Papers by Zeyu Wang
Turning the Tide: Repository-based Code Reflection (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Code large language models (LLMs) enhance programming by understanding and generating code across languages. |
| Approach: | a new benchmark evaluates code understanding and generation in repositories using code large language models. |
| Outcome: | The proposed model improves code understanding and generation in repositories by evaluating 1,888 test cases across 6 programming languages. |
DetectRL-X: Towards Reliable Multilingual and Real-World LLM-Generated Text Detection (2026.acl-long)
Copied to clipboard
Junchao Wu, Yefeng Liu, Chenyu Zhu, Hao Zhang, Zeyu Wu, Tianqi Shi, Yichao Du, Longyue Wang, Weihua Luo, Jinsong Su, Derek F. Wong
| Challenge: | Existing detectors are limited in their ability to detect large language models generated content in multilingual environments. |
| Approach: | They propose a multilingual benchmark to evaluate advanced detectors across 8 dimensions to better align with real-world applications. |
| Outcome: | The proposed benchmark encompasses 8 languages commonly used in commercial contexts and collects human-written texts from 6 domains highly susceptible to LLM misuse. |
Probing Across Time: What Does RoBERTa Know and When? (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Current approaches to natural language processing rely on fixed artifacts such as language models . current studies have focused on how these models acquire and demonstrate knowledge . |
| Approach: | They apply probing techniques to examine how language models acquire knowledge . they aim to inform future work on more efficient pretraining and understanding dependencies . |
| Outcome: | The proposed model learns linguistic abstractions, factual and commonsense knowledge, and reasoning abilities fast, stably, and robustly across domains. |
Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection (2022.emnlp-main)
Copied to clipboard
Suchin Gururangan, Dallas Card, Sarah Dreier, Emily Gade, Leroy Wang, Zeyu Wang, Luke Zettlemoyer, Noah A. Smith
| Challenge: | Language models rely on massive web crawls for diverse text data, but are rife with undesirable content. |
| Approach: | They analyze newspaper articles written by students from across the country to determine whose language is preferred by a quality filter. |
| Outcome: | The results show that newspapers from wealthier, educated, and urban zones are more likely to be classified as high quality. |
Detecting Urgency in Multilingual Medical SMS in Kenya (2022.aacl-srw)
Copied to clipboard
| Challenge: | Access to mobile phones has increased exponentially over the last 20 years, providing an opportunity to connect patients with healthcare interventions through mobile phones. |
| Approach: | They propose to use natural language processing to improve nurses' management of messages from pregnant and postpartum women in Kenya. |
| Outcome: | The proposed model did not reach the clinical usefulness threshold but could improve nurse workflow and responsiveness to urgent messages. |
HiGen: Hierarchy-Aware Sequence Generation for Hierarchical Text Classification (2024.eacl-long)
Copied to clipboard
| Challenge: | Hierarchical text classification is a complex subtask under multi-label text classification . the relevance of document sections can vary based on the hierarchy level, necessitating a dynamic document representation. |
| Approach: | They propose a text-generation-based framework that uses language models to encode dynamic text representations. |
| Outcome: | The proposed framework surpasses existing methods while handling data and mitigating class imbalance. |
Audio Jailbreak: An Open Comprehensive Benchmark for Jailbreaking Large Audio-Language Models (2026.acl-long)
Copied to clipboard
Zirui Song, Qian Jiang, Mingxuan Cui, Mingzhe Li, Lang Gao, Zeyu Zhang, Zixiang Xu, Yanbo Wang, Guangxian Ouyang, Zhenhao Chen, Xiuying Chen
| Challenge: | a recent study evaluated large audio-language models against jailbreak attacks . a new benchmark is being developed to evaluate LAM safety against jailbreaking attacks based on temporal and semantic nature of speech . |
| Approach: | They propose a benchmark to evaluate LAM jailbreak vulnerabilities in adversarial audio prompts . they use a dataset of 1,495 adversarials to evaluate their performance . |
| Outcome: | The proposed benchmark evaluates state-of-the-art LAMs against jailbreak attacks . it demonstrates that even small, semantically preserved perturbations can reduce safety . |
Towards Hierarchical Multi-Step Reward Models for Enhanced Reasoning in Large Language Models (2026.findings-acl)
Copied to clipboard
Teng Wang, Jiang Zhangyi, Zhenqi He, Hailei Gong, Shenyang Tong, Wenhan Yang, Zeyu Li, Yanan Zheng, Zifan He, Zewen Ye, Shengjie Ma, Jianping Zhang
