tent, a sleeping bag and a stove together. The older approach paid its cost on every query, stacking up autoregressive reasoning each time. This one trains a lightweight diffusion model once, via reinforcement learning, and then generates the whole slate instantly.
Making a model reason on every query is slow and expensive. Paying up front at training time makes serving fast and cheap, but the behaviour is fixed until you retrain. Other work runs the other way. HypoEvolve uses genetic algorithms to let multiple LLM agents search for scientific hypotheses, which effectively pushes more search onto inference. Which choice is right depends on the job. Search that must answer instantly and research that can afford to take its time do not want the cost in the same place.
The weakness of paying up front is what happens when the world changes after training ends. How Retrieve-for-Train reports the frequency and cost of retraining will decide whether this becomes the default shape of search or stays a fit for particular uses.