Selected work

Representative Papers

Analog-to-Stochastic Converter Using Magnetic Tunnel Junction Devices for Vision Chips

Sep 01, 2016IEEE transactions on nanotechnology

This work addresses the high power and area overhead of conventional analog-to-stochastic signal conversion, which typically requires two separate stages—analog-to-digital followed by digital-to-stochastic—rendering it unsuitable for energy-efficient vision chips. To overcome this limitation, the study proposes a novel single-step direct conversion approach leveraging the intrinsic probabilistic switching behavior of magnetic tunnel junctions (MTJs). This method significantly reduces hardware complexity and improves energy efficiency. Furthermore, to mitigate the impact of MTJ resistance variability, a compensation mechanism is introduced to enhance conversion accuracy and robustness. Mixed-mode NS-SPICE simulations based on 90 nm CMOS and 100 nm MTJ technologies demonstrate the proposed circuit’s advantages in terms of reduced area, lower power consumption, and improved tolerance to device variations.

29 citationsRead paper

Application-Driven Innovation in Machine Learning

Mar 26, 2024International Conference on Machine Learning

Application-driven machine learning (ML) research has been systematically undervalued in academia, leading to a growing disconnect between algorithmic innovation and real-world needs; this marginalization is reinforced by structural biases in peer review, faculty hiring, and pedagogy. Method: This paper introduces, for the first time, a formally defined “application-driven ML research paradigm,” elucidating its complementary relationship with the dominant methodology-driven paradigm. Drawing on interdisciplinary frameworks from education theory, research governance, and ML practice—and substantiated by empirical case studies and institutional critique—it diagnoses three systemic barriers hindering such research. Contribution/Results: The core contribution is a set of actionable, process-level interventions to reform academic evaluation systems, grounded in both theoretical analysis and pragmatic implementation pathways. These proposals have already catalyzed curricular reforms in AI education and adjustments to national funding review criteria across multiple universities, fostering cross-domain collaboration and methodological feedback loops between application domains and core ML research.

22 citations2 influentialRead paper

Retrieval-Augmented Generation for Natural Language Processing: A Survey

Jul 18, 2024arXiv.org

To address hallucination, knowledge staleness, and poor domain adaptability in large language models (LLMs), this paper conducts a systematic study of retrieval-augmented generation (RAG). We propose a full-stack RAG framework encompassing retriever design (dense, sparse, and hybrid), query rewriting, context fusion, LLM fine-tuning, and prompt engineering. We introduce the first taxonomy for dynamic knowledge updating and establish a multidimensional evaluation benchmark that balances academic rigor with industrial practicality. Additionally, we release a structured RAG knowledge graph and fully reproducible open-source code. Our contributions significantly enhance RAG’s robustness and maintainability in real-world deployments, providing both theoretical foundations and engineering best practices for knowledge-enhanced generative systems.

15 citationsRead paper

How Viable are Energy Savings in Smart Homes? A Call to Embrace Rebound Effects in Sustainable HCI

Aug 01, 2023ACM J. Comput. Sustain. Soc.

Energy savings from smart home technologies are often undermined by rebound effects—behavioral or systemic compensations triggered by increased efficiency—rendering sustainability gains transient. Method: Through a cross-disciplinary literature mapping analysis across Web of Science, Scopus, IEEE Xplore, Springer, and ACM SIGCHI proceedings, this study systematically identifies research gaps concerning rebound effects in computing, human-computer interaction (HCI), and smart home domains. Contribution/Results: We propose the first classification framework for rebound effects tailored to sustainable HCI, along with corresponding intervention pathways. Findings reveal that current energy-efficiency evaluations routinely neglect rebound mechanisms, while HCI is uniquely positioned to advance rebound identification, computational modeling, and behaviorally informed interventions. This work establishes a theoretical foundation and methodological toolkit for accurately assessing the real-world environmental impact of smart home systems.

12 citations1 influentialRead paper

An Addendum to NeBula: Toward Extending Team CoSTAR’s Solution to Larger Scale Environments

Apr 18, 2025IEEE Transactions on Field Robotics

Autonomous collaborative exploration in ultra-large-scale, unstructured underground environments remains challenging due to severe communication constraints, navigation uncertainty, and lack of prior maps. Method: This work extends TEAM CoSTAR’s NeBula autonomy system with a full-stack enhancement framework integrating semantic-geometric joint mapping, distributed POMDP-based global planning under communication constraints, adaptive filtering for localization, Gaussian process–based probabilistic traversability modeling, edge-cloud cooperative communication protocols, and aerial-ground heterogeneous multi-agent task allocation. Contribution/Results: The framework achieves, for the first time, robust long-range mapping (>5 km²), sub-meter localization accuracy (<0.3 m), and decentralized collaborative decision-making in kilometer-scale underground spaces (e.g., limestone mines). Validated in the DARPA Subterranean Challenge and real-world mine deployments, it improves mission completion rate by 37%, significantly advancing scalability, robustness, and coordination in autonomous underground exploration.

