Carnegie Mellon University says one way of training AI could replace four separate steps, which could make powerful Chatbots cheaper and easier for big labs to build without losing performance.
Key Takeaways
A paper from Carnegie Mellon and collaborators argued one likelihood-based algorithm can cover pretraining, midtraining, supervised fine-tuning and reinforcement learning
Researcher Omar Sanseviero posted the paper on X and the arXiv abstract calls it a recommended read
The arXiv abstract says the method improved the accuracy-compute tradeoff while preserving broader capabilities across controlled and real-world tests
The authors argue reinforcement learning objectives can be derived as maximum-likelihood fitting to model-generated target distributions
Researcher Omar Sanseviero posted the paper on X on Saturday, calling it a recommended read, and the underlying arXiv abstract says the method improved the accuracy-compute tradeoff while preserving broader capabilities across controlled and real-world tests.
Carnegie Mellon University is a leading U.S. research institution whose machine learning work often feeds into the tooling and theory later adopted by commercial AI labs. The paper’s core claim is not that every training phase disappears, but that the phases may share one mathematical frame instead of separate objective functions and separate software stacks.
Carnegie Mellon University Turns Training Into Scheduling
Most large models are still trained in stages.
Labs first feed huge text and image datasets into a model during pretraining. They then adjust it with narrower data in midtraining or supervised fine-tuning, and later use reinforcement learning, a method that rewards outputs that score well against a target.
The paper argues that this split is a habit, not a law.
Its authors write that suitable reinforcement learning objectives can be derived as maximum-likelihood fitting to model-generated target distributions. In practice, that means a lab could mix static datasets and model-generated rollouts inside one trainer, then focus on deciding which data enters the run, in what share and at what point.
That matters because training stages often create engineering duplication.
Different teams maintain different pipelines, checkpoints and evaluation rules. When a lab changes objectives midstream, it also risks losing earlier capabilities or spending extra compute to recover them.
Also Read: Verifiable Rewards Drive Critical Move Beyond RLHF At Frontier Labs
From Staged Recipes To One Trainer
The idea lands as AI labs face heavier pressure to justify how each training dollar gets spent.
Larger models have made data curation, reward design and post-training more expensive, especially as reinforcement learning moved from niche alignment work into mainstream model tuning.
That pressure has already changed how labs talk about training. Over the past year, model makers have shifted from bragging about parameter counts toward cost efficiency, token quality and post-training gains.
This paper fits that change because it treats the training lifecycle as a resource allocation problem, not a string of sacred stages.
The paper also arrives after several months of public debate over whether post-training now matters more than raw pretraining scale. Unified training does not settle that fight, but it does suggest a cleaner way to combine both.
How One Likelihood Rule Changes RL
In the paper, Carnegie Mellon University and its collaborators treat reinforcement learning as another form of matching the model to a target distribution, rather than a separate family of training tricks.
That is specific because RL in model training usually means sampling answers, scoring them, then nudging the model toward higher-scoring behavior.
The authors say that same loop can be rewritten inside a maximum-likelihood setup, which is the same broad family behind standard next-token training.
For AI labs like Anthropic or OpenAI, the practical prize is not elegance. It is fewer training handoffs.
A single trainer can reduce the number of checkpoint conversions, reward-model interfaces and brittle stage boundaries that complicate long runs. For infrastructure teams, that can mean less duplicated software and a clearer path to continual learning, where a model keeps improving without a full reset.
What to watch next is adoption outside the paper.
The arXiv result is a research claim, not a production rollout. The first proof that unified training matters will come if a major lab starts describing its post-training stack in these terms, or if open-source toolchains add mixed static-and-RL training as a default pattern.
Read Next: 3 Top AI Chatbots Ran Mouse Brain Surgery And Each Picked A Different Treatment