The transition to the fault-tolerant era exposes the limitations of homogeneous quantum systems, where no single qubit modality simultaneously offers optimal operation speed, connectivity, and scalability. In this work, we propose a strategic approach to Heterogeneous Quantum Architectures (HQA) that synthesizes the distinct advantages of the superconducting (SC) and neutral atom (NA) platforms. We explore two architectural role assignment strategies based on hardware characteristics: (1) We offload the latency-critical Magic State Factory (MSF) to fast SC devices while performing computation on scalable NA arrays, a design we term MagicAcc, which effectively mitigates the resource-preparation bottleneck. (2) We explore a Memory-Compute Separation (MCSep) paradigm that utilizes NA arrays for high-density qLDPC memory storage and SC devices for fast surface-code processing. Our evaluation, based on a comprehensive end-to-end cost model, demonstrates that principled heterogeneity yields significant performance gains. Specifically, our designs achieve $752\times$ speedup over NA-only baselines on average and reduce the physical qubit footprint by over $10\times$ compared to SC-only systems. These results chart a clear pathway for leveraging cross-modality interconnects to optimize the space-time efficiency of future fault-tolerant quantum computers. <<<

Current approaches to reducing undesired capabilities in language models are largely post hoc, and can thus be easily bypassed by adversaries. A natural alternative is to shape capabilities during pretraining itself. On the proxy task of removing medical capabilities, we show that the simple intervention of filtering pretraining data is highly effective, robust, and inexpensive at scale. Inspired by work on data attribution, we show that filtering tokens is more effective than filtering documents, achieving the same hit to undesired capabilities at a lower cost to benign ones. Training models spanning two orders of magnitude, we then demonstrate that filtering gets more effective with scale: for our largest models, token filtering leads to a 7000x compute slowdown on the forget domain. We also show that models trained with token filtering can still be aligned on the forget domain. Along the way, we introduce a methodology for labeling tokens with sparse autoencoders and distilling cheap, high-quality classifiers. We also demonstrate that filtering can be robust to noisy labels with sufficient pretraining compute. <<<

1<br> Abstract<br> Interactions between proteins and other biomolecules underlie nearly all biological processes, yet designing such interactions de novo remains challenging. Capturing their specific interactions and co-optimizing sequence and structure are difficult and often require extensive computation. We present Protein Hunter, a fast, fine-tuning-free framework for de novo protein design. Starting from an all-X sequence, we find diffusion-based structure prediction models hallucinate reasonable looking structures that can be further improved through iterative sequence re-design and structure re-prediction. This lightweight strategy achieves high AlphaFold3 in silico success rates across both unconditional and conditional generation tasks, including binders to proteins, cyclic peptides, small molecules, DNA, and RNA. Protein Hunter also supports multi-motif scaffolding and partial redesign, providing a general and efficient platform for de novo protein design across diverse molecular targets. <<<

This white paper discusses and explores the various points of intersection<br>between quantum computing and artificial intelligence (AI). It describes how<br>quantum computing could support the development of innovative AI solutions. It<br>also examines use cases of classical AI that can empower research and<br>development in quantum technologies, with a focus on quantum computing and<br>quantum sensing. The purpose of this white paper is to provide a long-term<br>research agenda aimed at addressing foundational questions about how AI and<br>quantum computing interact and benefit one another. It concludes with a set of<br>recommendations and challenges, including how to orchestrate the proposed<br>theoretical work, align quantum AI developments with quantum hardware roadmaps,<br>estimate both classical and quantum resources - especially with the goal of<br>mitigating and optimizing energy consumption - advance this emerging hybrid<br>software engineering discipline, and enhance European industrial<br>competitiveness while considering societal implications. <<<

This paper introduces Genesis, the first compiler designed to support<br>Hamiltonian Simulation on hybrid continuous-variable (CV) and discrete-variable<br>(DV) quantum computing systems. Genesis is a two-level compilation system. At<br>the first level, it decomposes an input Hamiltonian into basis gates using the<br>native instruction set of the target hybrid CV-DV quantum computer. At the<br>second level, it tackles the mapping and routing of qumodes/qubits to implement<br>long-range interactions for the gates decomposed from the first level. Rather<br>than a typical implementation that relies on SWAP primitives similar to<br>qubit-based (or DV-only) systems, we propose an integrated design of<br>connectivity-aware gate synthesis and beamsplitter SWAP insertion tailored for<br>hybrid CV-DV systems. We also introduce an OpenQASM-like domain-specific<br>language (DSL) named CVDV-QASM to represent Hamiltonian in terms of<br>Pauli-exponentials and basic gate sequences from the hybrid CV-DV gate set.<br>Genesis has successfully compiled several important Hamiltonians, including the<br>Bose-Hubbard model, $\mathbb{Z}_2-$Higgs model, Hubbard-Holstein model,<br>Heisenberg model and Electron-vibration coupling Hamiltonians, which are<br>critical in domains like quantum field theory, condensed matter physics, and<br>quantum chemistry. Our implementation is available at<br>Genesis-CVDV-Compiler(https://github.com/ruadapt/Genesis-CVDV-Compiler). <<<

