The newly developed approach uniquely integrates conventional neural network architectures with symbolic reasoning methods, mimicking human cognitive processes that break tasks and concepts into hierarchical steps and categorical abstractions. This hybrid model, pioneered under the guidance of Matthias Scheutz, Karol Family Applied Technology Professor, leverages neuro-symbolic AI to achieve energy efficiency improvements as high as a hundredfold compared to standard AI systems, all while delivering superior accuracy in complex task execution. The team plans to present their groundbreaking research at the upcoming International Conference of Robotics and Automation, reinforcing the crucial intersection of AI, robotics, and sustainable energy use.

Unlike mainstream large language models (LLMs) such as ChatGPT or Gemini, which primarily operate within screen-based environments, this research focuses on visual-language-action (VLA) models. VLAs extend conventional LLMs by incorporating visual and motor capabilities, enabling robots to perceive their environment using cameras and execute physical actions—like moving wheels, arms, or fingers—based on language commands. The combination of visual perception with linguistic processing is essential for robots interacting meaningfully with humans and their surroundings, but it traditionally comes with vast computational and energy expenses.

In real-world scenarios, resource-intensive VLA models frequently struggle with seemingly straightforward tasks. For example, a robot tasked with stacking blocks to build a tower must identify each block’s location, shape, and orientation while interpreting the instruction with high precision. Conventional neural networks may falter due to ambiguous visual inputs like shadows or reflections, resulting in errors such as incorrectly grasping or placing the blocks, potentially causing instability in the tower. Such failures parallel inaccuracies seen in LLM text outputs, where hallucinations—fabricated or false information—can degrade reliability drastically.

The neuro-symbolic system addresses these limitations by introducing a symbolic reasoning layer that leverages abstract grouping and rule-based generalization. Instead of relying purely on large-scale pattern recognition, this AI applies logical constraints and puzzle-specific heuristics that guide decision-making. This approach yields more robust and energy-efficient learning, minimizing the trial-and-error cycles that typically consume excessive computational power in conventional VLA models.

One compelling demonstration of this system’s superior performance involved the Tower of Hanoi—a classical puzzle used in AI research. Testing revealed the neuro-symbolic VLA achieved an impressive 95% success rate on the standard puzzle configuration, significantly outperforming the 34% success rate of traditional VLAs. Even more striking, when challenged with a novel, more complex variation of the puzzle it had never encountered before, the neuro-symbolic model maintained a 78% success rate, whereas the conventional models failed entirely to solve the puzzle. These results highlight the system’s potential for robust generalization beyond the training scenarios.

Beyond accuracy gains, the energy savings realized through neuro-symbolic AI are transformative. Training the neuro-symbolic VLA required only 34 minutes, a drastic reduction from the more than 36 hours needed for the conventional model’s training. Furthermore, the energy consumption for this training was merely 1% of that used by the standard approach. During operational use, the neuro-symbolic system continued to consume just around 5% of the energy needed by a conventional VLA, demonstrating remarkable efficiency across both learning and task execution stages.

The implications of adopting neuro-symbolic frameworks extend well beyond robotics. Scheutz points to analogous inefficiencies in popular LLM-powered applications, such as AI-driven search engine summaries, which can expend up to 100 times more energy than generating the underlying webpage content itself. These disproportionate energy costs raise broader concerns about the sustainability and scalability of current AI architectures, especially as demand increases exponentially and models become larger and more complex.

At present, the relentless drive toward ever-larger AI models and data center capacities fuels a competitive arms race, with monumental server facilities consuming power equivalent to that of entire small cities. Examples include Sandia National Laboratory’s servers, xAI Colossus in Memphis, and massive projects like Microsoft and OpenAI’s Stargate in construction. Such infrastructures require hundreds of megawatts to operate, posing a fundamental challenge to the sustainability of AI deployment if underlying computing models remain energy inefficient.

Given these realities, the researchers argue that while large language models and conventional VLAs have garnered immense popularity and demonstrated efficacy in many applications, they may be hitting a resource-driven bottleneck. It is likely that continuing to expand their scale and complexity without addressing fundamental computational inefficiencies will lead to untenable environmental and economic costs. Instead, a shift to hybrid neuro-symbolic AI architectures represents a promising alternative pathway that balances high performance with dramatically reduced energy footprints.

In summary, this innovative neuro-symbolic approach not only advances AI reliability and accuracy but pioneers a path toward environmentally sustainable artificial intelligence technologies. By mimicking human-like cognitive processes through the integration of neural networks with symbolic reasoning, the new system enables faster learning, greater task generalization, and substantial energy savings. As AI becomes increasingly embedded in everyday robotics and industrial automation, adopting such efficient methodologies will be critical to meet future performance demands without exacerbating global energy consumption and environmental impacts.

The upcoming presentation in Vienna at the International Conference of Robotics and Automation will further detail these insights and experimental validations, drawing attention to the pressing need for energy-conscious AI design. This research could serve as a blueprint for the next generation of AI systems—ones that are not only smarter but more sustainable and responsible in their resource use. The days of AI powered by sprawling, power-hungry data centers may soon give way to leaner, neuro-symbolic intelligences capable of thriving in resource-constrained environments while exceeding current performance benchmarks.

