UC San Diego Engineers Develop AI System for Improved Text and Image Understanding

New training method outperforms others on visual math reasoning tests, could enable smaller AI models to rival larger proprietary systems.

Published on Feb. 11, 2026

Got story updates? Submit your updates here. ›

Engineers at the University of California San Diego have developed a new way to train artificial intelligence systems to solve complex problems more reliably, particularly those that require interpreting both text and images. Their method, which evaluates how AI models reason through problems instead of just checking final answers, outperformed others in solving math word problems containing visual elements like charts and diagrams.

Why it matters

This work could lead to more capable AI tutors that walk students through solutions step by step while checking their logic, as well as more reliable automated analysis of business reports, complex charts or scientific papers with reduced risk of fabricated information or incorrect interpretations. It also democratizes AI by enabling smaller models to rival larger proprietary systems on difficult benchmarks.

The details

The team's approach has two key features: it evaluates the quality of training data so that higher-quality examples carry more weight during learning, and it rewards models for showing their work rather than just getting the right answer. This shift from "Did the AI get it right?" to "Did the AI think it through?" could provide a safety net for high-stakes applications like medical diagnosis, financial analysis and engineering.

  • The research was presented at the NeurIPS Conference in December 2025.
  • The team's AI model achieved a top public score of 85.2% on the MathVista benchmark in 2026.

The players

Pengtao Xie

Professor in the Department of Electrical and Computer Engineering at the UC San Diego Jacobs School of Engineering and senior author of the study.

Qi Cao

Co-author of the study, affiliated with UC San Diego.

Ruiyi Wang

Co-author of the study, affiliated with UC San Diego.

Ruiyi Zhang

Co-author of the study, affiliated with UC San Diego.

Sai Ashish Somayajula

Co-author of the study, affiliated with UC San Diego.

Got photos? Submit your photos here. ›

What they’re saying

“They are graded much like students taking a multiple-choice test. If they select the right answer, they still receive full credit, even if they guessed.”

— Pengtao Xie, Professor, UC San Diego (Mirage News)

“Our system doesn't just learn from everything. It learns what is worth learning from. It emphasizes quality over quantity.”

— Pengtao Xie, Professor, UC San Diego (Mirage News)

“You don't need a trillion-dollar computing cluster to get state-of-the-art reasoning.”

— Pengtao Xie, Professor, UC San Diego (Mirage News)

What’s next

The team is now refining the system by evaluating the quality of individual questions rather than entire datasets, and making the training process faster and less computationally demanding.

The takeaway

This new training method for AI systems that can reason with both text and images could lead to more capable and reliable AI assistants, tutors, and automated analysis tools, while also democratizing access to state-of-the-art AI capabilities.