📚 AI论文速递 2026-03-08

📄 To scale the solution of optimization and simulation problems, prior work has explored machine-learning surrogates that inexpensively map problem parameters to corresponding solutions. Commonly used approaches, including supervised and self-supervised learning with either soft or hard feasibility enforcement, face inherent challenges such as reliance on expensive, high-quality labels or difficult optimization landscapes. To address their trade-offs, we propose a novel framework that first collects "cheap" imperfect labels, then performs supervised pretraining, and finally refines the model through self-supervised learning to improve overall performance. Our theoretical analysis and merit-based criterion show that labeled data need only place the model within a basin of attraction, confirming that only modest numbers of inexact labels and training epochs are required. We empirically validate our simple three-stage strategy across challenging domains, including nonconvex constrained optimization, power-grid operation, and stiff dynamical systems, and show that it yields faster convergence; improved accuracy, feasibility, and optimality; and up to 59x reductions in total offline cost.

📄 为了扩展优化和模拟问题的解决方案，之前的工作探索了机器学习代理，这些代理能够以低成本将问题参数映射到相应的解决方案。常用方法包括监督学习和自监督学习，分别采用软或硬可行性约束，但面临固有挑战，例如依赖昂贵的高质量标签或困难的优化景观。为了解决这些权衡问题，我们提出了一个新的框架：首先收集"廉价"的不完美标签，然后进行监督预训练，最后通过自监督学习来完善模型以提高整体性能。我们的理论分析和基于优点的标准表明，标记数据只需要将模型放置在吸引域内，确认只需要少量不精确标签和训练轮次。我们通过具有挑战性的领域经验验证了我们简单的三阶段策略，包括非凸约束优化、电网运营和刚性动力系统，并表明它带来了更快的收敛；提高了准确性、可行性和最优性；并将总离线成本降低高达59倍。

📄 We introduce group surface codes, which are a natural generalization of the Z2 surface code, and equivalent to quantum double models of finite groups with specific boundary conditions. We show that group surface codes can be leveraged to perform non-Clifford gates in Z2 surface codes, thus enabling universal computation with well-established means of performing logical Clifford gates. Moreover, for suitably chosen groups, we demonstrate that arbitrary reversible classical gates can be implemented transversally in the group surface code. We present the logical operations in terms of a set of elementary logical operations, which include transversal logical gates, a means of transferring encoded information into and out of group surface codes, and preparation and readout. By composing these elementary operations, we implement a wide variety of logical gates and provide a unified perspective on recent constructions in the literature for sliding group surface codes and preparing magic states. We furthermore use tensor networks inspired by ZX-calculus to construct spacetime implementations of the elementary operations. This spacetime perspective also allows us to establish explicit correspondences with topological gauge theories. Our work extends recent efforts in performing universal quantum computation in topological orders without the braiding of anyons, and shows how certain group surface codes allow us to bypass the restrictions set by the Bravyi-König theorem, which limits the computational power of topological Pauli stabilizer models.

📄 我们引入了群表面码，这是Z2表面码的自然推广，等价于具有特定边界条件的有限群量子双模型。我们展示了群表面码可用于在Z2表面码中执行非Clifford门，从而通过执行逻辑Clifford门的成熟方法实现通用计算。此外，对于适当选择的群，我们证明任意可逆经典门可以在群表面码中横向实现。我们以一组基本逻辑操作的形式呈现逻辑操作，包括横向逻辑门、将编码信息传入和传出群表面码的方法，以及制备和读出。通过组合这些基本逻辑操作，我们实现了多种逻辑门，并为文献中关于滑动群表面码和制备魔术态的最新构造提供了统一视角。此外，我们使用受ZX演算启发的张量网络来构建基本操作的时空实现。这种时空视角还使我们能够建立与拓扑规范理论的明确对应关系。我们的工作扩展了最近在拓扑序中执行通用量子计算（无需任何子的编织）的努力，并展示了某些群表面码如何让我们绕过Bravyi-König定理的限制，该定理限制了拓扑Pauli稳定器模型的计算能力。

📄 We present POET-X, a novel approach for memory-efficient training of Large Language Models that leverages randomized numerical linear algebra and scales Orthogonal Rational Kernels. Our method reduces memory footprint by factorizing the backward pass through the use of adaptive stochastic gradient estimators, achieving a 4x reduction in memory usage while maintaining competitive model quality.