| Challenge: | Existing Process Reward Models (PRMs) are vulnerable to reward hacking and require expensive, large-scale annotation of reasoning steps. |
| Approach: | They propose a reward model approach which evaluates both individual and consecutive reasoning steps from fine-grained and coarse-grounded level. |
| Outcome: | Empirical results show that the proposed model performs better than existing PRMs and is more robust than existing models. |
TrendSim: Simulating Trending Topics in Social Media Under Poisoning Attacks with LLM-based Multi-agent System (2025.findings-naacl)
Copied to clipboard
Zeyu Zhang, Jianxun Lian, Chen Ma, Yaning Qu, Ye Luo, Lei Wang, Rui Li, Xu Chen, Yankai Lin, Le Wu, Xing Xie, Ji-Rong Wen
| Challenge: | Trending topics bring in a new channel for poisoning attacks, resulting in negative impacts on society. |
| Approach: | They propose an LLM-based multi-agent system to simulate trending topics in social media . they propose a time-aware interaction mechanism, centralized message dissemination, and an interactive system . |
| Outcome: | The proposed system simulates trending topics under poisoning attacks on social media platforms. |
Learning Gender-Neutral Word Embeddings (D18-1)
Copied to clipboard
| Challenge: | Word embeddings trained on human-generated corpora inherit strong gender stereotypes . prior studies show such embeddables exhibit social biases, such as gender stereotype . |
| Approach: | They propose a method to preserve gender information in certain dimensions of word vectors . they propose GN-GloVe, which is a gender-neutral variant of the word embedding model . |
| Outcome: | The proposed method preserves gender information in certain dimensions of word vectors while compelling other dimensions to be free of gender influence. |
DecompileBench: A Comprehensive Benchmark for Evaluating Decompilers in Real-World Scenarios (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches focus on syntactic correctness through synthetic micro-benchmarks or subjective human ratings, despite semantic fidelity and usability. |
| Approach: | They propose a framework that enables effective evaluation of decompilers in reverse engineering workflows . they compare six industrial-strength decompils and six recent LLM-powered approaches . |
| Outcome: | The proposed framework outperforms commercial tools in code understandability despite lower functionality correctness . it shows that it can transform human-centric reverse engineering workflows . |
Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models (2025.acl-long)
Copied to clipboard
Zihan Qiu, Zeyu Huang, Bo Zheng, Kaiyue Wen, Zekun Wang, Rui Men, Ivan Titov, Dayiheng Liu, Jingren Zhou, Junyang Lin
| Challenge: | Existing Mixture-of-Experts training frameworks use a micro-batch to calculate LBL . micro-batches are restricted to a single sequence, preventing expert specialization . |
| Approach: | They propose to use a global-batch to loosen the load balance constraint for MoEs models . they propose to synchronize fi across micro-batches and then use it to calculate the LBL . |
| Outcome: | The proposed global-batch LBL improves the domain specialization of experts . the micro-battery LBL is almost at the sequence level, and the router is pushed to distribute the token evenly . |
Powering Comparative Classification with Sentiment Analysis via Domain Adaptive Knowledge Transfer (2021.emnlp-main)
Copied to clipboard
| Challenge: | Comparative Preference Classification (CPC) is a natural language processing task that predicts whether a preference comparison exists between two entities in a given sentence . |
| Approach: | They propose a sentiment analyzer that learns sentiments to individual entities via domain adaptive knowledge transfer. |
| Outcome: | Experiments on the CompSent-19 dataset present a significant improvement on the F1 scores over the best existing CPC approaches. |
RGL: A Simple yet Effective Relation Graph Augmented Prompt-based Tuning Approach for Few-Shot Learning (2022.findings-naacl)
Copied to clipboard
| Challenge: | Pre-trained language models (PLMs) are a good starting point for downstream applications, but it is difficult to generalize them to new tasks given a few labeled samples. |
| Approach: | They propose to use Relation Graph augmented learning to improve the performance of few-shot natural language understanding tasks by rewriting the input sequence into a cloze question with masks. |
| Outcome: | Extensive experiments show that Relation Graph augmented learning (RGL) improves performance of prompt-based tuning strategies. |
Conditional Semantic Textual Similarity via Conditional Contrastive Learning (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods to assess similarity between sentences encounter over-estimation problem . compared to fuzzy representations, similarity is comparatively lower in terms of "The person's age". |
