arXiv subjects

Lan Liu

Publications and source records attributed to Lan Liu.

At least 19 recordsLinked to original sources

arXiv2026

Expert Streaming: Accelerating Low-Batch MoE Inference via Multi-chiplet Architecture and Dynamic Expert Trajectory Scheduling

Mixture-of-Experts is a promising approach for edge AI with low-batch inference. Yet, on-device deployments often face limited on-chip memory and severe workload imbalance; the prevalent use of offloading further incurs off-chip memory access bottlenecks. Moreover, MoE sparsity and dynamic gating shift distributed strategies toward much finer granularity and introduce runtime scheduling considerations. Recently, high die-to-die bandwidth chiplet interconnects have created new opportunities for multi-chiplet systems to address workload imbalance and offloading bottlenecks with fine-grained scheduling. In this paper, we propose Fully Sharded Expert Data Parallelism, a parallelization paradigm specifically architected for low-batch MoE inference on multi-chiplet accelerators. FSE-DP attains adaptive computation-communication overlap and balanced load by orchestrating fine-grained, complementary expert streams along dynamic trajectories across high-bandwidth D2D links. The attendant dataflow complexity is tamed by a minimal, hardware-amenable set of virtualization rules and a lightweight scheduling algorithm. Our approach achieves 1.22 to 2.00 times speedup over state-of-the-art baselines and saves up to 78.8 percent on-chip memory.

cs.AR↗

arXiv2025

SingingSDS: A Singing-Capable Spoken Dialogue System for Conversational Roleplay Applications

With recent advances in automatic speech recognition (ASR), large language models (LLMs), and text-to-speech (TTS) technologies, spoken dialogue systems (SDS) have become widely accessible. However, most existing SDS are limited to conventional spoken responses. We present SingingSDS, a cascaded SDS that responds through singing rather than speaking, fostering more affective, memorable, and pleasurable interactions in character-based roleplay and interactive entertainment scenarios. SingingSDS employs a modular ASR-LLM-SVS pipeline and supports a wide range of configurations across character personas, ASR and LLM backends, SVS models, melody sources, and voice profiles, tailored to different needs in terms of latency, quality, and musical style. SingingSDS is available as a plug-and-play web demo, featuring modular, open-source code that supports customization and extension. Demo: https://huggingface.co/spaces/espnet/SingingSDS. Code: https://github.com/SingingSDS/SingingSDS.

cs.SD↗

arXiv2025

SingMOS-Pro: An Comprehensive Benchmark for Singing Quality Assessment

Singing voice generation progresses rapidly, yet evaluating singing quality remains a critical challenge. Human subjective assessment, typically in the form of listening tests, is costly and time consuming, while existing objective metrics capture only limited perceptual aspects. In this work, we introduce SingMOS-Pro, a dataset for automatic singing quality assessment. Building on our preview version SingMOS, which provides only overall ratings, SingMOS-Pro extends the annotations of the additional data to include lyrics, melody, and overall quality, offering broader coverage and greater diversity. The dataset contains 7,981 singing clips generated by 41 models across 12 datasets, spanning from early systems to recent state-of-the-art approaches. Each clip is rated by at least five experienced annotators to ensure reliability and consistency. Furthermore, we investigate strategies for effectively utilizing MOS data annotated under heterogeneous standards and benchmark several widely used evaluation methods from related tasks on SingMOS-Pro, establishing strong baselines and practical references for future research. The dataset is publicly available at https://huggingface.co/datasets/TangRain/SingMOS-Pro.

cs.SD↗

arXiv2025

Dynamic read & write optimization with TurtleKV

High read and write performance is important for generic key-value stores, which are foundational to modern applications and databases. Yet, achieving high performance for mixed and dynamic workloads is challenging due to fundamental trade-offs between memory use and I/O for retrieval and updates. Past work emphasizes the trade-off between read- and write-optimization as expressed through primary data structure, in combination with read-memory trade-off mechanisms like caching and filtering. This raises re-tuning costs as optimal trade-off targets change, due to restructuring of stored data. We show that write-memory trade-off mechanisms are under-developed in current designs, and propose a new approach to dynamic key-value store optimization using a novel read-/write-balanced on-disk structure, the TurtleTree, and flexible read-memory & write-memory tuning knobs. We describe how the design of TurtleKV, our prototype, avoids in-memory bottlenecks to achieve high performance across a wide range of tuning parameters. When evaluated using YCSB, TurtleKV matches state-of-the-art SplinterDB for inserts, and is 5x/12x faster than RockDB/WiredTiger. In mixed workloads, TurtleKV is 16-25% faster than SplinterDB, >4x RocksDB, and 3-6x WiredTiger. TurtleKV is 2-9x faster than the others for point-query workloads, and has the best scan performance of the write-optimized systems tested.

