Search

Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark., Comment: 18pages

Self-consistency-based approaches, which involve repeatedly sampling multiple outputs and selecting the most consistent one as the final response, prove to be remarkably effective in improving the factual accuracy of large language models. Nonetheless, existing methods usually have strict constraints on the task format, largely limiting their applicability. In this paper, we present Integrative Decoding (ID), to unlock the potential of self-consistency in open-ended generation tasks. ID operates by constructing a set of inputs, each prepended with a previously sampled response, and then processes them concurrently, with the next token being selected by aggregating of all their corresponding predictions at each decoding step. In essence, this simple approach implicitly incorporates self-consistency in the decoding objective. Extensive evaluation shows that ID consistently enhances factuality over a wide range of language models, with substantial improvements on the TruthfulQA (+11.2%), Biographies (+15.4%) and LongFact (+8.5%) benchmarks. The performance gains amplify progressively as the number of sampled responses increases, indicating the potential of ID to scale up with repeated sampling.

Codesign, an integral part of computer architecture referring to the information interaction in hardware-software stack, is able to boost the algorithm mapping and execution in the computer hardware. This well applies to the noisy intermediate-scale quantum era, where quantum algorithms and quantum processors both need to be shaped to allow for advantages in experimental implementations. The state-of-the-art quantum adiabatic optimization algorithm faces challenges for scaling up, where the deteriorating optimization performance is not necessarily alleviated by increasing the circuit depth given the noise in the hardware. The counterdiabatic term can be introduced to accelerate the convergence, but decomposing the unitary operator corresponding to the counterdiabatic terms into one and two-qubit gates may add additional burden to the digital circuit depth. In this work, we focus on the counterdiabatic protocol with a codesigned approach to implement this algorithm on a photonic quantum processor. The tunable Mach-Zehnder interferometer mesh provides rich programmable parameters for local and global manipulation, making it able to perform arbitrary unitary evolutions. Accordingly, we directly implement the unitary operation associated to the counterdiabatic quantum optimization on our processor without prior digitization. Furthermore, we develop and implement an optimized counterdiabatic method by tackling the higher-order many-body interaction terms. Moreover, we benchmark the performance in the case of factorization, by comparing the final success probability and the convergence speed. In conclusion, we experimentally demonstrate the advantages of a codesigned mapping of counterdiabatic quantum dynamics for quantum computing on photonic platforms., Comment: 9 pages, 4 figures. Comments are welcome

In quantum mechanics, a long-standing question remains: How does a single photon traverse double slits? One intuitive picture suggests that the photon passes through only one slit, while its wavefunction splits into an ``empty" wave and a ``full" wave. However, the reality of this empty wave is yet to be verified. Here, we present a novel experimental configuration that combines quantum memory and weak measurement to investigate the nature of the empty wave. A single atomic excitation is probabilistically split between free space and a quantum memory, analogous to the two paths in a double-slit experiment. The quantum memory serves as a path detector, where single-photon Raman scattering is enhanced due to the presence of a stored spin wave, without collapsing the quantum state. This enhancement is recorded as classical information, and the spin wave stored in the quantum memory is retrieved twice, with an interference visibility of 79%. Unlike conventional weak measurement schemes, where weak values are detected during post-selection, our approach converts the weak value into classical information before interference takes place. Our results demonstrate the potential of quantum memory as a measurement device that preserves coherence while extracting partial information, offering new insights into quantum measurement.

The ability of Graph Neural Networks (GNNs) to capture long-range and global topology information is limited by the scope of conventional graph Laplacian, leading to unsatisfactory performance on some datasets, particularly on heterophilic graphs. To address this limitation, we propose a new class of parameterized Laplacian matrices, which provably offers more flexibility in controlling the diffusion distance between nodes than the conventional graph Laplacian, allowing long-range information to be adaptively captured through diffusion on graph. Specifically, we first prove that the diffusion distance and spectral distance on graph have an order-preserving relationship. With this result, we demonstrate that the parameterized Laplacian can accelerate the diffusion of long-range information, and the parameters in the Laplacian enable flexibility of the diffusion scopes. Based on the theoretical results, we propose topology-guided rewiring mechanism to capture helpful long-range neighborhood information for heterophilic graphs. With this mechanism and the new Laplacian, we propose two GNNs with flexible diffusion scopes: namely the Parameterized Diffusion based Graph Convolutional Networks (PD-GCN) and Graph Attention Networks (PD-GAT). Synthetic experiments reveal the high correlations between the parameters of the new Laplacian and the performance of parameterized GNNs under various graph homophily levels, which verifies that our new proposed GNNs indeed have the ability to adjust the parameters to adaptively capture the global information for different levels of heterophilic graphs. They also outperform the state-of-the-art (SOTA) models on 6 out of 7 real-world benchmark datasets, which further confirms their superiority., Comment: arXiv admin note: substantial text overlap with arXiv:2403.01475

