U.S. AI systems and data centers consumed roughly 415 terawatt hours of electricity in 2024 — more than 10% of the country’s total energy output that year, according to the International Energy Agency. The figure is expected to double by 2030. A team at Tufts University thinks it has a path around that wall.
Researchers at Tufts’ School of Engineering built a proof-of-concept AI system that uses just 1% of the energy required to train a conventional model, and only 5% of the energy to run one — while outperforming standard approaches on accuracy. The work comes from the lab of Matthias Scheutz, Karol Family Applied Technology Professor, and will be presented at the International Conference of Robotics and Automation in Vienna in May.
The approach is called neuro-symbolic AI. It pairs conventional neural networks with symbolic reasoning — the kind of structured, rule-based logic humans use when they break a problem into steps. Rather than brute-forcing a solution through massive statistical computation, the system applies explicit rules to narrow down its search, reaching correct answers faster and with far less compute.
To test it, the researchers used a Tower of Hanoi puzzle — a classic problem where a robot must stack discs in a specific order without ever placing a larger disc on a smaller one. Standard visual-language-action (VLA) models, the robot AI equivalents of large language models, solved it correctly 34% of the time. The neuro-symbolic system hit 95%. On a more complex version of the puzzle the robot had never seen in training, standard VLAs failed every single attempt. The neuro-symbolic system succeeded 78% of the time.
Training time tells the same story. The neuro-symbolic model was ready in 34 minutes. The conventional VLA took more than a day and a half. The energy gap was even starker — the neuro-symbolic approach used 1% of the training energy and 5% of the execution energy of its competitor.
“Like an LLM, VLA models act on statistical results from large training sets of similar scenarios, but that can lead to errors,” Scheutz said. “A neuro-symbolic VLA can apply rules that limit the amount of trial and error during learning and get to a solution much faster.”
Scheutz draws a direct line to the hallucination problem that has dogged chatbots since their commercial debut — fabricated legal citations, six-fingered portraits, confident wrong answers. “These systems are just trying to predict the next word or action in a sequence, but that can be imperfect, and they can come up with inaccurate results or hallucinations,” he said. “Their energy expense is often disproportionate to the task. When you search on Google, the AI summary at the top of the page consumes up to 100 times more energy than the generation of the website listings.”
The distinction matters because the team is not working on chatbots. They study visual-language-action models, which are designed to run robots in the physical world — systems that must interpret camera feeds, understand spoken instructions, and move arms, legs, and fingers in real space. These models carry the same statistical baggage as LLMs but operate in environments where a wrong answer means a dropped package or a tipped-over stack of blocks, not a hallucinated footnote.
The paper, available on arXiv and authored by Timothy Duggan and colleagues, concludes that current LLMs and VLAs may not be the right foundation for energy-efficient, reliable AI and could push the industry against hard resource limits. The researchers position hybrid neuro-symbolic AI as a more sustainable alternative — not a replacement for neural networks, but a way to impose structure on their otherwise unconstrained search for solutions.
The work is a preprint and has not yet been peer-reviewed. The full paper will appear in the ICRA 2026 conference proceedings.
