A model of artificial intelligence, after partial training on a quantum computer, gave correct answers where it previously made mistakes.
Researchers at Multiverse Computing have demonstrated that adding a very small number of parameters, executed on a quantum machine, can significantly improve the accuracy of a large language model. Their work, published on the platform
arXiv, constitutes a first proof that quantum computing can enhance AI systems already deployed at scale.
To understand the mechanism behind this improvement, we must look at a key indicator called "perplexity". The lower the perplexity of a model, the better it is at predicting the next word in a sentence. Traditionally, to reduce perplexity, the number of parameters was increased, which made the infrastructure heavier. But the researchers found an alternative: using specialized quantum blocks, called Cayley unitary adapters, which require only an infinitesimal increase in the number of parameters.
Concretely, the scientists took Meta's Llama 3.1 8B model, which has 8 billion parameters. They froze its original parameters and inserted the Cayley adapters, previously trained on a classical computer. The whole setup was executed on the IBM Quantum System Two quantum processor, equipped with 156 qubits.
The result? A 1.4% decrease in perplexity with only 6,000 additional parameters, which is an increase of 0.000075%. Such a level of gain with so few modifications is spectacular.
Thus, the tests revealed concrete improvements. For example, the original model incorrectly answered an astronomy question about giant planets, claiming that only Saturn has rings. After the quantum addition, the hybrid model correctly identified that all Jovian planets have rings. Similarly, in biology, a question about the consequences of gene flow was better handled by the improved version.
Furthermore, the authors of the study explain that this approach opens the door to hybrid AI systems, combining the best of classical and quantum computing. The ultimate goal is to achieve "quantum supremacy", where a quantum computer performs tasks impossible for any classical computer.
For now, the main difficulty in producing quantum computers at an industrial scale remains noise, i.e., errors caused by environmental disturbances during quantum computations. Only a few simple calculations on laboratory quantum computers are currently feasible.