Subject of Research:
Not applicable

Article Title:
The Price is Not Right: Neuro-Symbolic Approaches Significantly Outperform VLAs in Performance as well as Computational Cost and Energy Efficiency

News Publication Date:
5-Jun-2026

Image Credits:
Sandia National Laboratory

Keywords:
Artificial intelligence, Energy resources, Energy resources conservation, Water resources, Robotics

A groundbreaking study recently published in the prestigious journal Nature has unveiled significant progress in this field, showcasing collaborative efforts from leading global institutions including Politecnico di Milano, École Polytechnique Fédérale de Lausanne, Stanford University, University of Cambridge, and the Max Planck Institute. This research delves into the core challenge of training physical neural networks—an area that until now had largely been conceptual, hindered by the difficulty of adapting learning algorithms to operate within analogue physical substrates rather than traditional digital simulations.

At the heart of this innovation is the photonic microchip developed by the research team at Politecnico di Milano, which stands as a testament to the power of integrated photonic technologies. Unlike conventional electronic chips that rely on voltage and current states to represent and manipulate data, these photonic chips utilize the interference patterns of light waves to perform fundamental mathematical operations such as summations and multiplications. Remarkably, these operations occur on silicon microchips mere square millimeters in size, highlighting the extraordinary miniaturization and integration capabilities that photonics enables in hardware design.

This approach fundamentally transforms how data is processed. By bypassing the digital conversion steps that are traditionally necessary in electronic neural networks, photonic chips execute computational tasks natively in the analogue domain. According to Francesco Morichetti, the head of the Photonic Devices Lab at Politecnico di Milano, this leads to a profound reduction in both energy consumption and processing time. Such an advancement is pivotal, considering that modern AI data centers are notorious for their massive energy footprints. The capacity to perform AI computations in a more sustainable and efficient manner could redefine the environmental and economic landscape of machine learning.

Crucially, the study addresses one of the most vital aspects of neural network functionality: training. Training involves iteratively adjusting the network’s parameters until it successfully performs specific tasks or recognizes patterns in data. Traditionally, this process is computationally intensive and carried out within digital or software-based frameworks before deploying trained models onto hardware. The collaborative research introduces an innovative “in-situ” training methodology for photonic neural networks, whereby learning is conducted entirely within the physical system using light signals, without recourse to digital simulations.

This photonics-based in-situ training method embodies a paradigm shift by treating the physical network as the training medium itself. It leverages the intrinsic physics of light interference to update network weights and parameters directly on the photonic chip. Morichetti explains that this eliminates layers of data conversion and transfer between physical and digital domains, resulting in accelerated training speeds with enhanced robustness and efficiency. The potential to execute training faster, with lower latency, unlocks new possibilities for adaptive AI systems that can learn and evolve in real-time.

The implications of these advancements stretch far beyond laboratory prototypes. Photonic chips with optical neural network architectures could enable the creation of more sophisticated AI models capable of handling complex computations at the edge—on devices such as autonomous vehicles, drones, and portable sensors—without relying on centralized cloud processing. This decentralization of AI computation means systems can process real-time data locally, reducing communication lag, enhancing privacy, and improving resilience against network disruptions.

Moreover, photonic neural networks embody a confluence of multiple cutting-edge scientific disciplines, spanning applied physics, nanophotonics, optical materials, and computer science. The ability to engineer nanophotonic structures that manipulate photons with extreme precision enables the design of highly customizable neural architectures adapted for specific application domains. As research advances, this technology could birth new classes of AI hardware that operate fundamentally differently from silicon-based electronic processors, potentially surpassing them in speed, energy efficiency, and scalability.

One of the technical marvels enabling these developments is the exploitation of light interference—a phenomenon where multiple light waves overlap, amplifying or diminishing each other to yield precise mathematical outcomes on-chip. These analogue computations, inherently parallel and swift, contrast starkly with sequential electronic logic operations in conventional processors, offering significant advantages in throughput and power efficiency.

The collaborative nature of this research is also notable, bringing together experts from diverse institutions across Europe and the United States, each contributing unique expertise in photonics, AI, and applied sciences. Such interdisciplinary efforts underscore the complexity of transitioning physical neural networks from theoretical constructs to viable practical technologies. The paper, titled “Training of Physical Neural Networks,” provides a detailed commentary on the state of research in this domain and charts future pathways for overcoming remaining challenges.

In conclusion, the advent of physical neural networks powered by integrated photonic chips heralds a transformative step toward the next generation of intelligent machines. By leveraging the laws of physics to perform computational operations natively and efficiently, these systems promise to transcend the limitations of traditional digital computing architectures. Beyond merely accelerating AI workloads, they lay the groundwork for sustainable, adaptive, and decentralized intelligence embedded directly within everyday devices. The prospect of AI computations unfolding at the speed of light within miniature silicon photonic circuits marks a thrilling confluence of physics, engineering, and machine learning, poised to redefine how we conceive and deploy intelligent systems in the near future.