We construct $\varepsilon$-approximate unitary $k$-designs on $n$ qubits in<br>circuit depth $O(\log k \log \log n k / \varepsilon)$. The depth is<br>exponentially improved over all known results in all three parameters $n$, $k$,<br>$\varepsilon$. We further show that each dependence is optimal up to<br>exponentially smaller factors. Our construction uses $\tilde{{O}}(nk)$ ancilla<br>qubits and ${O}(nk)$ bits of randomness, which are also optimal up to $\log(n<br>k)$ factors. An alternative construction achieves a smaller ancilla count<br>$\tilde{{O}}(n)$ with circuit depth ${O}(k \log \log nk/\varepsilon)$. To<br>achieve these efficient unitary designs, we introduce a highly-structured<br>random unitary ensemble that leverages long-range two-qubit gates and low-depth<br>implementations of random classical hash functions. We also develop a new<br>analytical framework for bounding errors in quantum experiments involving many<br>queries to random unitaries. As an illustration of this framework's<br>versatility, we provide a succinct alternative proof of the existence of<br>pseudorandom unitaries. <<<

Artificial intelligence (AI) advancements over the past few years have had an<br>unprecedented and revolutionary impact across everyday application areas. Its<br>significance also extends to technical challenges within science and<br>engineering, including the nascent field of quantum computing (QC). The<br>counterintuitive nature and high-dimensional mathematics of QC make it a prime<br>candidate for AI's data-driven learning capabilities, and in fact, many of QC's<br>biggest scaling challenges may ultimately rest on developments in AI. However,<br>bringing leading techniques from AI to QC requires drawing on disparate<br>expertise from arguably two of the most advanced and esoteric areas of computer<br>science. Here we aim to encourage this cross-pollination by reviewing how<br>state-of-the-art AI techniques are already advancing challenges across the<br>hardware and software stack needed to develop useful QC - from device design to<br>applications. We then close by examining its future opportunities and obstacles<br>in this space. <<<

Many real-world applications require the prediction of long sequence<br>time-series, such as electricity consumption planning. Long sequence<br>time-series forecasting (LSTF) demands a high prediction capacity of the model,<br>which is the ability to capture precise long-range dependency coupling between<br>output and input efficiently. Recent studies have shown the potential of<br>Transformer to increase the prediction capacity. However, there are several<br>severe issues with Transformer that prevent it from being directly applicable<br>to LSTF, including quadratic time complexity, high memory usage, and inherent<br>limitation of the encoder-decoder architecture. To address these issues, we<br>design an efficient transformer-based model for LSTF, named Informer, with<br>three distinctive characteristics: (i) a $ProbSparse$ self-attention mechanism,<br>which achieves $O(L \log L)$ in time complexity and memory usage, and has<br>comparable performance on sequences' dependency alignment. (ii) the<br>self-attention distilling highlights dominating attention by halving cascading<br>layer input, and efficiently handles extreme long input sequences. (iii) the<br>generative style decoder, while conceptually simple, predicts the long<br>time-series sequences at one forward operation rather than a step-by-step way,<br>which drastically improves the inference speed of long-sequence predictions.<br>Extensive experiments on four large-scale datasets demonstrate that Informer<br>significantly outperforms existing methods and provides a new solution to the<br>LSTF problem. <<<

Generative AI technologies have been deployed in many places, such as<br>(multimodal) large language models and vision generative models. Their<br>remarkable performance should be attributed to massive training data and<br>emergent reasoning abilities. However, the models would memorize and generate<br>sensitive, biased, or dangerous information originated from the training data<br>especially those from web crawl. New machine unlearning (MU) techniques are<br>being developed to reduce or eliminate undesirable knowledge and its effects<br>from the models, because those that were designed for traditional<br>classification tasks could not be applied for Generative AI. We offer a<br>comprehensive survey on many things about MU in Generative AI, such as a new<br>problem formulation, evaluation methods, and a structured discussion on the<br>advantages and limitations of different kinds of MU techniques. It also<br>presents several critical challenges and promising directions in MU research. A<br>curated list of readings can be found:<br>https://github.com/franciscoliu/GenAI-MU-Reading. <<<