Over the past decade, Graph Neural Networks (GNNs) have achieved great success on machine learning tasks with relational data. However, recent studies have found that heterophily can cause significant performance degradation of GNNs, especially on node-level tasks. Numerous heterophilic benchmark datasets have been put forward to validate the efficacy of heterophily-specific GNNs and various homophily metrics have been designed to help people recognize these malignant datasets. Nevertheless, there still exist multiple pitfalls that severely hinder the proper evaluation of new models and metrics. In this paper, we point out three most serious pitfalls: 1) a lack of hyperparameter tuning; 2) insufficient model evaluation on the real challenging heterophilic datasets; 3) missing quantitative evaluation benchmark for homophily metrics on synthetic graphs. To overcome these challenges, we first train and fine-tune baseline models on $27$ most widely used benchmark datasets, categorize them into three distinct groups: malignant, benign and ambiguous heterophilic datasets, and identify the real challenging subsets of tasks. To our best knowledge, we are the first to propose such taxonomy. Then, we re-evaluate $10$ heterophily-specific state-of-the-arts (SOTA) GNNs with fine-tuned hyperparameters on different groups of heterophilic datasets. Based on the model performance, we reassess their effectiveness on addressing heterophily challenge. At last, we evaluate $11$ popular homophily metrics on synthetic graphs with three different generation approaches. To compare the metrics strictly, we propose the first quantitative evaluation method based on Fr\'echet distance., Comment: arXiv admin note: substantial text overlap with arXiv:2407.09618

Large Language Models (LLMs) exhibit remarkably powerful capabilities. One of the crucial factors to achieve success is aligning the LLM's output with human preferences. This alignment process often requires only a small amount of data to efficiently enhance the LLM's performance. While effective, research in this area spans multiple domains, and the methods involved are relatively complex to understand. The relationships between different methods have been under-explored, limiting the development of the preference alignment. In light of this, we break down the existing popular alignment strategies into different components and provide a unified framework to study the current alignment strategies, thereby establishing connections among them. In this survey, we decompose all the strategies in preference learning into four components: model, data, feedback, and algorithm. This unified view offers an in-depth understanding of existing alignment algorithms and also opens up possibilities to synergize the strengths of different strategies. Furthermore, we present detailed working examples of prevalent existing algorithms to facilitate a comprehensive understanding for the readers. Finally, based on our unified perspective, we explore the challenges and future research directions for aligning large language models with human preferences., Comment: 23 pages, 6 figures

Conventionally, atomic vapor is perceived as a non-living system governed by the principles of thermodynamics and statistical physics. However, the demarcation line between life and non-life appears to be less distinct than previously thought. In a study of amplifying spin waves, we observe a phenomenon reminiscent of life: The atomic spin wave stored in atomic vapor has a capability of absorbing energy from an external light source, and exhibits behaviors akin to active regeneration. We demonstrate that this regeneration significantly enhances the lifetime and retrieval efficiency of the spin wave, while concurrently the noise is effectively suppressed. Our results suggest that the regeneration mechanism holds promise for mitigating the pronounced decoherence typically encountered in spin waves carried by room-temperature media, therefore offering potential applications in the realms of quantum information and precision measurements at ambient conditions.

With the emergence of super-resolution lenses such as superlens and hyperlens, coupled with advancements in metamaterials, the diffraction limit of approximately half wavelength is no longer unbreakable. However, superlenses are easily affected by weak intrinsic losses and hyperlenses cannot achieve perfect imaging, significantly constraining their practical utility. To address these challenges, here we propose a perfect hyperlens based on the metric of de Sitter spacetime in cosmology. Importantly, perfect hyperlens is capable of self-focusing in geometrical optics while supporting propagating waves with exceptionally large wavenumbers, which endows it with key advantages such as ultra-high resolution, no aberration and strong robustness. Furthermore, we demonstrate the hyperbolic focusing performance and mimic the de Sitter spacetime in naturally in-plane hyperbolic polaritons of {\alpha}-MoO3 films numerically, which can be achieved with a gradient thickness profile. Our work provides cosmological insights into the field regulation in hyperbolic materials, greatly innovates the design principles of traditional imaging lenses., Comment: 22 pages, 8 figures

