Multiverse Computing Cuts LLM Perplexity 1.4% on 156-Qubit Processor

Researchers at Multiverse Computing have demonstrated a 1.4 percent improvement in perplexity, a measure of how well a language model predicts a sample, using a novel approach to enhance the performance of large language models on quantum hardware. The improvement was achieved by integrating Cayley-parameterised unitary adapters into a pre-trained 8-billion-parameter Llama 3.1 8B model and executing it on a 156-qubit IBM Quantum System Two superconducting processor. This marks a practical demonstration of quantum computing’s potential to address limitations in artificial intelligence model performance with a minimal increase of only 6,000 parameters. The researchers write that this method could overcome the unsustainable expansion of compute infrastructure required by increasingly large models. A systematic study also revealed 83 percent recovery of compression-induced degradation, and correct answers to questions that classical baselines failed to address.

LLM Parameter Scaling & Classical Limitations

The scaling of large language models (LLMs) is encountering a fundamental constraint; every trainable parameter occupies classical memory, and scaling a deployed model’s parameter count requires an unsustainable expansion of compute infrastructure, according to research published by Multiverse Computing. While techniques like quantisation and pruning offer some relief, they trade expressive capacity for reduced footprint, prompting exploration into alternative computational paradigms. Quantum computing presents a potential solution, leveraging the exponentially larger Hilbert space accessible with each added qubit, but practical demonstrations on models of significant scale have remained elusive until recently. Multiverse Computing has demonstrated an 8-billion-parameter model, and the team emphasizes this result is analogous to the experimental realisation of Shor’s algorithm, signifying a crucial step toward viable quantum-enhanced AI. Further investigation using SmolLM2, a 135-million parameter model, revealed improving perplexity with unitary block dimension, and 83 percent recovery of performance lost during compression.

This systematic study highlights the potential for quantum adapters to enhance model capacity and mitigate the drawbacks of existing compression techniques, offering a pathway toward more efficient and powerful LLMs. The researchers believe this work establishes a concrete basis for future scaling of quantum-classical hybrid models.

Cayley Unitary Adapters for LLM Integration

Multiverse Computing’s recent work demonstrates a novel approach to enhancing performance through quantum integration, beyond simply adding more parameters to large language models. The team pioneered the use of Cayley-parameterised unitary adapters, a specific quantum technique designed to be inserted into existing LLM architectures, rather than replacing them wholesale. This method introduces quantum-based parameters without demanding a complete overhaul of established AI infrastructure, a crucial step toward practical quantum-enhanced AI. The adapters are constructed to be hardware-efficient, sidestepping the exponential scaling issues that plague many quantum algorithms. A key innovation lies in the construction of these adapters, allowing for parallel execution of shallow circuits on current quantum hardware. Each block is a 2-qubit unitary, a depth-19 native-gate circuit on ibm_basquecountry, ensuring compatibility with existing coherence limits. The application of these adapters to the 8-billion-parameter Llama 3.1 8B model resulted in a 1.4 percent improvement in perplexity, using only 6,000 additional parameters. This provides a mechanistic understanding of how these adapters function, and validates the approach as a scalable route to quantum enhancement of contemporary LLMs.

SmolLM2 Perplexity Improvement with Unitary Blocks

Multiverse Computing researchers are actively investigating methods to enhance large language model performance through quantum processing, with recent work focusing on a systematic analysis of the SmolLM2 model. This 135-million parameter language model served as a crucial testbed for exploring the impact of quantum-enhanced components on language generation capabilities. Unlike attempts to scale up existing models directly, the team implemented quantum circuit blocks integrated into the model’s projection layers, and assessed their effect on perplexity, a standard measure of a language model’s predictive accuracy. The researchers achieved 83 percent recovery of compression-induced degradation, demonstrating the potential of these adapters to mitigate performance loss when models are compressed for efficiency. This is significant as model size remains a major constraint in the field.

Beyond improving performance, the team identified a sharp noise, expressivity phase transition identifying a concrete path to quantum utility at larger qubit scales, suggesting a clear pathway for scaling these quantum enhancements as hardware improves. This detailed analysis of SmolLM2 provides a mechanistic understanding that underpins the team’s broader success with the 8-billion-parameter Llama 3.1 8B model, where they demonstrated a 1.4 percent improvement in perplexity using a 156-qubit IBM Quantum System Two processor.

Hardware-Efficient Block-Diagonal Unitary Construction

The potential for quantum computing to accelerate artificial intelligence is increasingly focused on resource-efficient methods, and recent work from Multiverse Computing demonstrates a practical approach to integrating quantum processing into large language models. Rather than attempting full quantum implementations, researchers pioneered the use of block-diagonal unitaries (BDU) as adaptable components within existing LLMs, achieving a 1.4 percent improvement in perplexity with Llama 3.1 8B. This small gain is significant, suggesting a viable pathway to enhance AI performance without requiring a complete overhaul of current infrastructure. The design prioritises hardware efficiency; a generic unitary would demand exponentially growing computational resources, but the BDU construction allows for parallel execution of shallow circuits. The researchers explain that this allows for complex operations with manageable qubit requirements. The team views this as a foundational step, akin to early demonstrations of quantum algorithms, establishing a concrete basis for future scaling and exploration of quantum utility in artificial intelligence.

