Rethinking How AI Systems Learn and Act

Scheutz and his team focus on robots that interact directly with humans, setting their work apart from screen-based large language models (LLMs) like ChatGPT or Gemini. Instead, they develop visual-language-action (VLA) models, which expand on LLMs by integrating vision and motor capabilities. These systems can process camera input alongside language instructions and perform physical actions, such as moving wheels, arms, or fingers.

Figure 1. AI Energy Use Reduced by 100× in New Breakthrough.

However, traditional VLA systems are computationally intensive and often unreliable, even for simple tasks like stacking blocks. A robot must analyze its environment, determine each block’s position, shape, and orientation, and execute instructions accordingly. Small issues—like shadows affecting perception, misaligned blocks, or unstable structures—can lead to failure. Figure 1 shows AI Energy Use Reduced by 100× in New Breakthrough.

These limitations mirror the weaknesses of LLMs. Just as robots struggle with physical tasks, chatbots can generate incorrect or fabricated outputs, such as inventing legal references or producing images with unrealistic details like extra fingers.

Symbolic reasoning provides a more efficient alternative. By applying general rules and abstract concepts—such as geometry or center of mass—systems can plan actions more reliably while reducing trial-and-error learning.

How Neuro-Symbolic Systems Improve Performance

“Like LLMs, VLA models rely on statistical patterns learned from large datasets, which can lead to errors,” Scheutz explains. “Neuro-symbolic VLAs, however, incorporate rule-based reasoning that reduces trial and error, enabling faster and more accurate solutions.”

In experiments using the classic Tower of Hanoi puzzle, neuro-symbolic VLA systems achieved a 95% success rate, compared to just 34% for conventional models. Even when tested on a more complex, unfamiliar version of the puzzle, they maintained a 78% success rate, while standard systems failed entirely.

Training efficiency also improved dramatically. The neuro-symbolic system required only 34 minutes to train, whereas a conventional VLA model took over a day and a half. Energy consumption dropped just as significantly—training used only 1% of the energy required by standard models, and operational energy use was reduced to just 5%.

Scheutz draws a parallel with LLMs like ChatGPT and Gemini, noting that these systems predict the next word or action in a sequence, which can lead to inaccuracies or “hallucinations.” He also highlights their disproportionate energy use, pointing out that AI-generated summaries in search engines can consume up to 100 times more energy than generating standard website listings.

Toward a More Sustainable AI Future

As AI adoption accelerates, demand for large-scale data centers continues to rise, with some facilities consuming hundreds of megawatts of power—far exceeding the energy needs of many small cities.

The researchers argue that while LLMs and VLA systems have driven rapid progress, they may not offer a sustainable or reliable long-term solution. Instead, hybrid neuro-symbolic AI presents a more efficient and dependable path forward, with the potential to significantly reduce energy demands while improving performance.

Source: SciTECHDaily

Cite this article:

Priyadharshini S (2026), A 100× Power Reduction Breakthrough That May Fix AI’s Energy Problem, AnaTechMaz, pp.946