Universidad
Politécnica de Madrid

EELISA European University Universidades Españolas con Ucrania

Integrating AI and Internet of Things using more efficient, resilient, and sustainable systems

UPM Researchers from the School of Telecommunications Engineering introduce innovative solutions to detect failures and substantially reduce energy consumption.

18.03.2026

The increase of artificial intelligence (AI) presents important challenges regarding its operational reliability and high energy consumption, especially when integrated into critical safety systems or certain connected devices such as those in the Internet of Things (IoT). In response to this reality, a team from the School of Telecommunications Engineering (ETSIT) at the Universidad Politécnica de Madrid (UPM) showed two significant advances: a methodology for failure detection and a software framework for energy efficiency. The researchers, members of the Grupo de Internet de Nueva Generación, collaborated on both projects with colleagues from Tianjin University, the University of Electronic Science and Technology of China, and Northeastern University (USA).

 

 

The first innovation, called Concurrent Linguistic Error Detection (CLED), is based on a simple but effective principle: the text generated by a large language model (LLM) must be consistent and follow the rules of the language. When hardware errors occur, the model often generates text with abnormal patterns. The proposed methodology detects these errors, identifying more than 87% of them at a computational cost of less than 1%. “The great advantage of CLED is that it works even with commercial models, which we can only access through an API [application programming interface],” says professor Pedro Reviriego, co-author of the studies.

The second proposal, the Adjustable Sequence Length (ASL) scheme, addresses the problem of energy consumption in neural networks that use stochastic computing [characterized by using random bit sequences instead of precise binary values]. This technology, ideal for the IoT due to its low power consumption, processes information as bit streams. The ASL technique optimizes this process by intelligently adjusting the accuracy of each layer of the neural network: it maintains high accuracy in the initial, more sensitive layers and reduces it in the final layers. Thanks to this method, energy savings and latent reductions of over 60% have been achieved with virtually no loss of accuracy.

 

 

“These developments demonstrate that it is possible to make AI safer and more efficient without compromising its performance,” explains professor Javier Conde, also an author of the papers. He adds that these advances reaffirm UPM’s position at “the forefront of research in computing architectures for the next generation of intelligent and sustainable systems.”

  1. . J. Zhu, J. Conde, Z. Gao, P. Reviriego, S. Liu and F. Lombardi. "Concurrent Linguistic Error Detection (CLED): A New Methodology for Error Detection in Large Language Models,". IEEE Transactions on Computers, vol. 74, no. 11, pp. 3638-3651, Nov. 2025.
  2. . Wang et al. "Energy-Efficient Stochastic Computing (SC) Neural Networks for Internet of Things Devices With Layer-Wise Adjustable Sequence Length (ASL)", IEEE Internet of Things Journal, vol. 12, no. 14, pp. 26955-26967, 15 July15, 2025.