Papers by Ngai Wong
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models (2026.findings-acl)
Copied to clipboard
Hengyuan Zhang, Zhihao Zhang, Ercong Nie, Mingyang Wang, Zunhai Su, Yiwei Wang, Qianli Wang, Shuzhou Yuan, Xufeng Duan, Qibo Xue, Zeping Yu, Chenming Shang, Xiao Liang, Jing Xiong, Hui Shen, Chaofan Tao, Zhengwu Liu, Senjie Jin, Zhiheng Xi, Dongdong Zhang, Sophia Ananiadou, Tao Gui, Ruobing Xie, Hayden Kwok-Hay So, Hinrich Schuetze, Xuanjing Huang, Qi Zhang, Ngai Wong
| Challenge: | Existing literature on mechanistic interpretation (MI) treats it as an observational science, leaving practical applications underexplored. |
| Approach: | They propose a survey structured around the pipeline to identify and improve MI models. |
| Outcome: | The proposed framework enables tangible improvements in Alignment, Capability, and Efficiency. |
Structured Pruning for Efficient Generative Pre-trained Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Large-scale generative Pre-trained Language Models (PLMs) are limited in their deployment in real-world applications. |
| Approach: | They propose to prune the feed-forward networks of generative pre-trained language models to smaller widths without designing extra operators. |
| Outcome: | The proposed method achieves 1.51x/6.96x inference speedup on GPU/CPU with 67% size reduction. |
QuZO: Quantized Zeroth-Order Fine-Tuning for Large Language Models (2025.emnlp-main)
Copied to clipboard
Jiajun Zhou, Yifan Yang, Kai Zhen, Ziyue Liu, Yequan Zhao, Ershad Banijamali, Athanasios Mouchtaris, Ngai Wong, Zheng Zhang
| Challenge: | Large Language Models (LLMs) are quantized to lower precision to reduce memory cost and latency in inference. |
| Approach: | They propose a quantized zeroth-order framework for fine-tuning Large Language Models (LLMs) using low-precision forward passes. |
| Outcome: | The proposed method achieves comparable results to first-order methods in FP8 and superior accuracy in INT8 and INT4 training. |
Compression of Generative Pre-trained Language Models via Quantization (2022.acl-long)
Copied to clipboard
| Challenge: | Existing methods to compress generative pre-trained language models fail on generative tasks due to homogeneous word embeddings and limited memory. |
| Approach: | They propose a token-level contrastive distillation method to learn distinguishable word embeddings and a module-wise dynamic scaling method to make quantizers adaptive to different modules. |
| Outcome: | The proposed method outperforms the state-of-the-art compression methods on generative PLMs by a clear margin. |
TreeReview: A Dynamic Tree of Questions Framework for Deep and Efficient LLM-based Scientific Peer Review (2025.emnlp-main)
Copied to clipboard
Yuan Chang, Ziyue Li, Hengyuan Zhang, Yuanbo Kong, Yanru Wu, Hayden Kwok-Hay So, Zhijiang Guo, Liya Zhu, Ngai Wong
| Challenge: | Large Language Models (LLMs) have shown significant potential in assisting peer review, but current methods struggle to generate thorough and insightful reviews while maintaining efficiency. |
| Approach: | They propose a framework that models paper review as a hierarchical and bidirectional question-answering process. |
| Outcome: | The proposed framework outperforms baselines on full review generation and actionable feedback comments generation tasks while reducing LLM token usage by up to 80% compared to computationally intensive approaches. |
UNComp: Can Matrix Entropy Uncover Sparsity? — A Compressor Design from an Uncertainty-Aware Perspective (2025.emnlp-main)
Copied to clipboard
Jing Xiong, Jianghan Shen, Fanghua Ye, Chaofan Tao, Zhongwei Wan, Jianqiao Lu, Xun Wu, Chuanyang Zheng, Zhijiang Guo, Min Yang, Lingpeng Kong, Ngai Wong
| Challenge: | Deploying large language models (LLMs) for long-context inference remains challenging due to their substantial memory and computational demands. |
| Approach: | They propose an uncertainty-aware framework that leverages truncated matrix entropy to identify areas of low information content. |
| Outcome: | The proposed framework reduces the KV cache size to 4.74% of the original and achieves a 6% speedup. |
D-QRELO: Training- and Data-Free Delta Compression for Large Language Models via Quantization and Residual Low-Rank Approximation (2026.findings-acl)
Copied to clipboard
Junlin Li, Shuangyong Song, Guodong DU, Ngai Wong, Xuebo Liu, Yongxiang Li, Min Zhang, Jing Li, Xuelong Li
| Challenge: | Existing methods for fine-tuned large language models fail on fine-scale datasets . large data scale amplifies delta parameter magnitude, singular values, and entropy, causing compression errors. |
| Approach: | They propose a training- and data-free delta compression method that captures dominant delta structure and compensates residual low-rank approximation to recover fine-grained details from smaller residual error. |
| Outcome: | The proposed method outperforms existing methods on large-scale datasets on dense and MoE architectures. |
KG-RAG: Enhancing GUI Agent Decision-Making via Knowledge Graph-Driven Retrieval-Augmented Generation (2025.emnlp-main)
Copied to clipboard
Ziyi Guan, Jason Chun Lok Li, Zhijian Hou, Pingping Zhang, Donglai Xu, Yuzhi Zhao, Mengyang Wu, Jinpeng Chen, Thanh-Toan Nguyen, Pengfei Xian, Wenao Ma, Shengchao Qin, Graziano Chesi, Ngai Wong
| Challenge: | Recent advances in GUI agents have limited app-specific knowledge of complex mobile tasks. |
| Approach: | They propose a Knowledge Graph-driven Retrieval-Augmented Generation framework that transforms fragmented UTGs into structured vector databases for efficient real-time retrieval. |
| Outcome: | The proposed framework outperforms existing methods in a 75.8% success rate and 84.6% decision accuracy test across mobile apps. |
