The Secret to Better AI Training? Knowing When to Stop

Imagine you reward a customer service team based on how quickly they close tickets. At first, response times improve and customers are happier. But over time, some representatives start closing tickets before problems are actually solved. The scoreboard looks great. The customer experience doesn’t.

AI can fall into the same trap. Many of today’s AI models are fine-tuned using a scoring system that rewards answers people are likely to prefer. That score is meant to stand in for real quality, but it isn’t the same thing. Train a model long enough, and it can learn to chase the score instead of getting better at the job. Meanwhile, expensive computing power keeps running a model that has already peaked, because the score says everything is fine.

This raises a costly question for enterprises: how do you know when AI training has stopped making the model better? 

A recent Workday study answers that question. Our AI Research Team developed a method to check a model’s performance with a separate quality test during training, not just its training score—and halt training when quality started to slip. The study produced three big findings:

1. It’s possible to identify the point at which AI training has diminishing returns. While only a simulation, Workday’s method caught 99% of AI training jobs that had started making the model perform worse—and wrongly stopped only about 1.5% of jobs that were still improving the model’s performance.
2. Stopping at the right time frees up expensive computing power. Wasted Graphics Processing Unit (GPU) time dropped by about 22%, and jobs finished about 9% sooner, leaving more GPUs available for other teams.
3. A fixed cutoff is a risky shortcut. You might think the simplest fix is to stop every job after a set amount of training. But that approach cut off about two-thirds of the jobs that were still improving.

I sat down with Guilin Zhang, a principal AI researcher on Workday’s AI Research team, to talk about the study, conducted in partnership with Chuanyi Sun, Kai Zhao, Xu Chu, Shahryar Sarkani, and John M. Fossaceca. Here’s what Zhang had to say.

1. What was the main AI challenge you were looking to solve?

Zhang: We wanted to solve a problem in shared GPU clusters. A model’s training score can keep rising even when its actual quality starts to fall. The cluster may then keep giving that job GPUs, which wastes computing time. We wanted to detect that and stop the job at the right time.

2. Why does this matter for business leaders? 

Zhang: The reason is straightforward. GPUs are very expensive, and teams often share them. If a job keeps running while the model gets worse, it wastes computing time and money, and it makes other teams wait. Both AI quality and computing cost affect the business.

The cost comes mainly from continuing to train the model on GPUs, not from testing it. If a model has stopped improving, more training uses expensive computing time without making it better. Regular testing can help identify when it’s time to stop. That also frees up GPUs so other teams can use them right away.

I think this is also relevant to AI researchers and engineers who train and test models, and to the infrastructure teams who manage GPUs, as well as business leaders.

It’s not perfect, but a good test can show when a training score is going up while actual performance is going down.

3. Two terms come up a lot in your paper. “Reward hacking” and “world feedback.” Can you explain what these mean?

Zhang: “Reward hacking” happens when a model gets better at maximizing its training score without actually getting better at the task. For business leaders, that matters because you can spend more GPU time and money while the model’s real quality gets worse.

“World feedback” means checking how well the model actually performs on an evaluation, rather than relying only on the training score. That gives us a separate signal to detect when the score is rising but the quality is falling.

4. Tell me about the solution the team discovered. What can others learn from it?

Zhang: We call it EvalStop. We focused on the system that decides which jobs get GPUs, and we added a check to it. EvalStop tests the model on a separate evaluation during training. If the test score drops twice in a row, it stops the job, keeps the best version of the model, and frees the GPUs for other jobs.

There are three big takeaways. 

First, EvalStop caught almost all of the jobs that started getting worse, while rarely stopping jobs that were still improving. In the study, it caught 99% of the jobs that were getting worse, and it incorrectly stopped only about 1.5% of the jobs that were still improving. 

Second, it reduced wasted GPU time by about 22%, and jobs finished about 9% sooner. 

Third, we found that stopping every job at a fixed point was risky: that approach stopped about two-thirds of the jobs that were still improving.

5. What are the risks of stopping AI training too early versus stopping it too late?

Zhang: If we stop too early, we might miss a chance to make the model better. If we stop too late, we waste GPU time and the model gets worse. In our work, EvalStop tries to balance those risks by watching the test results and keeping the best version.

6. What does this research say about how companies should conduct their own AI training?

Zhang: Companies should not rely only on training scores. They should also test whether the AI is getting better at the tasks that matter to their users. They need to look beyond the training score.

One encouraging finding in our study is that a simple check can help us use AI computing resources more efficiently. In our simulation, it reduced wasted GPU time by about 22% while rarely stopping jobs that were still improving.

7. What are some things that business leaders can do to ensure that AI quality is actually improving?

Zhang: First, define what good results look like for users. That’s the most important one. 

Second, ask teams to test AI on real tasks, not just track training scores. 

Third, keep people involved in reviewing the results and improving those tests. Humans still need to decide what good performance means and design tests that reflect it. Teams should use those tests, but they don’t replace human judgment.

8. How does this research contribute to responsible AI?

Zhang: This research supports the broader goal of building AI that works well for people while using computing resources responsibly. It focuses on actual performance, not just training scores. 

9. Is there anything else you would like to add?

Zhang: A useful question for business leaders is: why can’t we simply stop every job after a set amount of time? In our study, that approach stopped about two-thirds of the jobs that were still improving. EvalStop checks the model’s test results instead, so it can make a better decision about when to stop.

Looking for more insights on how to make AI trustworthy, useful, and efficient for the enterprise? Visit workday.com/ai-research for a complete list of Workday AI research papers.