Engineers at UC San Diego have created a better way to teach artificial intelligence (AI) systems to handle tough problems mixing words and pictures, like math questions with charts.
This approach makes AI more dependable by checking its step-by-step thinking, not just the final answer. It also picks the best training examples, ignoring poor ones, to help the AI learn faster and smarter. In tests, AI trained this way did better on math problems with images than other models.
The method could lead to helpful AI tools, such as tutors that guide students through homework while spotting mistakes in logic. It might also improve automatic reviews of reports, graphs, or science papers, with less chance of errors or made-up facts.
This research was presentedat at the NeurIPS Conference in late 2025.
How the new method improves AI learning
Usually, AI gets points only for right answers, like guessing on a quiz. But this system rewards clear, logical steps, even if the end result is wrong sometimes. It punishes bad reasoning to build stronger skills.
A big issue is uneven training data. Some is useful and hard, like advanced lessons, while other is simple or noisy, like basic kids' books. The system acts like a filter, giving more weight to good data and less to weak ones. It tests itself on separate problems to fine-tune what it learns from.
In benchmarks, the trained AI scored 85.2% on MathVista, a test of visual math skills, beating top scores. This lets smaller AI run on regular computers match big ones from companies, without huge costs.
Now, the creators are tweaking it to judge single questions better and make training quicker. The study, called "DreamPRM: Domain-Reweighted Process Reward Model for Multimodal Reasoning," shows promise for safe AI in fields like medicine or finance, where wrong steps could cause harm.