arXiv subjects

Yao Fu

Publications and source records attributed to Yao Fu.

At least 37 records · Page 2Linked to original sources

arXiv2025

Room temperature optical control of spin states in organic diradicals

We report a family of luminescent alternant diradicals which, at room temperature, support a ground-state spin-triplet, near-unity photoluminescence quantum yields, and optical spin addressability. These diradicals comprise trityl groups meta-linked via pyridyl or phenyl groups, enabling optically bright triplet-to-triplet and singlet-to-singlet transitions. At room temperature, we observe optically detected magnetic resonance in these systems at zero magnetic field and a strong magneto-photoluminescence (10% modulation at 2 mT applied magnetic field). Distinct photoluminescence bands (at 630 nm and 700 nm) show opposite-sign spin-optical responses, arising from spin-selective intersystem crossing between triplet and singlet manifolds. These bright, all-organic diradicals offer a new set of chemically tunable materials for room temperature spin-optical interfaces, paving the way for application as quantum sensors.

cond-mat.mtrl-sci↗

arXiv2025

Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs

Neural network pruning has emerged as a promising approach for deploying LLMs in low-resource scenarios while preserving downstream task performance. However, for the first time, we reveal that such pruning disrupts LLMs' internal activation features crucial for lie detection, where probing classifiers (typically small logistic regression models) trained on these features assess the truthfulness of LLM-generated statements. This discovery raises a crucial open question: how can we prune LLMs without sacrificing these critical lie detection capabilities? Our investigation further reveals that naively adjusting layer-wise pruning sparsity based on importance inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. To address this issue, we propose Truthful Pruning aligned by Layer-wise Outliers (TPLO), which places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. This preserves LLMs' original performance while retaining critical features of inner states needed for robust lie detection. Moreover, we introduce a prompting rule to enrich the TruthfulQA benchmark for better calibrating LLM pruning. Empirical results show that our approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA.

cs.LG↗

arXiv2025

Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs

Quantization enables efficient deployment of large language models (LLMs) in resource-constrained environments by significantly reducing memory and computation costs. While quantized LLMs often maintain performance on perplexity and zero-shot tasks, their impact on truthfulness-whether generating truthful or deceptive responses-remains largely unexplored. In this work, we introduce TruthfulnessEval, a comprehensive evaluation framework for assessing the truthfulness of quantized LLMs across three dimensions: (1) Truthfulness on Logical Reasoning; (2) Truthfulness on Common Sense; and (3) Truthfulness on Imitative Falsehoods. Using this framework, we examine mainstream quantization techniques (ranging from 4-bit to extreme 2-bit) across several open-source LLMs. Surprisingly, we find that while quantized models retain internally truthful representations, they are more susceptible to producing false outputs under misleading prompts. To probe this vulnerability, we test 15 rephrased variants of "honest", "neutral" and "deceptive" prompts and observe that "deceptive" prompts can override truth-consistent behavior, whereas "honest" and "neutral" prompts maintain stable outputs. Further, we reveal that quantized models "know" the truth internally yet still produce false outputs when guided by "deceptive" prompts via layer-wise probing and PCA visualizations. Our findings provide insights into future designs of quantization-aware alignment and truthfulness interventions.

cs.AI↗

arXiv2025

FAEDKV: Infinite-Window Fourier Transform for Unbiased KV Cache Compression

The efficacy of Large Language Models (LLMs) in long-context tasks is often hampered by the substantial memory footprint and computational demands of the Key-Value (KV) cache. Current compression strategies, including token eviction and learned projections, frequently lead to biased representations -- either by overemphasizing recent/high-attention tokens or by repeatedly degrading information from earlier context -- and may require costly model retraining. We present FAEDKV (Frequency-Adaptive Infinite-Window for KV cache), a novel, training-free KV cache compression framework that ensures unbiased information retention. FAEDKV operates by transforming the KV cache into the frequency domain using a proposed Infinite-Window Fourier Transform (IWDFT). This approach allows for the equalized contribution of all tokens to the compressed representation, effectively preserving both early and recent contextual information. A preliminary frequency ablation study identifies critical spectral components for layer-wise, targeted compression. Experiments on LongBench benchmark demonstrate FAEDKV's superiority over existing methods by up to 22\%. In addition, our method shows superior, position-agnostic retrieval accuracy on the Needle-In-A-Haystack task compared to compression based approaches.

cs.CL↗

arXiv2025

MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently, but it depends on heterogeneous compute and memory resources. These factors jointly affect system Cost, Accuracy, and Performance (CAP), making trade-offs inevitable. Existing benchmarks often fail to capture these trade-offs accurately, complicating practical deployment decisions. To address this, we introduce MoE-CAP, a benchmark specifically designed for MoE systems. Our analysis reveals that achieving an optimal balance across CAP is difficult with current hardware; MoE systems typically optimize two of the three dimensions at the expense of the third-a dynamic we term the MoE-CAP trade-off. To visualize this, we propose the CAP Radar Diagram. We further introduce sparsity-aware performance metrics-Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU)-to enable accurate performance benchmarking of MoE systems across diverse hardware platforms and deployment scenarios.

cs.LG↗

arXiv2025

HybridServe: Efficient Serving of Large AI Models with Confidence-Based Cascade Routing

Giant Deep Neural Networks (DNNs), have become indispensable for accurate and robust support of large-scale cloud based AI services. However, serving giant DNNs is prohibitively expensive from an energy consumption viewpoint easily exceeding that of training, due to the enormous scale of GPU clusters needed to hold giant DNN model partitions and replicas. Existing approaches can either optimize energy efficiency or inference accuracy but not both. To overcome this status quo, we propose HybridServe, a novel hybrid DNN model serving system that leverages multiple sized versions (small to giant) of the model to be served in tandem. Through a confidence based hybrid model serving dataflow, HybridServe prefers to serve inference requests with energy-efficient smaller models so long as accuracy is not compromised, thereby reducing the number of replicas needed for giant DNNs. HybridServe also features a dataflow planner for efficient partitioning and replication of candidate models to maximize serving system throughput. Experimental results using a prototype implementation of HybridServe show that it reduces energy footprint by up to 19.8x compared to the state-of-the-art DNN model serving systems while matching the accuracy of serving solely with giant DNNs.

cs.LG↗

arXiv2025

MoE-Infinity: Efficient MoE Inference on Personal Machines with Sparsity-Aware Expert Cache

This paper presents MoE-Infinity, an efficient MoE inference system designed for personal machines with limited GPU memory capacity. The key idea for MoE-Infinity is that on personal machines, which are often single-user environments, MoE-based LLMs typically operate with a batch size of one. In this setting, MoE models exhibit a high degree of activation sparsity, meaning a small number of experts are frequently reused in generating tokens during the decode phase. Leveraging this idea, we design a sparsity-aware expert cache, which can trace the sparse activation of experts during inference and carefully select the trace that represents the sparsity pattern. By analyzing these selected traces, MoE-Infinity guides the replacement and prefetching of the expert cache, providing 3.1-16.7x per-token latency improvements over numerous state-of-the-art systems, including vLLM, Ollama, DeepSpeed and BrainStorm across various MoE models (DeepSeek and Mixtral) when handling different LLM tasks. MoE-Infinity's source code is publicly available at https://github.com/EfficientMoE/MoE-Infinity

cs.LG↗

arXiv2025

Hierarchical search method for gravitational waves from stellar-mass binary black holes in noisy space-based detector data

Future space-based laser interferometric detectors, such as LISA, will be able to detect gravitational waves (GWs) generated during the inspiral phase of stellar-mass binary black holes (SmBBHs). The detection and characterization of GWs from SmBBHs poses a formidable data analysis challenge, arising from the large number of wave cycles that make the search extremely sensitive to mismatches in signal and template parameters in a likelihood-based approach. This makes the search for the maximum of the likelihood function over the signal parameter space an extremely difficult task. We present a data analysis method that addresses this problem using both algorithmic innovations and hardware acceleration driven by GPUs. The method follows a hierarchical approach in which a semi-coherent $\mathcal{F}$-statistic is computed with different numbers of frequency domain partitions at different stages, with multiple particle swarm optimization (PSO) runs used in each stage for global optimization. An important step in the method is the judicious partitioning of the parameter space at each stage to improve the convergence probability of PSO and avoid premature convergence to noise-induced secondary maxima. The hierarchy of stages confines the semi-coherent searches to progressively smaller parameter ranges, with the final stage performing a search for the global maximum of the fully-coherent $\mathcal{F}$-statistic. We test our method on 2.5 years of a single LISA TDI combination and find that for an injected SmBBH signal with a SNR between $\approx 11$ and $\approx 14$, the method can estimate (i) the chirp mass with a relative error of $\lesssim 0.01\%$, (ii) the time of coalescence within $\approx 100$ sec, (iii) the sky location within $\approx 0.2$ ${\rm deg}^2$, and (iv) orbital eccentricity at a fiducial signal frequency of 10 mHz with a relative error of $\lesssim 1\%$. (abr.)

