A team of researchers at Tufts University has developed a neuro-symbolic AI that can reduce energy requirements during exercise by up to 99 percent. Instead of 36 hours, she only needed 34 minutes for a test.
The hunger for energy is growing rapidly thanks to artificial intelligence. In the US alone, AI systems and data centers consumed approximately 415 terawatt hours of electricity in 2024. That’s more than the UK’s entire electricity consumption. Data centers and AI systems therefore account for more than ten percent of total national energy production.
According to forecasts by the International Energy Agency, this figure is expected to double by 2030. That’s why researchers are looking for ways to make systems more efficient without causing electricity costs to skyrocket. A new technical approach from Tufts University promises a significant remedy here.
What makes neuro-symbolic AI different from ChatGPT?
The work of Matthias Scheutz and his team at the Tufts University School of Engineering is behind the development. The scientists use so-called neuro-symbolic AI. They combine classic neural networks with fixed logical rules, similar to how people solve problems in steps and categories.
The research focuses primarily on robots that interact directly with people. The researchers use so-called visual language action models (VLA), which expand conventional language models such as ChatGPT or Gemini to include vision and movement. By applying general rules, these systems understand concepts such as the shape or center of gravity of an object much better.
This is how the new approach beats classic AI
Matthias Scheutz, Professor of Applied Technology, said: “Similar to an LLM, VLA models rely on statistical results from large training data sets with similar scenarios, but this can lead to errors. A neurosymbolic VLA can apply rules that limit the amount of trial and error during the learning process and thus arrive at a solution much more quickly. Not only does it complete the task much faster, but the time required to train the system is also significantly reduced.”
In experiments with the Tower of Hanoi puzzle, Scheutz’s system achieved a success rate of 95 percent. Conventional models only achieved 34 percent on the same task. Even with completely unknown tasks, the new technology excelled with 78 percent success, while conventional AI systems aborted every attempt.
The hybrid approach noticeably improves planning and makes it more reliable overall. In addition, the neuro-symbolic method significantly reduces the amount of trial and error required during the learning phase. The experiments delivered measurable successes in terms of time spent on the necessary learning.
While standard models took over a day and a half to train, the new system was ready in just 34 minutes. The time required for the learning process is reduced significantly. A significant increase in efficiency affects both computing time and direct power consumption.
One percent energy: The savings in detail
The savings in actual power consumption of the technology were just as significant. The training required only one percent of the energy of conventional models. During ongoing operation, Matthias Scheutz’s technology only consumed five percent of the energy.
Current AI summaries in search engines often use 100 times more energy than simple hit lists. This is exactly where the solution from Tufts comes in to make such computationally intensive tasks sustainable through rule-based methods. For many everyday tasks, the current massive use of energy is disproportionate to the benefits.
According to the researchers, today’s models do not provide a long-term, sustainable foundation for the future. Timothy Duggan, Pierrick Lorang, Hong Lu and Matthias Scheutz published their results on arXiv in February 2026. Their concept serves as a necessary alternative to previous resource-intensive models.
Why data centers now have to recalculate
With such innovations, sustainable operation of large data centers is once again within reach. The neuro-symbolic method offers an efficient basis for future developments. The results are forcing companies to completely recalculate the energy requirements of future data centers.
The research paper has not yet gone through the so-called peer review process – in other words: it has not yet been checked by independent scientists. While the results seem impressive, it is important to put them in perspective.
The tests have so far been carried out with a comparatively simple puzzle, not with large language models such as ChatGPT or Google Gemini. It remains to be seen whether the savings of 99 percent can also be transferred to models with billions of parameters. Nevertheless, the approach provides promising food for thought – precisely because it shows that not every AI problem has to be solved with raw computing power.
Also interesting: