Zhengwu Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Zhengwu Liu has a total of 168 co-authors. Here are the top co-authors by number of shared publications:
- Bin Gao
Tsinghua University
17 shared publications - Huaqiang Wu
Tsinghua University
17 shared publications - Jianshi Tang
Tsinghua University
16 shared publications - Ngai Wong
Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China
16 shared publications He Qian
School of Integrated Circuits, Beijing Innovation Center for Integrated Circuits, Tsinghua University, Beijing, China
12 shared publicationsPeng Yao
School of Integrated Circuits, Beijing Advanced Innovation Center for Integrated Circuits, Tsinghua University, Beijing, China
10 shared publications
Latest Publications
- Open Access
- 20 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 8006
Memristive singular value decomposition
Singular value decomposition (SVD) underpins low-rank representation across scientific computing, signal processing, and machine learning. However, iterative computations in SVD are energy-intensive on conventional von Neumann architectures with separate storage and computation units, posing significant challenges for complex information processing. Here, we present memristive SVD (MSVD), built on compute-in-memory (CIM) memristor chips, enabled by a selective representation enhanced architecture (SREA) that ensures numerical fidelity across iterations. We demonstrate MSVD across three tiers of increasing complexity: low-rank approximation for image enhancement and epidemiological data reconstruction; user-scalable recognition where incremental MSVD exploits persistent in-memory storage to incorporate new users without remapping existing information; and large language model weight decomposition where MSVD-based initialization consistently outperforms the standard fine-tuning method on mathematical benchmarks. Beyond software-comparable accuracy, SREA reduces energy overhead and accelerates convergence over unenhanced MSVD, and the full system achieves order-of-magnitude gains in energy efficiency and speed over conventional hardware across all demonstrated scenarios, with these advantages growing progressively with each update cycle in the incremental setting. This work accelerates SVD across various scenarios and extends memristor-based systems towards general computing applications.
- 20 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 8006
Open AccessMemristive singular value decomposition
Singular value decomposition (SVD) underpins low-rank representation across scientific computing, signal processing, and machine learning. However, iterative computations in SVD are energy-intensive on conventional von Neumann architectures with separate storage and computation units, posing significant challenges for complex information processing. Here, we present memristive SVD (MSVD), built on compute-in-memory (CIM) memristor chips, enabled by a selective representation enhanced architecture (SREA) that ensures numerical fidelity across iterations. We demonstrate MSVD across three tiers of increasing complexity: low-rank approximation for image enhancement and epidemiological data reconstruction; user-scalable recognition where incremental MSVD exploits persistent in-memory storage to incorporate new users without remapping existing information; and large language model weight decomposition where MSVD-based initialization consistently outperforms the standard fine-tuning method on mathematical benchmarks. Beyond software-comparable accuracy, SREA reduces energy overhead and accelerates convergence over unenhanced MSVD, and the full system achieves order-of-magnitude gains in energy efficiency and speed over conventional hardware across all demonstrated scenarios, with these advantages growing progressively with each update cycle in the incremental setting. This work accelerates SVD across various scenarios and extends memristor-based systems towards general computing applications.