astro-ph.IM↗

arXiv2025

Multi-field quantum conferencing overcomes the network capacity limit

Quantum conferencing enables multiple nodes within a quantum network to share a secure group key for private message broadcasting. The key rate, however, is limited by the repeaterless capacity to distribute multiparticle entangled states across the network. Currently, in the finite-size regime, no feasible schemes utilizing existing experimental techniques can overcome the fundamental rate-distance limit of quantum conferencing in quantum networks without repeaters. Here, we propose a practical, multi-field scheme that breaks this limit, involving virtually establishing Greenberger-Horne-Zeilinger states through post-measurement coincidence matching. This proposal features a measurement-device-independent characteristic and can directly scale to support any number of users. Simulations show that the fundamental limitation on the group key rate can be overcome in a reasonable running time of sending $10^{14}$ pulses. We predict that it offers an efficient design for long-distance broadcast communication in future quantum networks.

quant-ph↗

arXiv2024

Security and Privacy of Digital Twins for Advanced Manufacturing: A Survey

In Industry 4.0, the digital twin is one of the emerging technologies, offering simulation abilities to predict, refine, and interpret conditions and operations, where it is crucial to emphasize a heightened concentration on the associated security and privacy risks. To be more specific, the adoption of digital twins in the manufacturing industry relies on integrating technologies like cyber-physical systems, the Industrial Internet of Things, virtualization, and advanced manufacturing. The interactions of these technologies give rise to numerous security and privacy vulnerabilities that remain inadequately explored. Towards that end, this paper analyzes the cybersecurity threats of digital twins for advanced manufacturing in the context of data collection, data sharing, machine learning and deep learning, and system-level security and privacy. We also provide several solutions to the threats in those four categories that can help establish more trust in digital twins.

eess.SY↗

arXiv2024

AutoGuide: Automated Generation and Selection of Context-Aware Guidelines for Large Language Model Agents

Recent advances in large language models (LLMs) have empowered AI agents capable of performing various sequential decision-making tasks. However, effectively guiding LLMs to perform well in unfamiliar domains like web navigation, where they lack sufficient knowledge, has proven to be difficult with the demonstration-based in-context learning paradigm. In this paper, we introduce a novel framework, called AutoGuide, which addresses this limitation by automatically generating context-aware guidelines from offline experiences. Importantly, each context-aware guideline is expressed in concise natural language and follows a conditional structure, clearly describing the context where it is applicable. As a result, our guidelines facilitate the provision of relevant knowledge for the agent's current decision-making process, overcoming the limitations of the conventional demonstration-based learning paradigm. Our evaluation demonstrates that AutoGuide significantly outperforms competitive baselines in complex benchmark domains, including real-world web navigation.

cs.CL↗

arXiv2024

Dynamic Self-Distillation via Previous Mini-batches for Fine-tuning Small Language Models

Knowledge distillation (KD) has become a widely adopted approach for compressing large language models (LLMs) to reduce computational costs and memory footprints. However, the availability of complex teacher models is a prerequisite for running most KD pipelines. Thus, the traditional KD procedure can be unachievable or budget-unfriendly, particularly when relying on commercial LLMs like GPT4. In this regard, Self-distillation (SelfD) emerges as an advisable alternative, enabling student models to learn without teachers' guidance. Nonetheless, existing SelfD approaches for LMs often involve architectural modifications, assuming the models are open-source, which may not always be practical. In this work, we introduce a model-agnostic and task-agnostic method named dynamic SelfD from the previous minibatch (DynSDPB), which realizes current iterations' distillation from the last ones' generated logits. Additionally, to address prediction inaccuracies during the early iterations, we dynamically adjust the distillation influence and temperature values to enhance the adaptability of fine-tuning. Furthermore, DynSDPB is a novel fine-tuning policy that facilitates the seamless integration of existing self-correction and self-training techniques for small language models (SLMs) because they all require updating SLMs' parameters. We demonstrate the superior performance of DynSDPB on both encoder-only LMs (e.g., BERT model families) and decoder-only LMs (e.g., LLaMA model families), validating its effectiveness across natural language understanding (NLU) and natural language generation (NLG) benchmarks.