cs.DB↗

arXiv2025

CiteEval: Principle-Driven Citation Evaluation for Source Attribution

Citation quality is crucial in information-seeking systems, directly influencing trust and the effectiveness of information access. Current evaluation frameworks, both human and automatic, mainly rely on Natural Language Inference (NLI) to assess binary or ternary supportiveness from cited sources, which we argue is a suboptimal proxy for citation evaluation. In this work we introduce CiteEval, a citation evaluation framework driven by principles focusing on fine-grained citation assessment within a broad context, encompassing not only the cited sources but the full retrieval context, user query, and generated text. Guided by the proposed framework, we construct CiteBench, a multi-domain benchmark with high-quality human annotations on citation quality. To enable efficient evaluation, we further develop CiteEval-Auto, a suite of model-based metrics that exhibit strong correlation with human judgments. Experiments across diverse systems demonstrate CiteEval-Auto's superior ability to capture the multifaceted nature of citations compared to existing metrics, offering a principled and scalable approach to evaluate and improve model-generated citations.

cs.CL↗

arXiv2024

Visualizing Distributed Traces in Aggregate

Distributed systems are comprised of many components that communicate together to form an application. Distributed tracing gives us visibility into these complex interactions, but it can be difficult to reason about the system's behavior, even with traces. Systems collect large amounts of tracing data even with low sampling rates. Even when there are patterns in the system, it is often difficult to detect similarities in traces since current tools mainly allow developers to visualize individual traces. Debugging and system optimization is difficult for developers without an understanding of the whole trace dataset. In order to help present these similarities, this paper proposes a method to aggregate traces in a way that groups together and visualizes similar traces. We do so by assigning a few traces that are representative of each set. We suggest that traces can be grouped based on how many services they share, how many levels the graph has, how structurally similar they are, or how close their latencies are. We also develop an aggregate trace data structure as a way to comprehensively visualize these groups and a method for filtering out incomplete traces if a more complete version of the trace exists. The unique traces of each group are especially useful to developers for troubleshooting. Overall, our approach allows for a more efficient method of analyzing system behavior.

cs.DC↗

arXiv2024

RAG-QA Arena: Evaluating Domain Robustness for Long-form Retrieval Augmented Question Answering

Question answering based on retrieval augmented generation (RAG-QA) is an important research topic in NLP and has a wide range of real-world applications. However, most existing datasets for this task are either constructed using a single source corpus or consist of short extractive answers, which fall short of evaluating large language model (LLM) based RAG-QA systems on cross-domain generalization. To address these limitations, we create Long-form RobustQA (LFRQA), a new dataset comprising human-written long-form answers that integrate short extractive answers from multiple documents into a single, coherent narrative, covering 26K queries and large corpora across seven different domains. We further propose RAG-QA Arena by directly comparing model-generated answers against LFRQA's answers using LLMs as evaluators. We show via extensive experiments that RAG-QA Arena and human judgments on answer quality are highly correlated. Moreover, only 41.3% of the most competitive LLM's answers are preferred to LFRQA's answers, demonstrating RAG-QA Arena as a challenging evaluation platform for future research.

cs.CL↗

arXiv2023

Mediation analysis with unmeasured confounding between parallel mediators and outcome

Mediation analysis extending beyond single mediators has gained significant attention in recent years. However, related methods often assume the absence of unmeasured mediator-outcome confounding. To address this, we develop a mediation analysis framework that accounts for such confounding within a linear structural equation model with parallel mediators. Specifically, we introduce a pseudo proxy variable to capture unmeasured confounding, allowing us to identify causal parameters. Leveraging this proxy, we propose a partially penalized method to identify mediators that significantly affect the outcome. The resultant estimates are consistent, and the estimates of nonzero parameters are asymptotically normal. Motivated by these results, we further introduce a procedure that can consistently select active mediation pathways with large probability. Simulation studies demonstrate the superior performance of the proposed approach. Finally, we apply our approach to genomic data, identifying gene expressions that potentially mediate the impact of a genetic variant on mouse obesity.

