← News
Tech for Good Tech for Good Frontiers

The Optimizer That Teaches AI to Walk the Tightrope — Not Just Balance on Training Wheels

The Optimizer That Teaches AI to Walk the Tightrope — Not Just Balance on Training Wheels
Across Multiple TTF Settings test-time performance improvements Training gap reduced
Thomas Zhang Et Al. Authors
UC Berkeley & Carnegie Mellon Research institutions
Gradient-Wise + Activation-Wise Preconditioning Method combines
ArXiv, June 2026 Published
No Data/Architecture Change Needed drop-in improvement

Imagine training a chef only by watching them plate a single dish, then expecting them to cook a five-course meal perfectly. That's essentially how most AI systems are taught today — and it's a problem researchers have been wrestling with for years.

Thomas Zhang and collaborators at a team based at the University of California, Berkeley and Carnegie Mellon University have developed a new way to train AI systems that could change that. Their method, called DoPr (short for "double-preconditioning"), helps AI models perform better not just on paper, but in real-world situations where they must rely on their own earlier decisions to make later ones.

The challenge is what researchers call "test-time feedback" — the gap between how well an AI performs during training and how well it actually works when deployed. A robot trained to take one correct step might stumble after taking ten. A language model that writes a coherent paragraph might lose the thread entirely when generating a long essay. This gap tends to grow longer the more steps an AI must take.

Traditional training methods optimize for a single prediction at a time, using familiar math tricks like the Adam optimizer. But DoPr adds a second layer of intelligence borrowed from an older technique called Kronecker-Factored Approximate Curvature (KFAC). In plain terms, DoPr teaches AI systems to think ahead — to consider not just whether their next action is correct, but whether that action will set them up for success afterward.

In experiments, models trained with DoPr performed better across several real-world scenarios, from robot movement to language generation. But here's the twist: the improvements didn't always show up in standard validation scores. The researchers say this raises important new questions about how we measure whether an AI is truly learning well.

"We show that the addition of activation-wise preconditioning yields a drop-in intervention for increasing downstream model performance," the researchers wrote, noting that traditional metrics may be "insufficient" for capturing what matters in long-horizon tasks.

For everyday users, the implications could be significant. More robust AI systems — ones that don't unravel after a few wrong steps — could mean better robot assistants, more reliable AI tools, and AI that maintains quality across longer tasks. The work was published on arXiv in June 2026.

View on map →