Drifting models are attractive for language generation because they produce a sample with a single function evaluation, which makes them well suited to fast parallel generation. The trade-off, according to the new arXiv paper ReLaDrift, is that this one-step formulation does not support iterative refinement.
The paper, announced as new on arXiv (2610.23225v1), proposes Recurrent Language Generation via Drifting — a recurrent variant that adds the missing refinement loop. The abstract frames this as a step toward test-time scaling for drifting models, though the full method and experimental results are not detailed in the abstract itself.
Because this digest draws on a single source, there are no conflicting accounts to reconcile. The main open question — how the recurrent formulation compares empirically with other test-time scaling approaches — is not addressed in the available text.