Llama 3.1 8B Enhancement on IBM Quantum System Two

Researchers successfully enhanced the Llama 3.1 8B large language model, achieving a 1.4 percent improvement in perplexity, a metric of predictive accuracy, by integrating quantum processing into its architecture. The team’s approach centers on block-diagonal unitaries, a construction designed for hardware efficiency. Unlike generic unitary transformations requiring exponentially scaling resources, these adapters leverage the parallel processing capabilities of the quantum computer, utilizing shallow, two-qubit circuits. This allows for end-to-end inference directly on the QPU, validating the concept with a production-scale LLM, as the researchers state. Further investigation using the smaller SmolLM2 model (135 million parameters) revealed that perplexity improved monotonically with increasing unitary block dimension, and importantly, showed 83 percent recovery of compression-induced degradation. This systematic study provides mechanistic insight into the observed performance gains and suggests a pathway toward realizing quantum utility at larger qubit scales, even with the inherent challenges of noise present in current quantum hardware.

Noise-Expressivity Phase Transition & Quantum Utility

The ability of quantum computers to enhance artificial intelligence isn’t solely about achieving higher numbers; it’s about navigating a delicate balance between quantum noise and expressive power. Researchers, led by Borja Aizpurua at Multiverse Computing, have identified a “sharp noise–expressivity phase transition” crucial for realizing practical quantum utility, particularly in large language models. Their work, detailed in a recent pre-print, demonstrates that beyond a certain qubit scale, the benefits of quantum computation become increasingly apparent, even amidst inherent hardware imperfections. This transition is revealed through a systematic study employing the SmolLM2 model (135 million parameters), allowing for exhaustive experimentation impossible with larger, more complex LLMs. By varying the size of the quantum adapters, the team observed a measure of language model prediction accuracy. This isn’t simply about squeezing more performance from existing models, but about unlocking capabilities previously inaccessible, with an 8-billion-parameter model achieved with only 6,000 additional parameters. The team’s approach, utilizing Cayley-parameterised unitary adapters, offers a hardware-efficient method for integrating quantum parameters into classical LLMs, paving the way for future scaling and innovation.

Prior Quantum Approaches to Language Models

Quantum-enhanced language models are not merely a theoretical pursuit; prior investigations have already begun to map a path toward practical implementation, though challenges remain. Before the demonstration of improvements with Cayley-parameterised unitary adapters, researchers explored various avenues for integrating quantum mechanics with natural language processing. Early work focused on restricted regimes, including quantum machine learning applied to text classification, and quantum natural language processing limited to simplified grammatical structures. Variational sequence models and hybrid methods for LLM fine-tuning also emerged, with one study demonstrating quantum self-attention on a 72-qubit processor specifically for text classification. These initial forays, while promising, often faced limitations.

Approaches were frequently confined to simulations rather than real quantum hardware, or concentrated on classification tasks instead of the more complex autoregressive generation central to modern LLMs, and many operated at a linguistic scale far removed from the parameters of production-scale models. To the best of our knowledge, no prior work has demonstrated quantum enhancement of a production-scale, pre-trained LLM for autoregressive language generation on real gate-based quantum hardware, highlighting the novelty of recent advancements. Other techniques included quantum knowledge distillation, multi-architecture frameworks for quantum-enhanced natural language generation, and even reinterpretations of transformer layers as unitary operators. However, these methods, while innovative, still faced hurdles in scaling to models with billions of parameters. The current work builds upon this foundation, introducing a block-diagonal unitary approach designed for hardware efficiency and scalability, aiming to overcome the limitations of earlier quantum-classical hybrid models.

Authors & Affiliations, Multiverse Computing Collaboration

Borja Aizpurua of Multiverse Computing, alongside colleagues, spearheaded the research detailed in a recent publication concerning quantum-enhanced large language models. Aizpurua’s affiliation extends to the Parque Científico y Tecnológico de Gipuzkoa in Spain and the Department of Basic Sciences at Tecnun, University of Navarra, highlighting a collaborative approach between industry and academia. Sukhbinder Singh, also at Multiverse Computing with a base for Social Innovation in Toronto, Canada, contributed to the work, as did Augustine Kshetrimayum and Saeed S. The team’s expertise is further enriched by Román Orús, affiliated with both Multiverse Computing and the Donostia International Physics Center in San Sebastián, Spain, and the Ikerbasque Foundation for Science in Bilbao. This multi-institutional involvement underscores the complexity of bridging quantum computing and artificial intelligence. The research was published on arXiv.org under a perpetual non-exclusive license, detailing the integration of Cayley-parameterised unitary adapters into an 8-billion-parameter model, achieving a 1.4 percent improvement in perplexity. This advancement demonstrates the viability of the physical approach and establishes a basis for future scaling.