Image inversion is a fundamental task in generative models, aiming to map<br>images back to their latent representations to enable downstream applications<br>such as editing, restoration, and style transfer. This paper provides a<br>comprehensive review of the latest advancements in image inversion techniques,<br>focusing on two main paradigms: Generative Adversarial Network (GAN) inversion<br>and diffusion model inversion. We categorize these techniques based on their<br>optimization methods. For GAN inversion, we systematically classify existing<br>methods into encoder-based approaches, latent optimization approaches, and<br>hybrid approaches, analyzing their theoretical foundations, technical<br>innovations, and practical trade-offs. For diffusion model inversion, we<br>explore training-free strategies, fine-tuning methods, and the design of<br>additional trainable modules, highlighting their unique advantages and<br>limitations. Additionally, we discuss several popular downstream applications<br>and emerging applications beyond image tasks, identifying current challenges<br>and future research directions. By synthesizing the latest developments, this<br>paper aims to provide researchers and practitioners with a valuable reference<br>resource, promoting further advancements in the field of image inversion. We<br>keep track of the latest works at https://github.com/RyanChenYN/ImageInversion <<<

Denoising diffusion probabilistic models (DDPM) are powerful hierarchical<br>latent variable models with remarkable sample generation quality and training<br>stability. These properties can be attributed to parameter sharing in the<br>generative hierarchy, as well as a parameter-free diffusion-based inference<br>procedure. In this paper, we present Few-Shot Diffusion Models (FSDM), a<br>framework for few-shot generation leveraging conditional DDPMs. FSDMs are<br>trained to adapt the generative process conditioned on a small set of images<br>from a given class by aggregating image patch information using a set-based<br>Vision Transformer (ViT). At test time, the model is able to generate samples<br>from previously unseen classes conditioned on as few as 5 samples from that<br>class. We empirically show that FSDM can perform few-shot generation and<br>transfer to new datasets. We benchmark variants of our method on complex vision<br>datasets for few-shot learning and compare to unconditional and conditional<br>DDPM baselines. Additionally, we show how conditioning the model on patch-based<br>input set information improves training convergence. <<<

We present an approach for estimating the fraction of text in a large corpus<br>which is likely to be substantially modified or produced by a large language<br>model (LLM). Our maximum likelihood model leverages expert-written and<br>AI-generated reference texts to accurately and efficiently examine real-world<br>LLM-use at the corpus level. We apply this approach to a case study of<br>scientific peer review in AI conferences that took place after the release of<br>ChatGPT: ICLR 2024, NeurIPS 2023, CoRL 2023 and EMNLP 2023. Our results suggest<br>that between 6.5% and 16.9% of text submitted as peer reviews to these<br>conferences could have been substantially modified by LLMs, i.e. beyond<br>spell-checking or minor writing updates. The circumstances in which generated<br>text occurs offer insight into user behavior: the estimated fraction of<br>LLM-generated text is higher in reviews which report lower confidence, were<br>submitted close to the deadline, and from reviewers who are less likely to<br>respond to author rebuttals. We also observe corpus-level trends in generated<br>text which may be too subtle to detect at the individual level, and discuss the<br>implications of such trends on peer review. We call for future<br>interdisciplinary work to examine how LLM use is changing our information and<br>knowledge practices. <<<

PySCF is a general-purpose electronic structure platform designed from the<br>ground up to emphasize code simplicity, both to aid new method development, as<br>well as for flexibility in computational workflow. The package provides a wide<br>range of tools to support simulations of finite size systems, extended systems<br>with periodic boundary conditions, low dimensional periodic systems, and custom<br>Hamiltonians, using mean-field and post-mean-field methods with standard<br>Gaussian basis functions. To ensure easy of extensibility, PySCF uses the<br>Python language to implement almost all its features, while computationally<br>critical paths are implemented with heavily optimized C routines. Using this<br>combined Python/C implementation, the package is as efficient as the best<br>existing C or Fortran based quantum chemistry programs. In this paper we<br>document the capabilities and design philosophy of the current version of the<br>PySCF package. <<<

Recent advancements in quantum computing have opened new avenues for tackling<br>long-standing complex combinatorial optimization problems that are intractable<br>for classical computers. Predicting secondary structure of mRNA is one such<br>notoriously difficult problem that can benefit from the ever-increasing<br>maturity of quantum computing technology. Accurate prediction of mRNA secondary<br>structure is critical in designing RNA-based therapeutics as it dictates<br>various steps of an mRNA life cycle, including transcription, translation, and<br>decay. The current generation of quantum computers have reached utility-scale,<br>allowing us to explore relatively large problem sizes. In this paper, we<br>examine the feasibility of solving mRNA secondary structures on a quantum<br>computer with sequence length up to 60 nucleotides representing problems in the<br>qubit range of 10 to 80. We use Conditional Value at Risk (CVaR)-based VQE<br>algorithm to solve the optimization problems, originating from the mRNA<br>structure prediction problem, on the IBM Eagle and Heron quantum processors. To<br>our encouragement, even with ``minimal'' error mitigation and fixed-depth<br>circuits, our hardware runs yield accurate predictions of minimum free energy<br>(MFE) structures that match the results of the classical solver CPLEX. Our<br>results provide sufficient evidence for the viability of solving mRNA structure<br>prediction problems on a quantum computer and motivate continued research in<br>this direction. <<<