Quantum Techniques for Improving AI Model Training

Quantum computing promises to making LLMs more efficient. And it's already working on real hardware. Efficient fine-tuning of large language models remains a critical bottleneck in AI development, with most researchers focused on purely classical computing approaches. A new paper from Chinese researchers demonstrates how quantum computing principles can dramatically reduce the parameters needed while improving model performance. The team introduces Quantum Weighted Tensor Hybrid Network (QWTHN), which combines quantum neural networks with tensor decomposition techniques to overcome the expressive limitations of traditional Low-Rank Adaptation (LoRA). By leveraging quantum state superposition and entanglement, their approach achieves remarkable efficiency: reducing trainable parameters by 76% while simultaneously improving performance by up to 15% on benchmark datasets. Most importantly, this isn't just theoretical - they've successfully implemented inference on actual quantum computing hardware. This represents a tangible advancement in making quantum computing practical for AI applications, demonstrating that even current-generation quantum devices can enhance the capabilities of billion-parameter language models. The integration of quantum techniques into traditional deep learning frameworks might become standard practice for resource-efficient AI development in the future. More on Quantum Hybrid Networks and other AI highlights in this week's LLM Watch:

Training ML on ground-state physics using quantum computing experiments has so far been limited to small systems or highly structured states. In our new preprint, we push this paradigm to 2D systems of up to 115 qubits. Out on the arXiv today https://lnkd.in/eUzgeWHH, we demonstrate learning on ground states of the interacting 2D Heisenberg XXZ model. For system sizes of 57 and 115 qubits, we generate a dataset of 1-, 2- and 12-local Pauli observables that agree with DMRG to within a few percent across most of the antiferromagnetic phase. Training neural networks on this data, we find they can accurately predict specific spatial patterns of observables across the lattice for previously unseen Hamiltonians. The models generalise the behaviour of the underlying quantum experiments, both within the training distribution and in an out-of-distribution regime approaching the phase boundary. Ultimately, our results help us push towards a regime where AI may be trained on data from quantum processors that are fundamentally beyond the reach of classical approximation methods. A huge thanks to my co-authors across IBM Quantum, the University of Oxford, and the STFC Hartree Centre including Freya Shah, Minjun Jeon, M. Emre Şahin, Christa Zoufal, Kunal Sharma IBM STFC

🚀 New Paper on arXiv! I’m excited to share our latest work: “Learning to Program Quantum Measurements for Machine Learning” 📌 arXiv: https://lnkd.in/euRhBQJM 👥 With Huan-Hsin Tseng (Brookhaven National Lab), Hsin-Yi Lin (Seton Hall University), and Shinjae Yoo (BNL) In this paper, we challenge a long-standing limitation in quantum machine learning: static measurements. Most QML models rely on fixed observables (e.g., Pauli-Z), limiting the expressivity of the output space. We take this one step further--by making the quantum observable (Hermitian matrix) a learnable, input-conditioned component, programmed dynamically by a neural network. 🧠 Our approach integrates: 1. A Fast Weight Programmer (FWP) that generates both VQC rotation parameters and quantum observables 2. A differentiable, end-to-end architecture for measurement programming 3. A geometric formulation based on Hermitian fiber bundles to describe quantum measurements over data manifolds 🧪 Experiments on noisy datasets (make_moons, make_circles, and high-dimensional classification) show that our dual-generator model outperforms all traditional baselines—achieving faster convergence, higher accuracy, and stronger generalization even under severe noise. We believe this work opens the door to adaptive quantum measurements and paves the way toward more expressive and robust QML models. If you're working on QML, differentiable quantum programming, or quantum meta-learning, I’d love to connect! #QuantumMachineLearning #QuantumComputing #QML #FastWeightProgrammer #DifferentiableQuantumProgramming #arXiv #HybridAI #AI #Quantum

