Tufts Researchers Build AI That Uses 100x Less Energy With Higher Accuracy

Researchers at Tufts University have developed a neuro-symbolic AI system that uses up to 100 times less energy than conventional deep learning models while simultaneously achieving higher task accuracy -- a combination that has long been considered out of reach. The work, led by Professor Matthias Scheutz at the Tufts School of Engineering, blends traditional neural networks with symbolic reasoning in a way that mirrors how humans break complex problems into ordered, rule-governed steps.

The approach centers on separating high-level planning from low-level motor control. A symbolic planner written in the Planning Domain Definition Language (PDDL) handles task sequencing -- determining the correct order of actions based on explicit rules and goals -- while a compact learned neural module handles the physical execution of each step. This division of labor allows the system to train the neural component on a much narrower, well-defined problem space, dramatically cutting the compute needed. In tests on the Tower of Hanoi puzzle, a benchmark for sequential planning, the neuro-symbolic model achieved a 95% success rate compared to just 34% for standard end-to-end neural approaches. Training time dropped from more than 36 hours to just 34 minutes.

The energy implications are significant. Conventional large-scale AI training and inference already account for a growing share of global electricity consumption, and the trend is accelerating as model sizes grow. A system that delivers better results at a fraction of the energy cost has obvious appeal for embedded and edge applications -- mobile robots, wearable devices, on-device AI agents -- where battery life is a hard constraint. Scheutz notes that the architecture is designed to generalize: "We are building a framework that can be adapted to a wide range of robotic and autonomous tasks, not just toy problems."

The research will be formally presented at the International Conference on Robotics and Automation (ICRA) in Vienna in May 2026, where it is expected to attract significant attention from both the robotics and the broader AI efficiency communities. It arrives at a moment when the field is actively searching for architectures that can deliver capable AI without the runaway energy costs associated with scaling transformer-based models. Neuro-symbolic AI, long considered a niche academic pursuit, is increasingly being reexamined as a practical path toward efficient, interpretable, and robust intelligence.