A new preprint on arXiv outlines a method for reinforcement learning from unlabeled data by combining behavior foundation models with generative modeling. The authors describe an appealing recipe: pre-train a generative model to model the data, then use that model to extract a policy directly. The approach is called "drifting," and it aims to make the extraction process a one-step generation rather than a multi-step procedure.

The abstract is truncated in the source, so full technical details are not yet available. What is clear is the motivation: leveraging large amounts of unlabeled behavioral data to learn policies without explicit rewards or labels. The proposed drifting mechanism appears to be the key innovation, allowing the generative model to be converted into a policy in a single step.

As with many preprints, the claims have not yet been peer-reviewed, and the missing portion of the abstract limits a full assessment. Still, the direction aligns with a broader trend toward foundation models for decision-making, and the one-step generation idea could have implications for efficiency in policy extraction.