NVIDIA’s launch of "Ising" marks the introduction of the world’s first open-source #AI model family purpose-built for #quantum #computing workflows. The platform targets two of the most critical bottlenecks in quantum systems—processor calibration and real-time error correction—by embedding AI directly into quantum control loops. Released across developer ecosystems (GitHub, Hugging Face) and integrated with CUDA-Q, Ising positions AI as the #orchestration layer for hybrid quantum-classical computing. Early adoption by institutions such as Fermilab and Harvard University signals immediate traction in #research. Strategically, this launch reframes AI not just as an application layer, but as foundational infrastructure for scalable, fault-tolerant quantum systems. Ising is fundamentally differentiated by its dual-model architecture: a 35B-parameter vision-language model for automated quantum calibration and a #3D CNN-based decoder for real-time quantum error correction. This architecture replaces manual calibration workflows with agentic AI pipelines, achieving up to 2.5× faster and 3× more accurate decoding while requiring significantly less training #data. Technically, it integrates tightly with NVIDIA’s CUDA-Q stack and NVQLink interconnect, enabling low-latency coupling between GPUs and quantum processing units (QPUs). Unlike generative AI models, Ising operates as a physics-aware control system, optimized for noisy qubit environments and scalable to millions of qubits, effectively acting as an AI control plane for quantum hardware. The Ising launch materially reshapes the quantum ecosystem by positioning NVIDIA as the control-plane leader in quantum computing, despite not manufacturing quantum hardware. It accelerates commercialization timelines by addressing error correction—widely seen as the primary barrier to the development of useful quantum systems. Market response was immediate, with quantum stocks (IonQ, Rigetti Computing, D-Wave) surging on expectations of faster industry maturation. Strategically, Ising challenges incumbents by shifting value from hardware-centric differentiation to AI-driven orchestration, thereby reinforcing a hybrid architecture in which GPUs and QPUs co-evolve. This positions NVIDIA as a central enabler across competing quantum vendors, potentially standardizing its ecosystem as the de facto operating layer for quantum-AI #convergence. These architectures intensify system autonomy and complexity, requiring dynamic governance models and adaptive #cyber-#ethics to continuously monitor, audit, and recalibrate #risks across hybrid quantum-AI control planes. #strategy #governance #business #investments #technology #future #digital

Quantum-inspired machine learning will continue to play a significant role improving Artificial Intelligence performance. This is primarily due to myriad available applications utilizing efficient tensor network approximations that can be applied to complex or quantum systems. [01] Tensor networks, alongside active areas of dequantized algorithms and quantum variational algorithms represent effective 'QiML 2.0' software run on traditional computers. [02-04] In addition, quantum-inspired analogues will likely be further extended akin to Physics-Informed Machine Learning to 'assist machine learning tasks, representation of physical prior, and methods for incorporating physical prior.' [05] Recent ground breaking literature shown below has elevated tensor networks from efficient research tools to now increasing the pace of AI across disciplines. A) Tensor network/neural network hybrid performed better than standalone tensor networks or neural networks by NSF, MIT, and Harvard researchers. [06] B) More explainable and controllable compression of a Generative AI LLM to a fraction of its size by Multiverse Computing. [07] C) Researchers outperformed a leading quantum computer experiment in speed, precision, and accuracy - with scaling now corresponding to an infinite number of quantum bits on traditional hardware by Flatiron Institute, NYU. [08] Leading software platforms ITensor on C++ and Julia, and TeNPy on Python have been featured in a number of 2024 papers, and both maintain discussion forums to assist with tensor network developments. [09-12] In summary, High dimensional data in AI can now be distributed across tensor networks in more informed ways due to recent literature advancements and software library improvements. References [01] Tensor networks: https://lnkd.in/gipfeK_q [02] Ewin Tang: https://lnkd.in/gfgNSfKY [03] VQA: https://lnkd.in/gitb6TSq [04] QiML survey: https://lnkd.in/g97vr3_r [05] Physics-Informed ML: https://lnkd.in/gffmUFSx [06] NSF, MIT, Harvard: https://lnkd.in/gNkXEUtW [07] Multiverse: https://lnkd.in/gjNsqWJu [08] Flatiron, NYU: https://lnkd.in/gZmyJckE   [09] ITensor: https://lnkd.in/gXJWFNCU  [10] TeNPy: https://lnkd.in/g3Ciyyxs [11] ITensor: https://lnkd.in/gwhAp4BE [12] TeNPy: https://lnkd.in/gU5ceMFd

Interesting research in Quantum Machine Learning addresses key challenges in scalability and data encoding. The GitHub repository is included for further reference. A recent study titled "An Efficient Quantum Classifier Based on Hamiltonian Representations" (Tiblias et al.) proposes a novel approach to quantum classification. The study tackles the limitations of current QML methods that often rely on toy datasets or significant feature reduction due to hardware constraints and the high costs of encoding dense vector representations on quantum devices. The researchers introduce an efficient approach called the Hamiltonian classifier, which circumvents the costs of data encoding by mapping inputs to a finite set of Pauli strings and making predictions based on their expectation values. They also present two classifier variants, PEFF and SIM, with different trade-offs in terms of parameters and sample complexity. Key outcomes of this work include: * A new encoding scheme achieving logarithmic complexity in both qubits and quantum gates relative to the input dimensionality. * The development of classifier variants (PEFF and SIM) offers different performance-cost trade-offs. PEFF reduces model size, while SIM boasts better sample complexity. * The Simplified Hamiltonian (SIM) variant achieves logarithmic scaling in qubit and gate complexity along with a constant sample complexity, making it a strong candidate for practical implementation on Noisy Intermediate-Scale Quantum (NISQ) devices. * Experiments showed that increasing the number of Pauli strings in the SIM model leads to better performance and more stable training dynamics, with models using 500 to 1000 Pauli strings often matching the performance of classical baselines. You can find the GitHub repo here: https://lnkd.in/dN38CFPv. The article here: https://lnkd.in/dG4agXap #quantumcomputing #machinelearning #quantummachinelearning #artificialintelligence #research #nlp #imageclassification #datascience

