A new research paper from Carnegie Mellon University argues that pretraining, fine-tuning, and reinforcement learning can run as a single algorithm rather than three separate training stages.

Most foundation models—the large, general-purpose models that serve as the base for many AI systems—are built like a three-course meal: pretraining first, fine-tuning second, and reinforcement learning for dessert. The Carnegie Mellon study suggests those courses might work better served on one plate.

The paper, titled “Training Foundation Models with a Single Algorithm,” was posted to arXiv on October 29, 2026, a preprint server where research is commonly released ahead of formal peer review. Its lead author is Yuda Song of Carnegie Mellon University, who worked alongside collaborators from Johns Hopkins University and Aurora Innovation. According to the study, blending all three types of training data into one unified method can beat the traditional staged approach.

What the researchers actually did

The core idea is a reframing. The team treats every training stage as maximum-likelihood fitting of model-generated data.

That shift moves the important decisions somewhere new. Instead of a rigid sequence of distinct phases, the framework puts the emphasis on data composition and scheduling. Under this view, the difference between pretraining, fine-tuning, and reinforcement learning becomes a question of which data is used and when, rather than which algorithm is applied. Static datasets—the kind used in pretraining and supervised fine-tuning—are folded together with the data management used in reinforcement learning, and everything flows through one training framework.

The results so far

According to the paper, the unified approach improved the tradeoff between accuracy and compute. In practical terms, the models delivered more performance out of the same processing budget.

The method also outperformed staged training across several measures. These included in-distribution evaluations, generalization, and readiness for continual learning. The researchers also report that unified training helped models retain broader capabilities across many tasks.

The framework removes the need for separate infrastructure built for each training stage, which the authors say simplifies training systems overall. The experiments covered models with around 124 million parameters and around 0.76 billion parameters—scales well below those of the largest models commonly described as foundation models, so the paper's supporting evidence comes from a comparatively modest slice of the model-size spectrum.

Why the three-stage recipe exists in the first place

The staged pipeline became standard because each phase appeared to do a distinct job. Pretraining builds broad knowledge, fine-tuning teaches the model to follow instructions, and reinforcement learning shapes behavior toward preferred outcomes.

The CMU paper challenges that separation directly. It is described as the first public presentation of a unified-training framework backed by supporting experiments.

The authors go a step further. They argue that such a framework could open new ways to understand how capabilities emerge in models, and how those capabilities are retained as models evolve.

What this means for AI developers

As of October 30, 2026, the research had produced no notable commercial applications or market partnerships. Discussion has stayed mostly within the academic community, focused on what a single training system could mean for model development. The paper's stated practical implication for developers is that a single training framework could replace the separate systems currently built for each stage. As it stands, the work remains a research result rather than a deployed technique.

For anyone tracking the work, two concrete markers follow from the paper's current status: it has not yet been through formal peer review, and its reported experiments stop at the parameter scales described above. Independent replication and follow-up studies at larger scales are the visible checkpoints as the research circulates.