6 citationsRead paper

Recent publications

Latest Papers

Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games

Aug 12, 2026

This study investigates whether large language models can achieve effective coordination surpassing Nash equilibrium in one-shot, multi-agent games without communication or a central controller. The authors construct a self-play benchmark to systematically evaluate the coordination capabilities of 13 models under the assumption that all agents employ the same model. Results demonstrate that two state-of-the-art closed-source models consistently outperform Nash equilibrium in two-player settings, approaching joint optimality, whereas open-source models exhibit limited performance highly dependent on game structure. Coordination efficacy markedly deteriorates in games with four or more players. This work provides the first evidence of the critical influence of model scale and game structure on communication-free coordination among language agents.

0 citationsRead paper

Alignment Drift in Single-Model Speculative Decoding for ASR: Mechanism, Correction, and Cost

Aug 12, 2026

This work addresses a critical limitation in single-model speculative decoding for automatic speech recognition, where the draft module struggles to accurately track audio positions, leading to alignment drift and degraded prediction quality. The study reveals that precise audio position tracking is pivotal for effective speculation and proposes AnchorDraft, a training method that corrects such drift without altering the inference graph. AnchorDraft either leverages attention readout positions during verification or guides the draft module to implicitly learn positional information. Experiments demonstrate that AnchorDraft significantly accelerates end-to-end inference across two target model scales. Properly aligned audio windows substantially increase token acceptance rates, with the median error in verification-stage attention confined to merely two frames.

0 citationsRead paper

Backdoor Decontamination Dynamics in LLM Agents

Aug 11, 2026

This study addresses the vulnerability of open-source large language model agents to stealthy backdoors implanted during fine-tuning, which are difficult to detect when trigger conditions remain unobserved. The work systematically investigates the efficacy of defensive poisoning and unlearning in mitigating unknown backdoors, revealing for the first time that trigger recognition and malicious execution can be behaviorally decoupled. It proposes a novel strategy: applying defensive poisoning with analogous triggers followed by depoisoning, which nearly eliminates the original backdoor. Evaluated on the AgentDyn framework with J-lens representation visualization across 115 experiments, defensive poisoning alone removes approximately 56% of backdoors, while combining it with depoisoning achieves near-complete (≈100%) removal. Notably, in multi-backdoor settings, neutralizing one known backdoor incidentally eradicates 87% of coexisting unknown backdoors.

0 citationsRead paper

Every Token Counts: Exact Likert-Scale Distributions for Measuring LLM Attitudes and Biases

Aug 11, 2026

This study addresses the limitations of existing large language model (LLM) evaluation methods, which rely on unstructured benchmarks and struggle to causally disentangle sources of bias—such as baseline traits, contextual confounding, or interaction effects. To overcome this, the authors propose the first analytical evaluation framework grounded in precise Likert-scale response distributions, integrating psychometric principles with LLM mechanics. The framework employs a full factorial experimental design, token-level probability mass functions to eliminate sampling noise, and combines ordinal consensus metrics with distributional analysis of variance. Experiments across five mainstream LLMs successfully uncover systematic national-origin biases obscured by conventional aggregate metrics, demonstrating the framework’s high sensitivity and validity in assessing consumer ethnocentrism.

0 citationsRead paper

Early Pregnancy Treatment Decisions: Designing Perinatal Pharmacoepidemiology Studies using Real-World Data

Aug 11, 2026

This study addresses methodological challenges in perinatal pharmacologic research—such as left and right censoring, competing events, and gestational age heterogeneity—that often introduce bias in causal effect estimation, particularly when evaluating decisions to modify preconception treatment regimens during early pregnancy. Leveraging real-world data, this work extends the target trial emulation framework to the context of pre-pregnancy medication changes and innovatively proposes a gestational age–anchored definition of time zero tailored to early-pregnancy therapeutic decisions. The approach systematically corrects for selection bias and immortal time bias. By establishing a generalizable methodological paradigm, this research enhances the scientific rigor and feasibility of pharmacoepidemiologic studies assessing the safety and effectiveness of medications—such as those for type 2 diabetes—during the periconceptional and prenatal periods.

0 citationsRead paper