What if the best use of a quantum computer isn't solving the problem, but teaching a classical AI how to? Genius! A group of researchers ran a real quantum computer (up to 115 qubits) to approximate ground states of a 2D magnetic material model, then measured properties like spin correlations. Instead of running the quantum computer for every new scenario, they trained a neural network on this quantum-generated data. The network then accurately predicted material properties for new, unseen conditions, showing quantum computers can generate useful training data for classical AI at real-world scale. Full white paper: https://lnkd.in/expXyBaJ

Exciting work from Caltech, Google Quantum AI, MIT, and Oratomic on quantum advantage for classical machine learning. The long standing question: can quantum computers offer a rigorous advantage in large scale classical data processing, not just specialized problems like cryptography or quantum simulation? This paper gives rigorous results for formalized machine learning tasks. In the benchmarks they report, a quantum computer with fewer than 60 logical qubits performs classification and dimension reduction on massive datasets using 4 to 6 orders of magnitude less memory than the classical and QRAM based baselines in the paper. The key idea is quantum oracle sketching. Instead of loading an entire dataset into quantum memory, it streams classical samples one at a time, applies small quantum rotations, and discards each sample immediately. These operations coherently build an approximate quantum oracle that can then be used in downstream quantum algorithms. The authors present numerical experiments on IMDb sentiment analysis and single cell RNA sequencing that are consistent with the theory. What makes this notable: - A provable quantum memory advantage for classification and dimension reduction - The advantage is framed as a theorem under the paper's learning model, not just a conjecture or empirical trend - The approach is designed to work with streaming, noisy, and time varying classical data Read the paper here: https://lnkd.in/g77PuZzQ

🚨 𝗤𝘂𝗮𝗻𝘁𝘂𝗺 𝗠𝗟 𝗷𝘂𝘀𝘁 𝗴𝗼𝘁 𝘃𝗲𝗿𝘆 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴... A new preprint from Google Quantum AI + Caltech + MIT + Oratomic drops a bold claim: 👉 A quantum computer with < 60 logical qubits could outperform any classical machine — even those with exponentially more memory — on real ML tasks. Yes, real ones: 🧠 IMDb sentiment analysis 🧬 Single-cell RNA classification Let that sink in. 💡 The breakthrough? Something called quantum oracle sketching Instead of trying to shove an entire dataset into fragile quantum memory (the usual bottleneck 😵💫), this approach does something smarter: ✨ It streams data one sample at a time ✨ Each data point nudges the quantum state with a tiny rotation ✨ Those microscopic updates accumulate into a full dataset representation No massive memory. No full data loading. Just… elegant physics. ⚡ Bonus: It avoids a major pain point in quantum ML Because the circuit is built directly from data (not trained via gradients), it sidesteps the dreaded: 🕳️ Barren plateau problem — where optimization just… dies.

If you've been doubting whether quantum computers will ever do anything useful beyond breaking encryption, this one's for you. A quantum computer with fewer than 60 logical qubits can run AI on massive real-world datasets using ten thousand to a million times less memory than any classical machine. Movie review sentiment analysis. Cell type classification from RNA sequencing. Real AI tasks, real data. This is not a storage trick. The quantum computer runs the full ML pipeline. An algorithm called quantum oracle sketching streams data through the processor one sample at a time. Each sample applies a small quantum rotation, then gets discarded. The accumulated rotations build a compressed quantum model of the entire dataset in a handful of qubits. Quantum algorithms then run classification and dimensionality reduction directly on that model. A readout protocol extracts the results. Data in, model built, inference done, predictions out. All on a tiny quantum chip. A classical machine matching this provably needs exponentially more memory, and that proof is unconditional. It relies only on quantum superposition being real. It holds even if you give classical machines unlimited time. Think about what this means for the age of AI. The world generates more data every day than it can store. Every sensor, every device, every interaction. Classical AI has to choose: store less and learn worse, or build bigger data centers and burn more energy. A quantum ML pipeline that learns from streaming data without storing it sidesteps that tradeoff entirely. But to be clear: This is a theoretical proof validated through numerical simulations. It has not been demonstrated on actual quantum hardware. Yet, fewer than 60 logical qubits is in the range that near-term error-corrected machines are targeting. We are finally getting the use-case evidence this field needed. 📸 Credits: Haimeng Zhao, Caltech Alexander Zlokapa Hsin-Yuan (Robert) Huang John Preskill Ryan Babbush Jarrod McClean Hartmut Neven Paper on arXiv:2604.07639 Deep dive on this live on X (@drmichaela_e). Newsletter version at 5pm CET today, link on my website.