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
PEACE: Algorithm-Hardware Co-Design for PE-Granular Energy-Aware Compressed GEMM Acceleration
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessPEACE: Algorithm-Hardware Co-Design for PE-Granular Energy-Aware Compressed GEMM Acceleration
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
EdgeSC: Universal Stochastic Computing Architecture for Efficient Edge Detection
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessEdgeSC: Universal Stochastic Computing Architecture for Efficient Edge Detection
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (7) , 2518-2532
From SMURF to HI-SMURF: Scalable Multivariate Nonlinear Function Approximation via Compact Stochastic Architectures
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (7) , 2518-2532
From SMURF to HI-SMURF: Scalable Multivariate Nonlinear Function Approximation via Compact Stochastic Architectures
- Open Access
- 11 April 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (5) , 1-23
Hardware-aware Low-Rank Adaptation for Large Language Models Based on Hybrid Compute-in-Memory Architecture
Low-rank adaptation (LoRA) is a predominant parameter-efficient finetuning method for adapting large language models (LLMs) to downstream tasks. Meanwhile, Compute-in-Memory (CIM) architectures demonstrate superior energy efficiency due to their array-level parallel in-memory computing designs. In this article, we propose deploying the LoRA-finetuned LLMs on the hybrid CIM architecture (i.e., pretrained weights onto energy-efficient Resistive Random-Access Memory (RRAM) and LoRA branches onto noise-free Static Random-Access Memory (SRAM)), reducing the energy cost to about 3% compared with the Nvidia A100 GPU. However, the inherent noise of RRAM on the saved weights leads to performance degradation, simultaneously. To address this issue, we design a novel Hardware-aware Low-rank Adaptation (HaLoRA) method. The key insight is to train a LoRA branch that is robust toward such noise and then deploy it on noise-free SRAM, while the extra cost is negligible since the parameters of LoRAs are much fewer than pretrained weights (e.g., 0.15% for LLaMA-3.2 1B model). To improve the robustness towards the noise, we theoretically analyze the gap between the optimization trajectories of the LoRA branch under both ideal and noisy conditions and further design an extra loss to minimize the upper bound of this gap. Therefore, we can enjoy both energy efficiency and accuracy during inference. Experiments finetuning the Qwen and LLaMA series demonstrate the effectiveness of HaLoRA across multiple reasoning tasks, achieving up to 22.7 improvement in average score while maintaining robustness at various noise types and noise levels.
- 11 April 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (5) , 1-23
Open AccessHardware-aware Low-Rank Adaptation for Large Language Models Based on Hybrid Compute-in-Memory Architecture
Low-rank adaptation (LoRA) is a predominant parameter-efficient finetuning method for adapting large language models (LLMs) to downstream tasks. Meanwhile, Compute-in-Memory (CIM) architectures demonstrate superior energy efficiency due to their array-level parallel in-memory computing designs. In this article, we propose deploying the LoRA-finetuned LLMs on the hybrid CIM architecture (i.e., pretrained weights onto energy-efficient Resistive Random-Access Memory (RRAM) and LoRA branches onto noise-free Static Random-Access Memory (SRAM)), reducing the energy cost to about 3% compared with the Nvidia A100 GPU. However, the inherent noise of RRAM on the saved weights leads to performance degradation, simultaneously. To address this issue, we design a novel Hardware-aware Low-rank Adaptation (HaLoRA) method. The key insight is to train a LoRA branch that is robust toward such noise and then deploy it on noise-free SRAM, while the extra cost is negligible since the parameters of LoRAs are much fewer than pretrained weights (e.g., 0.15% for LLaMA-3.2 1B model). To improve the robustness towards the noise, we theoretically analyze the gap between the optimization trajectories of the LoRA branch under both ideal and noisy conditions and further design an extra loss to minimize the upper bound of this gap. Therefore, we can enjoy both energy efficiency and accuracy during inference. Experiments finetuning the Qwen and LLaMA series demonstrate the effectiveness of HaLoRA across multiple reasoning tasks, achieving up to 22.7 improvement in average score while maintaining robustness at various noise types and noise levels.