Homophily principle, \ie{} nodes with the same labels or similar attributes are more likely to be connected, has been commonly believed to be the main reason for the superiority of Graph Neural Networks (GNNs) over traditional Neural Networks (NNs) on graph-structured data, especially on node-level tasks. However, recent work has identified a non-trivial set of datasets where GNN's performance compared to the NN's is not satisfactory. Heterophily, i.e. low homophily, has been considered the main cause of this empirical observation. People have begun to revisit and re-evaluate most existing graph models, including graph transformer and its variants, in the heterophily scenario across various kinds of graphs, e.g. heterogeneous graphs, temporal graphs and hypergraphs. Moreover, numerous graph-related applications are found to be closely related to the heterophily problem. In the past few years, considerable effort has been devoted to studying and addressing the heterophily issue. In this survey, we provide a comprehensive review of the latest progress on heterophilic graph learning, including an extensive summary of benchmark datasets and evaluation of homophily metrics on synthetic graphs, meticulous classification of the most updated supervised and unsupervised learning methods, thorough digestion of the theoretical analysis on homophily/heterophily, and broad exploration of the heterophily-related applications. Notably, through detailed experiments, we are the first to categorize benchmark heterophilic datasets into three sub-categories: malignant, benign and ambiguous heterophily. Malignant and ambiguous datasets are identified as the real challenging datasets to test the effectiveness of new models on the heterophily challenge. Finally, we propose several challenges and future directions for heterophilic graph representation learning., Comment: Suggestions and comments are welcomed at sitao.luan@mail.mcgill.ca!

In recent progress, mathematical verifiers have achieved success in mathematical reasoning tasks by validating the correctness of solutions generated by policy models. However, existing verifiers are trained with binary classification labels, which are not informative enough for the model to accurately assess the solutions. To mitigate the aforementioned insufficiency of binary labels, we introduce step-wise natural language feedback as rationale labels, that is, the correctness of each step and the detailed explanations. In this paper, we propose Math-Minos, a natural language feedback-enhanced verifier by constructing automatically generated training data and a two-stage training paradigm for effective training and efficient inference. Our experiments reveal that a small set of natural language feedback can significantly boost the performance of the verifier in both verification and reinforcement learning. We have released the code and data for further exploration., Comment: 15 pages

Large language models (LLMs), including both proprietary and open-source models, have showcased remarkable capabilities in addressing a wide range of downstream tasks. Nonetheless, when it comes to practical Chinese legal tasks, these models fail to meet the actual requirements. Proprietary models do not ensure data privacy for sensitive legal cases, while open-source models demonstrate unsatisfactory performance due to their lack of legal knowledge. To address this problem, we introduce LawGPT, the first open-source model specifically designed for Chinese legal applications. LawGPT comprises two key components: legal-oriented pre-training and legal supervised fine-tuning. Specifically, we employ large-scale Chinese legal documents for legal-oriented pre-training to incorporate legal domain knowledge. To further improve the model's performance on downstream legal tasks, we create a knowledge-driven instruction dataset for legal supervised fine-tuning. Our experimental results demonstrate that LawGPT outperforms the open-source LLaMA 7B model. Our code and resources are publicly available at https://github.com/pengxiao-song/LaWGPT and have received 5.7K stars on GitHub., Comment: Technical Report

Room-temperature (RT) two-dimensional (2D) van der Waals (vdW) ferromagnets hold immense promise for next-generation spintronic devices for information storage and processing. To achieve high-density energy-efficient spintronic devices, it is essential to understand local magnetic properties of RT 2D vdW magnets. In this work, we realize noninvasive in situ magnetic detection in vdW-layered ferromagnet Fe3GaTe2 using divacancy spins quantum sensor in silicon carbide (SiC) at RT. The structural features and magnetic properties of the Fe3GaTe2 are characterized utilizing Raman spectrum, magnetization and magneto-transport measurements. Further detailed analysis of temperature- and magnetic field-dependent optically detected magnetic resonances of the PL6 divacancy near the Fe3GaTe2 reveal that, the Curie temperature (Tc) of Fe3GaTe2 is ~360K, and the magnetization increases with external magnetic fields. Additionally, spin relaxometry technology is employed to probe the magnetic fluctuations of Fe3GaTe2, revealing a peak in the spin relaxation rate around Tc. These experiments give insights into the intriguing local magnetic properties of 2D vdW RT ferromagnet Fe3GaTe2 and pave the way for the application of SiC quantum sensors in noninvasive in situ magnetic detection of related 2D vdW magnets., Comment: 15 pages, 4 figures

In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100% Acc. performance, matching that of a full KV cache.

