foundation models for RL

A foundation model is a single large network pretrained on huge, broad data and then reused for many downstream tasks — the way one language model backs a thousand applications. Foundation models for RL ask whether decision-making can ride the same wave: take a big model pretrained on diverse experience, video, text, or trajectories, and reuse it as the backbone for an agent rather than training every policy from scratch.

In practice this shows up several ways. Sequence models trained on logged trajectories cast control as next-token prediction, as in decision transformers. Vision and language foundation models are bolted on to give an agent perception and instruction-following it never had to learn. And world models pretrained on video aim to give planning a rich, transferable simulator of how things move. The shared bet is that broad pretraining buys generalization and sample efficiency that narrow RL cannot.

It is one of the hottest directions in the field, but the fit is imperfect. RL data is interactive, sequential, and reward-driven in a way web text is not; collecting diverse experience at internet scale is far harder than scraping documents; and a frozen backbone may encode the wrong abstractions for control. Whether decision-making gets its own clean foundation-model moment is genuinely open.