- Top Cited Papers
- 23 March 2026
- Published by Springer Nature in Nature Electronics
- Vol. 9 (6) , 652-664
In situ spectral reconstruction based on a memristor chip for energy-efficient computational spectrometry
- 23 March 2026
- Published by Springer Nature in Nature Electronics
- Vol. 9 (6) , 652-664
Top Cited PapersIn situ spectral reconstruction based on a memristor chip for energy-efficient computational spectrometry
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
TinnitusLLM: A Multimodal Large Language Model Framework for Tinnitus Diagnosis Through EEG-fMRI Fusion Learning
Accurate tinnitus diagnosis is crucial for enabling timely therapeutic intervention and longitudinal treatment monitoring. While non-invasive neuroimaging modalities-particularly electroencephalography (EEG) with millisecond temporal resolution and functional magnetic resonance imaging (fMRI) with millimeter spatial resolution- provide complementary neural features, existing diagnostic approaches remain constrained to unimodal analysis of EEG or fMRI data, inherently limiting diagnostic precision and clinical generalizability. This paper introduces TinnitusLLM, the first multimodal large language model (LLM) framework that synergistically integrates EEG and fMRI features for tinnitus diagnosis. To enable LLM-based interpretation of neural signals, this framework integrates three key components: (1) a neuroinspired positional encoding mechanism that injects neurophysiological priors into the embedding space, enabling neurologically grounded, dynamic positional mapping of EEG and fMRI tokens; (2) multimodal autoregressive pretraining on more than 500 hours of EEG and 250 hours of fMRI data to learn causally informed predictive representations; and (3) fine-tuning with a cross-modal, subject-invariant adversarial learning strategy that enforces subject-independent constraints in the shared cross-modal feature space, thereby substantially improving diagnostic robustness across subjects. We validate TinnitusLLM through comprehensive experiments on a rigorously collected multimodal dataset containing 20 participants. Quantitative evaluations demonstrate that TinnitusLLM achieves superior cross-subject diagnostic accuracy compared to the state-of-the-art baseline methods. These results underscore TinnitusLLM's potential as a clinically viable framework for objective tinnitus assessment through multimodal neural decoding.
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
TinnitusLLM: A Multimodal Large Language Model Framework for Tinnitus Diagnosis Through EEG-fMRI Fusion Learning
Accurate tinnitus diagnosis is crucial for enabling timely therapeutic intervention and longitudinal treatment monitoring. While non-invasive neuroimaging modalities-particularly electroencephalography (EEG) with millisecond temporal resolution and functional magnetic resonance imaging (fMRI) with millimeter spatial resolution- provide complementary neural features, existing diagnostic approaches remain constrained to unimodal analysis of EEG or fMRI data, inherently limiting diagnostic precision and clinical generalizability. This paper introduces TinnitusLLM, the first multimodal large language model (LLM) framework that synergistically integrates EEG and fMRI features for tinnitus diagnosis. To enable LLM-based interpretation of neural signals, this framework integrates three key components: (1) a neuroinspired positional encoding mechanism that injects neurophysiological priors into the embedding space, enabling neurologically grounded, dynamic positional mapping of EEG and fMRI tokens; (2) multimodal autoregressive pretraining on more than 500 hours of EEG and 250 hours of fMRI data to learn causally informed predictive representations; and (3) fine-tuning with a cross-modal, subject-invariant adversarial learning strategy that enforces subject-independent constraints in the shared cross-modal feature space, thereby substantially improving diagnostic robustness across subjects. We validate TinnitusLLM through comprehensive experiments on a rigorously collected multimodal dataset containing 20 participants. Quantitative evaluations demonstrate that TinnitusLLM achieves superior cross-subject diagnostic accuracy compared to the state-of-the-art baseline methods. These results underscore TinnitusLLM's potential as a clinically viable framework for objective tinnitus assessment through multimodal neural decoding.
- Open Access
- 6 February 2026
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 12 (6) , eady5485
Privacy-preserving data analysis using a memristor chip with colocated authentication and processing
Privacy-preserving data analysis is essential in health care applications to safeguard sensitive patient information while enabling medical monitoring and diagnostics. However, existing solutions generally separate security from analysis modules and memory from computation units, creating hardware and energy overheads that constrain their use in resource-limited medical devices. Here, we introduce the memristor-based colocated authentication and processing (CLAP) system, which achieves security-analysis integration through embedding physical unclonable functions within compute-in-memory architecture. To resolve the incompatibilities between these two features, we propose a differential stochastic mapping method by applying information theory principles. We demonstrate CLAP on a 130-nanometer memristor chip, validating its versatility across diverse information processing tasks. In an electrocardiogram data collection task, CLAP achieves device authentication with an area under the curve of 99.46% and efficient signal compression with a software-level percentage root mean square difference. CLAP demonstrates 146.0-fold energy efficiency gain and 17.6-fold area reduction, providing intrinsically secure hardware solutions that enhance both privacy preservation and computational efficiency for health care applications.