Mixture-of-Subspaces in Low-Rank Adaptation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Using a subspace-inspired Low-Rank Adaptation method, large language models can be optimized for downstream tasks using parameter-efficient finetuning. |
| Approach: | They propose a subspace-inspired Low-Rank Adaptation method that decomposes LoRA weights into two subspaces and merges them into the frozen original weight. |
| Outcome: | The proposed method outperforms LoRA on commonsense reasoning, visual instruction tuning, and subject-driven text-to-image generation tasks. |
Gradually Excavating External Knowledge for Implicit Complex Question Answering (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have gained attention for their human-comparable capabilities but they may not solve open-domain implicit questions due to out-of-date domain knowledge, one-shot generation and restricted comprehensiveness. |
| Approach: | They propose a gradual knowledge excavation framework for open-domain complex question answering using extrinsic knowledge and historical knowledge. |
| Outcome: | The proposed framework achieves 78.17% accuracy with less than 6% parameters of its competitors, setting new SOTA in the 10B LLM class. |
Revisiting Model Interpolation for Efficient Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models that interpolate weights of two specialized models can be abused for efficient reasoning. |
| Approach: | They propose to merge two specialized models and create a model that combines efficiency and efficiency. |
| Outcome: | The proposed method outperforms existing models on efficiency and effectiveness. |
Rethinking Kullback-Leibler Divergence in Knowledge Distillation for Large Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have been used in Knowledge Distillation (KD) to compress large models. |
| Approach: | They propose a Kullback-Leiber divergence method which adaptively allocates weights to combine RKL and FKL to reduce the size of Large Language Models (LLMs). |
| Outcome: | The proposed method outperforms baselines and improves diversity and quality of generated responses. |
Exploring Layer-wise Information Effectiveness for Post-Training Quantization in Small Language Models (2026.findings-acl)
Copied to clipboard
He Xiao, Qingyao Yang, Dirui Xie, Wendong XU, Zunhai Su, Runming Yang, Haobo Liu, Wenyong Zhou, Zhengwu Liu, Ngai Wong
| Challenge: | Large language models with billions of parameters are often over-provisioned . smaller models exhibit lower robustness under extreme low-bit quantization . |
| Approach: | They propose a hardware-native, metric-driven post-training quantization framework that keeps uniform bit-width within each layer while mixing precision across layers. |
| Outcome: | LieQ reduces large accuracy gap observed for large language models with billions of parameters while preserving standard multiplication kernels. |
Edge-free but Structure-aware: Prototype-Guided Knowledge Distillation from GNNs to MLPs (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods to train low-latency multilayer perceptrons (MLPs) on graph tasks are based on graph nodes and lack graph structural information. |
| Approach: | They propose to distill graph structural information from Graph Neural Networks (GNNs) to low-latency multilayer perceptrons (MLPs) on graph tasks. |
| Outcome: | The proposed method does not require graph edges (edge-free setting) yet learns structure-aware MLPs. |
Find Your Optimal Teacher: Personalized Data Synthesis via Router-Guided Multi-Teacher Distillation (2026.acl-long)
Copied to clipboard
Hengyuan Zhang, Shiping Yang, Xiao Liang, Chenming Shang, Yuxuan Jiang, Chaofan Tao, Jing Xiong, Hayden Kwok-Hay So, Ruobing Xie, Angel X. Chang, Ngai Wong
| Challenge: | Existing studies show that stronger models are not always optimal teachers, suggesting a mismatch between the teacher’s output and the student’s learning ability. |
| Approach: | They propose a method that routes each prompt to its optimal teacher via a query-level router that jointly considers the student models’ learnability and teacher models’ response quality. |
| Outcome: | The proposed method outperforms baselines on six benchmarks including instruct tuning and math reasoning settings. |
LoRETTA: Low-Rank Economic Tensor-Train Adaptation for Ultra-Low-Parameter Fine-Tuning of Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for parameter-efficient fine-tuning are limited by the growing number of trainable parameters with the rapid deployment of Large Language Models (LLMs). |
| Approach: | They propose a parameter-efficient framework that reduces trainable parameters through tensor-train decomposition. |
| Outcome: | The proposed methods achieve comparable or better performance than most widely used methods with up to 100 fewer parameters on the LLaMA-2-7B models. |
GuiLoMo: Allocating Experts and Ranks for LoRA-MoE via Bilevel Optimization with GuidedSelection Vectors (2025.findings-emnlp)
Copied to clipboard
Xinrong Chen, Hengyuan Zhang, Yingmin Qiu, Xiao Liang, Ziyue Li, Guanyu Wang, Weiping Li, Tong Mo, Hayden Kwok-Hay So, Ngai Wong
| Challenge: | Low-Rank Adaptation (LoRA) methods are efficient for a large language model with reduced computational costs. |
| Approach: | They propose a layer-wise expert numbers and ranks allocation strategy with GuidedSelection Vectors. |
| Outcome: | The proposed method achieves superior or comparable performance to all baselines on three backbone models. |
Weight-Inherited Distillation for Task-Agnostic BERT Compression (2024.findings-naacl)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) is a predominant approach for BERT compression. |
| Approach: | They propose a weight-inherited distillation method which directly transfers knowledge from the teacher to a compact student model by inheriting the weights. |
| Outcome: | The proposed method outperforms state-of-the-art KD-based methods on GLUE and SQUAD benchmarks. |