stat.ME↗

arXiv2023

On the Temporal-spatial Analysis of Estimating Urban Traffic Patterns Via GPS Trace Data of Car-hailing Vehicles

Car-hailing services have become a prominent data source for urban traffic studies. Extracting useful information from car-hailing trace data is essential for effective traffic management, while discrepancies between car-hailing vehicles and urban traffic should be considered. This paper proposes a generic framework for estimating and analyzing urban traffic patterns using car-hailing trace data. The framework consists of three layers: the data layer, the interactive software layer, and the processing method layer. By pre-processing car-hailing GPS trace data with operations such as data cutting, map matching, and trace correction, the framework generates tensor matrices that estimate traffic patterns for car-hailing vehicle flow and average road speed. An analysis block based on these matrices examines the relationships and differences between car-hailing vehicles and urban traffic patterns, which have been overlooked in previous research. Experimental results demonstrate the effectiveness of the proposed framework in examining temporal-spatial patterns of car-hailing vehicles and urban traffic. For temporal analysis, urban road traffic displays a bimodal characteristic while car-hailing flow exhibits a 'multi-peak' pattern, fluctuating significantly during holidays and thus generating a hierarchical structure. For spatial analysis, the heat maps generated from the matrices exhibit certain discrepancies, but the spatial distribution of hotspots and vehicle aggregation areas remains similar.

stat.AP↗

arXiv2023

An Adaptive Ensemble Framework for Addressing Concept Drift in IoT Data Streams

In the modern era of digital transformation, the evolution of the fifth-generation (5G) wireless network has played a pivotal role in revolutionizing communication technology and accelerating the growth of smart technology applications. Enabled by the high-speed, low-latency characteristics of 5G, these applications have shown significant potential in various sectors, from healthcare and transportation to energy management and beyond. As a crucial component of smart technology, IoT systems for service delivery often face concept drift issues in network data stream analytics due to dynamic IoT environments, resulting in performance degradation. In this article, we propose a drift-adaptive framework called Adaptive Exponentially Weighted Average Ensemble (AEWAE) consisting of three stages: IoT data preprocessing, base model learning, and online ensembling. It is a data stream analytics framework that integrates dynamic adjustments of ensemble methods to tackle various scenarios. Experimental results on two public IoT datasets demonstrate that our proposed framework outperforms state-of-the-art methods, achieving high accuracy and efficiency in IoT data stream analytics.

cs.CR↗

arXiv2022

Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks

Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we identify the issue of tokenization inconsistency that is commonly neglected in training these models. This issue damages the extractive nature of these tasks after the input and output are tokenized inconsistently by the tokenizer, and thus leads to performance drop as well as hallucination. We propose a simple yet effective fix to this issue and conduct a case study on extractive QA. We show that, with consistent tokenization, the model performs better in both in-domain and out-of-domain datasets, with a notable average of +1.7 F2 gain when a BART model is trained on SQuAD and evaluated on 8 QA datasets. Further, the model converges faster, and becomes less likely to generate out-of-context answers. With these findings, we would like to call for more attention on how tokenization should be done when solving extractive tasks and recommend applying consistent tokenization during training.

cs.CL↗

arXiv2022

Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations

There has been great progress in unifying various table-to-text tasks using a single encoder-decoder model trained via multi-task learning (Xie et al., 2022). However, existing methods typically encode task information with a simple dataset name as a prefix to the encoder. This not only limits the effectiveness of multi-task learning, but also hinders the model's ability to generalize to new domains or tasks that were not seen during training, which is crucial for real-world applications. In this paper, we propose compositional task configurations, a set of prompts prepended to the encoder to improve cross-task generalization of unified models. We design the task configurations to explicitly specify the task type, as well as its input and output types. We show that this not only allows the model to better learn shared knowledge across different tasks at training, but also allows us to control the model by composing new configurations that apply novel input-output combinations in a zero-shot manner. We demonstrate via experiments over ten table-to-text tasks that our method outperforms the UnifiedSKG baseline by noticeable margins in both in-domain and zero-shot settings, with average improvements of +0.5 and +12.6 from using a T5-large backbone, respectively.

cs.CL↗

arXiv2022

Semiparametric Efficient Dimension Reduction in multivariate regression with an Inner Envelope

