IBM Accelerates AI Model Using Quantum Processor

Researchers from Multiverse Computing have successfully demonstrated a quantum enhancement of a large language model (LLM) using IBM's technology, specifically a 156-qubit Heron processor. This experiment marks a significant milestone as the first comprehensive quantum enhancement of an LLM on a superconducting processor, focusing on autoregressive text generation.

The study utilized Meta's Llama 3.1 8B model, which remained unaltered with frozen parameters. Instead, quantum adapters known as Cayley-parameterized unitary adapters (CUA) were incorporated, initially trained using classical methods before being integrated into a hybrid quantum-classical framework.

The testing was performed on the IBM Quantum System Two, an architecture tailored for hybrid quantum applications, leveraging the capabilities of the 156-qubit Heron chip. The hybrid model achieved a 1.4% reduction in perplexity for Llama 3.1 8B by introducing around 6,000 additional parameters, which is a mere 0.000075% of the model's total size.

In practical applications, the quantum-enhanced version of Llama demonstrated superior performance, accurately answering questions about astronomy and biology that the base model struggled with, such as details regarding gas giant planets and their rings.

Lead author Borja Aispuru emphasized that this work serves as a proof of concept, showing that quantum components can yield more precise predictions for text generation while using fewer computational resources. The team is optimistic about achieving further reductions in perplexity and enhancing accuracy with fewer parameters compared to traditional classical methods.

In related news, shares in quantum technology companies saw an uptick in May following the U.S. Department of Commerce's announcement of a $2 billion funding initiative for American firms under the CHIPS R&D program.