Subject of Research: Not applicable

Article Title: Training of physical neural networks

News Publication Date: 3-Sep-2025

Web References: DOI: 10.1038/s41586-025-09384-2

Image Credits: Politecnico di Milano, DEIB – Department of Electronics, Information and Bioengineering

Keywords

Artificial neural networks, Applied sciences and engineering, Physical sciences, Physics, Optical materials, Nanophotonics, Photonics, Applied optics, Neural net processing, Computer science, Computer processing, Generative AI, Machine learning

Neurosymbolic AI merges the strengths of traditional symbolic reasoning with the robust capabilities of data-driven neural networks. This hybrid model draws inspiration from the human brain, whose efficient operations require only about 20 watts of power while demonstrating rapid and reflective thinking. The brain’s ability to process information efficiently stands in stark contrast to the energy-hungry operations of current AI systems, which often necessitate extensive computational resources. By examining the underlying principles of cognitive functioning, researchers envision a more sustainable AI landscape where smaller entities can compete with larger corporations that currently dominate the scene.

At the core of neurosymbolic AI is the utilization of semantically meaningful symbols to structure and manipulate knowledge. These symbolic approaches—grounded in logic and mathematical principles, such as differential equations—can streamline the cognitive processes of AI systems. Rather than relying solely on vast datasets to uncover correlations, neurosymbolic models can learn foundational axioms or facts from smaller data samples. This empowers AI systems to infer related truths through the application of symbolic logic, thereby drastically reducing the volume of data and computational demands typically required for generating insights.

Consider the process in which traditional LLMs learn complex relations from considerable data inputs. The range of training necessary often leads to the consumption of resources on a monumental scale. Conversely, neurosymbolic AI offers a more efficient pathway, where AI systems can derive straightforward and profound truths—much like humans do. For instance, from an axiom such as “All men are mortal” and the fact that “Socrates is a man,” the system can ascertain the derived conclusion that “Socrates is mortal.” Such capabilities illustrate the potential of integrating symbolic reasoning with machine learning, presenting a compelling case for reducing the size and energy consumption of AI models.

Analyses conducted by Velasquez and his team suggest that neurosymbolic models could be up to 100 times smaller than contemporary leading LLMs, heralding a significant shift in the AI landscape. This qualitative change not only has the potential to democratize AI technology—allowing smaller companies and research institutions to participate—but also promotes an arena where resource allocation is far more equitable. The implications of such a shift extend beyond corporate competition; they embody a vision of technology that aligns with environmental stewardship and responsibility.

As AI researchers continue to explore the full potential of neurosymbolic AI, the prospects for enhancing trust and reliability in AI systems grow increasingly optimistic. Trust in AI technologies is paramount, especially as these systems become more integrated into everyday life. The principles underlying neurosymbolic AI support the creation of transparent and interpretable models that mirror human reasoning processes. This transparency is essential for establishing confidence in AI outputs, which can often appear opaque or enigmatic, particularly within traditional neural network frameworks.

Moreover, given the mounting concerns regarding the environmental impact of extensive computing operations, the time has come to rethink how society approaches the development and deployment of AI technologies. Neurosymbolic AI embodies a critical step toward a more sustainable model, enabling the delivery of advanced capabilities while minimizing energy consumption. As stakeholders across various sectors assess their tech-related responsibilities, the emergence of neurosymbolic AI fosters a long overdue dialogue about the ethical ramifications of artificial intelligence and its role in society.

Improvements in efficiency and sustainability are especially crucial given the rapid pace of technological advancement. As businesses rush to harness AI capabilities, harnessing the potential of neurosymbolic principles could offer a crucial lifeline in somewhat turbulent waters. Strategic advancements in this field hold the promise of creating AI systems that do not merely reflect the status quo but redefine how technology interacts with human needs and environmental concerns.

Furthermore, operationalizing neurosymbolic methodologies also beckons a reexamination of data governance and accessibility. By requiring less data to train effective models, neurosymbolic AI makes strides not only in performance but also in the ethical dimensions of data usage. This less resource-intensive approach reduces the risk of pervasive surveillance and the monopolization of data, issues that have emerged alongside the rise of AI technologies driven by large datasets.

In conclusion, as the AI industry grapples with its energy demands and ecological footprint, innovations like neurosymbolic AI present an empowering vision for the future. By championing a hybrid approach, researchers pave the way for more democratized access to AI technology that does not sacrifice environmental sustainability or ethical rigor. Embracing the principles of neurosymbolic AI could very well revolutionize the ecological narrative surrounding AI development, empowering a diverse range of contributors to engage in this transformative journey.

The implications of these advancements are boundless, marking a potential turning point in the relationship between technology and nature. It is not only a promise of sustainable AI but also a call to action for society to foster responsible innovation. As we strive to develop technology that aligns with the needs of the planet and its inhabitants, neurosymbolic AI may well become the hallmark of a new era in artificial intelligence.

Subject of Research: Neurosymbolic AI
Article Title: Neurosymbolic AI as an antithesis to scaling laws
News Publication Date: 20-May-2025
Web References:
References:
Image Credits:

Keywords

Artificial intelligence