- 6 February 2026
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 12 (6) , eady5485
Open AccessPrivacy-preserving data analysis using a memristor chip with colocated authentication and processing
Privacy-preserving data analysis is essential in health care applications to safeguard sensitive patient information while enabling medical monitoring and diagnostics. However, existing solutions generally separate security from analysis modules and memory from computation units, creating hardware and energy overheads that constrain their use in resource-limited medical devices. Here, we introduce the memristor-based colocated authentication and processing (CLAP) system, which achieves security-analysis integration through embedding physical unclonable functions within compute-in-memory architecture. To resolve the incompatibilities between these two features, we propose a differential stochastic mapping method by applying information theory principles. We demonstrate CLAP on a 130-nanometer memristor chip, validating its versatility across diverse information processing tasks. In an electrocardiogram data collection task, CLAP achieves device authentication with an area under the curve of 99.46% and efficient signal compression with a software-level percentage root mean square difference. CLAP demonstrates 146.0-fold energy efficiency gain and 17.6-fold area reduction, providing intrinsically secure hardware solutions that enhance both privacy preservation and computational efficiency for health care applications.
- Open Access
- 21 January 2026
- Published by MDPI AG in Preprints.org
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. With actionable mechanistic interpretability evolving at a fast pace, we pledge to keep this survey up to date, ensuring it reflects the cutting-edge advances in this area.
- 21 January 2026
- Published by MDPI AG in Preprints.org
Open AccessLocate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. With actionable mechanistic interpretability evolving at a fast pace, we pledge to keep this survey up to date, ensuring it reflects the cutting-edge advances in this area.
- Open Access
- 17 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (2) , 1-24
PPD: A Portable and Highly Parallel Dispatching System for Deep Learning
The acceleration of inference process for deep learning models is closely tied with the parallelization capability of computational graph operators and the parallel scheduling strategies. Most existing deep learning compilers focus on optimizing intra-operator parallelism, while neglecting inter-operator parallelism. Furthermore, most industrial inference engines, such as PyTorch and TensorFlow, utilize a dataflow-based model to describe tasks and schedule operators. They are computationally expensive and operate in a topological order and are parallelized to run within a single CUDA stream. However, they fail to fully exploit the parallelism capabilities of multiple CUDA streams. In this article, we propose PPD, a portable, highly parallel dispatching system. It boosts the inference performance by dividing the computational graph into multiple taskflow-based subgraphs. Additionally, PPD entails a dispatching algorithm on a single GPU with multiple CUDA streams to enhance the parallelism and performance of model inference. PPD offers users a lightweight model definition and an inference C++ interface, allowing for seamless integration into any context. We also verify the feasibility of PPD on AMD and other graphics cards. We validate our approach on widely adopted neural network models with varying degrees of parallelism, and compare it with industrial inference engines. Experiments demonstrate that PPD outperforms SOTA methods by up to 2.28× .
- 17 December 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (2) , 1-24
Open AccessPPD: A Portable and Highly Parallel Dispatching System for Deep Learning
The acceleration of inference process for deep learning models is closely tied with the parallelization capability of computational graph operators and the parallel scheduling strategies. Most existing deep learning compilers focus on optimizing intra-operator parallelism, while neglecting inter-operator parallelism. Furthermore, most industrial inference engines, such as PyTorch and TensorFlow, utilize a dataflow-based model to describe tasks and schedule operators. They are computationally expensive and operate in a topological order and are parallelized to run within a single CUDA stream. However, they fail to fully exploit the parallelism capabilities of multiple CUDA streams. In this article, we propose PPD, a portable, highly parallel dispatching system. It boosts the inference performance by dividing the computational graph into multiple taskflow-based subgraphs. Additionally, PPD entails a dispatching algorithm on a single GPU with multiple CUDA streams to enhance the parallelism and performance of model inference. PPD offers users a lightweight model definition and an inference C++ interface, allowing for seamless integration into any context. We also verify the feasibility of PPD on AMD and other graphics cards. We validate our approach on widely adopted neural network models with varying degrees of parallelism, and compare it with industrial inference engines. Experiments demonstrate that PPD outperforms SOTA methods by up to 2.28× .