cs.CL↗

arXiv2024

Interactive and Expressive Code-Augmented Planning with Large Language Models

Large Language Models (LLMs) demonstrate strong abilities in common-sense reasoning and interactive decision-making, but often struggle with complex, long-horizon planning tasks. Recent techniques have sought to structure LLM outputs using control flow and other code-adjacent techniques to improve planning performance. These techniques include using variables (to track important information) and functions (to divide complex tasks into smaller re-usable sub-tasks). However, purely code-based approaches can be error-prone and insufficient for handling ambiguous or unstructured data. To address these challenges, we propose REPL-Plan, an LLM planning approach that is fully code-expressive (it can utilize all the benefits of code) while also being dynamic (it can flexibly adapt from errors and use the LLM for fuzzy situations). In REPL-Plan, an LLM solves tasks by interacting with a Read-Eval-Print Loop (REPL), which iteratively executes and evaluates code, similar to language shells or interactive code notebooks, allowing the model to flexibly correct errors and handle tasks dynamically. We demonstrate that REPL-Plan achieves strong results across various planning domains compared to previous methods.

cs.CL↗

arXiv2024

DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

cs.CL↗

arXiv2024

ServerlessLLM: Low-Latency Serverless Inference for Large Language Models

This paper presents ServerlessLLM, a distributed system designed to support low-latency serverless inference for Large Language Models (LLMs). By harnessing the substantial near-GPU storage and memory capacities of inference servers, ServerlessLLM achieves effective local checkpoint storage, minimizing the need for remote checkpoint downloads and ensuring efficient checkpoint loading. The design of ServerlessLLM features three core contributions: (i) \emph{fast multi-tier checkpoint loading}, featuring a new loading-optimized checkpoint format and a multi-tier loading system, fully utilizing the bandwidth of complex storage hierarchies on GPU servers; (ii) \emph{efficient live migration of LLM inference}, which enables newly initiated inferences to capitalize on local checkpoint storage while ensuring minimal user interruption; and (iii) \emph{startup-time-optimized model scheduling}, which assesses the locality statuses of checkpoints on each server and schedules the model onto servers that minimize the time to start the inference. Comprehensive evaluations, including microbenchmarks and real-world scenarios, demonstrate that ServerlessLLM dramatically outperforms state-of-the-art serverless systems, reducing latency by 10 - 200X across various LLM inference workloads.

cs.LG↗

arXiv2024

Source-independent quantum secret sharing with entangled photon pair networks

The large-scale deployment of quantum secret sharing (QSS) in quantum networks is currently challenging due to the requirements for the generation and distribution of multipartite entanglement states. Here we present an efficient source-independent QSS protocol utilizing entangled photon pairs in quantum networks. Through the post-matching method, which means the measurement events in the same basis are matched, the key rate is almost independent of the number of participants. In addition, the unconditional security of our QSS against internal and external eavesdroppers can be proved by introducing an equivalent virtual protocol. Our protocol has great performance and technical advantages in future quantum networks.

quant-ph↗

arXiv2024

Efficient source-independent quantum conference key agreement

Quantum conference key agreement (QCKA) enables the unconditional secure distribution of conference keys among multiple participants. Due to challenges in high-fidelity preparation and long-distance distribution of multi-photon entanglement, entanglement-based QCKA is facing severe limitations in both key rate and scalability. Here, we propose a source-independent QCKA scheme utilizing the post-matching method, feasible within the entangled photon pair distribution network. We introduce an equivalent distributing virtual multi-photon entanglement protocol for providing the unconditional security proof even in the case of coherent attacks. For the symmetry star-network, comparing with previous $n$-photon entanglement protocol, the conference key rate is improved from $O(η^{n})$ to $O(η^{2})$, where $η$ is the transmittance from the entanglement source to one participant. Simulation results show that the performance of our protocol has multiple orders of magnitude advantages in the intercity distance. We anticipate that our approach will demonstrate its potential in the implementation of quantum networks.

quant-ph↗

arXiv2024

Optical read and write of spin states in organic diradicals

Optical control and read-out of the ground state spin structure has been demonstrated for defect states in crystalline semiconductors, including the diamond NV- center, and these are promising systems for quantum technologies. Molecular organic semiconductors offer synthetic control of spin placement, in contrast to current limitations in these crystalline systems. Here we report the discovery of spin-optical addressability in a diradical molecule that comprises two trityl radical groups coupled via a fluorene bridge. We demonstrate the three important properties that enable operation as a spin-photon interface: (i) triplet and singlet spin states show photoluminescence peaked at 640 and 700 nm respectively; this allows easy optical measurement of ground state spin. (ii) the ground state spin exchange is small (~60 μeV) that allows preparation of ground state spin population. This can be achieved by spin-selective excited state intersystem crossing, and we report up to 8% microwave-driven contrast in photoluminescence. (iii) both singlet and triplet manifolds have near-unity photoluminescence quantum yield, which is in contrast to the near-zero quantum yields in prior reports of molecular diradicals. Our results establish these tuneable open-shell organic molecules as a platform to engineer tailor-made spin-optical interfaces.

cond-mat.mtrl-sci↗