Recent studies reveal that Large Language Models (LLMs) face challenges in balancing safety with utility, particularly when processing long texts for NLP tasks like summarization and translation. Despite defenses against malicious short questions, the ability of LLMs to safely handle dangerous long content, such as manuals teaching illicit activities, remains unclear. Our work aims to develop robust defenses for LLMs in processing malicious documents alongside benign NLP task queries. We introduce a defense dataset comprised of safety-related examples and propose single-task and mixed-task losses for instruction tuning. Our empirical results demonstrate that LLMs can significantly enhance their capacity to safely manage dangerous content with appropriate instruction tuning. Additionally, strengthening the defenses of tasks most susceptible to misuse is effective in protecting LLMs against processing harmful information. We also observe that trade-offs between utility and safety exist in defense strategies, where Llama2, utilizing our proposed approach, displays a significantly better balance compared to Llama1., Comment: accepted to NAACL2024 TrustNLP workshop

Massive MIMO (multiple-input multiple-output) detection is an important topic in wireless communication and various machine learning based methods have been developed recently for this task. Expectation Propagation (EP) and its variants are widely used for MIMO detection and have achieved the best performance. However, EP-based solvers fail to capture the correlation between unknown variables, leading to a loss of information, and in addition, they are computationally expensive. In this paper, we show that the real-valued system can be modeled as spectral signal convolution on graph, through which the correlation between unknown variables can be captured. Based on such analysis, we propose graph convolution-enhanced expectation propagation (GCEPNet). GCEPNet incorporates data-dependent attention scores into Chebyshev polynomial for powerful graph convolution with better generalization capacity. It enables a better estimation of the cavity distribution for EP and empirically achieves the state-of-the-art (SOTA) MIMO detection performance with much faster inference speed. To our knowledge, we are the first to shed light on the connection between the system model and graph convolution, and the first to design the data-dependent coefficients for graph convolution., Comment: In IEEE GLOBECOM 2024 Conference Proceedings

Quantum communication networks are crucial for both secure communication and cryptographic networked tasks. Building quantum communication networks in a scalable and cost-effective way is essential for their widespread adoption, among which a stable and miniaturized high-quality quantum light source is a key component. Here, we establish a complete polarization entanglement-based fully connected network, which features an ultrabright integrated Bragg reflection waveguide quantum source, managed by an untrusted service provider, and a streamlined polarization analysis module, which requires only one single-photon detector for each end user. We perform a continuously working quantum entanglement distribution and create correlated bit strings between users. Within the framework of one-time universal hashing, we provide the first experimental implementation of source-independent quantum digital signatures using imperfect keys circumventing the necessity for private amplification. More importantly, we further beat the 1/3 fault-tolerance bound in Byzantine agreement, achieving unconditional security without relying on sophisticated techniques. Our results offer an affordable and practical route for addressing consensus challenges within the emerging quantum network landscape.

Graph Attention Network (GAT) is one of the most popular Graph Neural Network (GNN) architecture, which employs the attention mechanism to learn edge weights and has demonstrated promising performance in various applications. However, since it only incorporates information from immediate neighborhood, it lacks the ability to capture long-range and global graph information, leading to unsatisfactory performance on some datasets, particularly on heterophilic graphs. To address this limitation, we propose the Directional Graph Attention Network (DGAT) in this paper. DGAT is able to combine the feature-based attention with the global directional information extracted from the graph topology. To this end, a new class of Laplacian matrices is proposed which can provably reduce the diffusion distance between nodes. Based on the new Laplacian, topology-guided neighbour pruning and edge adding mechanisms are proposed to remove the noisy and capture the helpful long-range neighborhood information. Besides, a global directional attention is designed to enable a topological-aware information propagation. The superiority of the proposed DGAT over the baseline GAT has also been verified through experiments on real-world benchmarks and synthetic data sets. It also outperforms the state-of-the-art (SOTA) models on 6 out of 7 real-world benchmark datasets.

We consider the success probability of the $L_0$-regularized box-constrained Babai point, which is a suboptimal solution to the $L_0$-regularized box-constrained integer least squares problem and can be used for MIMO detection. First, we derive formulas for the success probability of both $L_0$-regularized and unregularized box-constrained Babai points. Then we investigate the properties of the $L_0$-regularized box-constrained Babai point, including the optimality of the regularization parameter, the monotonicity of its success probability, and the monotonicity of the ratio of the two success probabilities. A bound on the success probability of the $L_0$-regularized Babai point is derived. After that, we analyze the effect of the LLL-P permutation strategy on the success probability of the $L_0$-regularized Babai point. Then we propose some success probability based column permutation strategies to increase the success probability of the $L_0$-regularized box-constrained Babai point. Finally, we present numerical tests to confirm our theoretical results and to show the advantage of the $L_0$ regularization and the effectiveness of the proposed column permutation algorithms compared to existing strategies., Comment: 37 pages, 1 figure including 2 subfigures