📄 我们提出了POET-X，这是一种利用随机数值线性代数并缩放正交有理核的大语言模型内存高效训练新方法。我们的方法通过使用自适应随机梯度估计器分解反向传播来减少内存占用，在保持模型质量的同时实现了4倍的内存使用减少。

📄 We study how to observe and control internal features in Vision-Language-Action (VLA) models. We propose methods to identify semantic features within VLA representations and use them for model editing and constraint enforcement. Our experiments on robotic manipulation tasks show that this approach enables targeted modifications to model behavior without full retraining.

📄 我们研究如何在视觉-语言-动作（VLA）模型中观察和控制内部特征。我们提出了识别VLA表示中的语义特征并将其用于模型编辑和约束 enforcement的方法。我们在机器人操作任务上的实验表明，这种方法可以在无需完全重新训练的情况下对模型行为进行针对性修改。

📄 We explore using CLIP (Contrastive Language-Image Pretraining) representations for speech recognition tasks. By adapting CLIP's vision encoder for audio spectrograms, we achieve significant improvements in low-resource speech recognition benchmarks compared to traditional acoustic models.

📄 我们探索将CLIP（对比语言-图像预训练）表示用于语音识别任务。通过将CLIP的视觉编码器适配到音频频谱图，我们在低资源语音识别基准测试中相比传统声学模型取得了显著改进。

📄 We present MMLU-Pro, an enhanced version of the Massive Multitask Language Understanding benchmark. MMLU-Pro includes more challenging questions, longer context, and requires more complex reasoning than the original MMLU. Our evaluation of 20+ language models shows that MMLU-Pro provides a more differentiated measure of model capabilities.

📄 我们提出了MMLU-Pro，这是大规模多任务语言理解基准的增强版本。MMLU-Pro包含比原始MMLU更具挑战性的问题、更长的上下文，需要更复杂的推理。我们对20多个语言模型的评估表明，MMLU-Pro提供了更差异化的模型能力衡量标准。

📄 We propose Q-DiT, a post-training quantization method specifically designed for Diffusion Transformers. Unlike previous methods that focus on CNN-based diffusion models, Q-DiT addresses the unique challenges of transformer architectures in diffusion models, achieving state-of-the-art results in both weight and activation quantization.

📄 我们提出了Q-DiT，这是一种专门为扩散变换器设计的训练后量化方法。与之前专注于基于CNN的扩散模型的方法不同，Q-DiT解决了扩散模型中变换器架构的独特挑战，在权重和激活量化方面都取得了最先进的结果。

📄 This survey provides a comprehensive overview of methods for efficient LLM inference. We cover techniques including quantization, pruning, knowledge distillation, architecture optimization, and system-level optimizations. We also discuss benchmark datasets and evaluation protocols for measuring inference efficiency.

📄 本综述全面概述了大语言模型高效推理的方法。我们涵盖了量化、剪枝、知识蒸馏、架构优化和系统级优化等技术。我们还讨论了用于衡量推理效率的基准数据集和评估协议。

📄 We propose a hierarchical perspective on offline reinforcement learning that goes beyond reward optimization. Our approach models the task as a hierarchy of subtasks, allowing for more sample-efficient learning and better transfer across tasks. Experiments show significant improvements over reward-based methods in sparse reward environments.

📄 我们提出了一个超越奖励优化的离线强化学习层级视角。我们的方法将任务建模为子任务层次结构，从而实现更高的样本效率学习和更好的任务间迁移。实验表明，在稀疏奖励环境中，相比基于奖励的方法有显著改进。