Recently, Su and Cook proposed a dimension reduction technique called the inner envelope which can be substantially more efficient than the original envelope or existing dimension reduction techniques for multivariate regression. However, their technique relied on a linear model with normally distributed error, which may be violated in practice. In this work, we propose a semiparametric variant of the inner envelope that does not rely on the linear model nor the normality assumption. We show that our proposal leads to globally and locally efficient estimators of the inner envelope spaces. We also present a computationally tractable algorithm to estimate the inner envelope. Our simulations and real data analysis show that our method is both robust and efficient compared to existing dimension reduction methods in a diverse array of settings.

stat.ME↗

arXiv2021

Magnetic Feshbach resonances in collisions of $^{23}$Na$^{40}$K with $^{40}$K

We present measurements of more than 80 magnetic Feshbach resonances in collisions of ultracold $^{23}$Na$^{40}$K with $^{40}$K. We assign quantum numbers to a group of low-field resonances and show that they are due to long-range states of the triatomic complex in which the quantum numbers of the separated atom and molecule are approximately preserved. The resonant states are not members of chaotic bath of short-range states. Similar resonances are expected to be a common feature of alkali-metal diatom + atom systems.

physics.atom-ph↗

arXiv2021

Envelope Methods with Ignorable Missing Data

Envelope method was recently proposed as a method to reduce the dimension of responses in multivariate regressions. However, when there exists missing data, the envelope method using the complete case observations may lead to biased and inefficient results. In this paper, we generalize the envelope estimation when the predictors and/or the responses are missing at random. Specifically, we incorporate the envelope structure in the expectation-maximization (EM) algorithm. As the parameters under the envelope method are not pointwise identifiable, the EM algorithm for the envelope method was not straightforward and requires a special decomposition. Our method is guaranteed to be more efficient, or at least as efficient as, the standard EM algorithm. Moreover, our method has the potential to outperform the full data MLE. We give asymptotic properties of our method under both normal and non-normal cases. The efficiency gain over the standard EM is confirmed in simulation studies and in an application to the Chronic Renal Insufficiency Cohort (CRIC) study.

stat.ME↗

arXiv2021

Mixed Effects Envelope Models

When multiple measures are collected repeatedly over time, redundancy typically exists among responses. The envelope method was recently proposed to reduce the dimension of responses without loss of information in regression with multivariate responses. It can gain substantial efficiency over the standard least squares estimator. In this paper, we generalize the envelope method to mixed effects models for longitudinal data with possibly unbalanced design and time-varying predictors. We show that our model provides more efficient estimators than the standard estimators in mixed effects models. Improved accuracy and efficiency of the proposed method over the standard mixed effects model estimator are observed in both the simulations and the Action to Control Cardiovascular Risk in Diabetes (ACCORD) study.

stat.ME↗

arXiv2021

Envelopes for multivariate linear regression with linearly constrained coefficients

A constrained multivariate linear model is a multivariate linear model with the columns of its coefficient matrix constrained to lie in a known subspace. This class of models includes those typically used to study growth curves and longitudinal data. Envelope methods have been proposed to improve estimation efficiency in the class of unconstrained multivariate linear models, but have not yet been developed for constrained models that we develop in this article. We first compare the standard envelope estimator based on an unconstrained multivariate model with the standard estimator arising from a constrained multivariate model in terms of bias and efficiency. Then, to further improve efficiency, we propose a novel envelope estimator based on a constrained multivariate model. Novel envelope-based testing methods are also proposed. We provide support for our proposals by simulations and by studying the classical dental data and data from the China Health and Nutrition Survey and a study of probiotic capacity to reduced Salmonella infection.

stat.ME↗

arXiv2020

AWS CORD-19 Search: A Neural Search Engine for COVID-19 Literature

Coronavirus disease (COVID-19) has been declared as a pandemic by WHO with thousands of cases being reported each day. Numerous scientific articles are being published on the disease raising the need for a service which can organize, and query them in a reliable fashion. To support this cause we present AWS CORD-19 Search (ACS), a public, COVID-19 specific, neural search engine that is powered by several machine learning systems to support natural language based searches. ACS with capabilities such as document ranking, passage ranking, question answering and topic classification provides a scalable solution to COVID-19 researchers and policy makers in their search and discovery for answers to high priority scientific questions. We present a quantitative evaluation and qualitative analysis of the system against other leading COVID-19 search platforms. ACS is top performing across these systems yielding quality results which we detail with relevant examples in this work.

cs.IR↗