| Approach: | They propose a conditional contrastive learning framework that constructs positive and negative samples from two perspectives. |
| Outcome: | The proposed method achieves state-of-the-art performance with five models based on bi-encoder and tri-encoding architectures. |
Virtual Compiler Is All You Need For Assembly Code Search (2024.acl-long)
Copied to clipboard
| Challenge: | Using a large dataset, we find that assembly code search is a significant task for reverse engineers. |
| Approach: | They propose to train a Large Language Model (LLM) to emulate a general compiler. |
| Outcome: | The proposed model surpasses the baseline by 26%. |
Plot2Code: A Comprehensive Benchmark for Evaluating Multi-modal Large Language Models in Code Generation from Scientific Plots (2025.findings-naacl)
Copied to clipboard
| Challenge: | Multi-modal Large Language Models have shown remarkable progress in visual contexts, yet their ability to convert visual figures into executable code remains underexplored. |
| Approach: | They propose to use a set of visual coding metrics to assess MLLMs' visual . pass rate, text-match ratio, and GPT-4V rating judgement to assess the quality of generated code and rendered images. |
| Outcome: | The proposed benchmark includes 132 high-quality matplotlib plots across six plot types, as well as 150 and 86 plots from Python’s and R’s plotly libraries respectively, totaling 368 plots. |
Fraud-R1 : A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements (2025.findings-acl)
Copied to clipboard
Shu Yang, Shenzhe Zhu, Zeyu Wu, Keyu Wang, Junchi Yao, Junchao Wu, Lijie Hu, Mengdi Li, Derek F. Wong, Di Wang
| Challenge: | Existing fraud detection benchmarks focus on single-turn classification tasks, failing to capture dynamic nature of real-world fraud attempts. |
| Approach: | They propose a bilingual benchmark to assess LLMs' ability to resist fraud and phishing attacks across five key fraud categories: Fraudulent Services, Impersonation, Phishing Scams, Fake Job Postings, and Online Relationships. |
| Outcome: | The proposed model improves in role-play settings and in e-commerce and recommendation systems. |
Towards Objective Fine-tuning: How LLMs’ Prior Knowledge Causes Potential Poor Calibration? (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have enabled powerful domain-specific applications through supervised fine-tuning. |
| Approach: | They propose a cognition-aware framework that applies targeted learning strategies according to the model’s prior knowledge to improve calibration. |
| Outcome: | The proposed framework significantly improves calibration while maintaining performance, achieving an average 57% reduction in ECE compared to standard fine-tuning in Llama3-8B. |
RQT: Hierarchical Residual Quantization for Multi-Model Compression (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for decomposing fine-tuned LLMs are sensitive to the magnitude of delta values. |
| Approach: | They propose a hierarchical quantization framework that shares low-bit integer weights across similar models. |
| Outcome: | The proposed framework achieves an average accuracy degradation of approximately 3% on fine-tuned models across mathematics, coding, chatbot, and Chinese LLMs. |
CoV: Chain-of-View Prompting for Spatial Reasoning (2026.findings-acl)
Copied to clipboard
Haoyu Zhao, Akide Liu, Zeyu Zhang, Weijie Wang, Feng Chen, Ruihan Zhu, Gholamreza Haffari, Bohan Zhuang
| Challenge: | Embodied question answering requires collecting context that is distributed across multiple viewpoints . most recent vision–language models (VLMs) are constrained to a fixed and finite set of input views . |
| Approach: | They propose a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. |
| Outcome: | The proposed framework improves LLM-Match performance by 11.98% on four mainstream VLMs. |
LLaMA Pro: Progressive LLaMA with Block Expansion (2024.acl-long)
Copied to clipboard
| Challenge: | Existing studies have demonstrated that pre-trained LLMs are limited in certain domains, such as programming, mathematics, biomedical, or finance. |
| Approach: | They propose a new post-pretraining method with an expansion of Transformer blocks to tune the expanded blocks using only new corpus, efficiently and effectively improving the model’s knowledge while mitigating forgetting. |
| Outcome: | The proposed model outperforms existing models in programming and math and its instruction-following counterpart LLaMA Pro-8.3B in general tasks, programming, and mathematics. |
A Closer Look into Mixture-of-Experts in Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Mixture-of-experts (MoE) architectures are gaining increasing attention for their unique properties and remarkable performance. |
| Approach: | They propose a mixture-of-experts architecture that allows for model scaling without sacrificing computational efficiency. |
| Outcome: | The proposed model increases model size without sacrificing computational efficiency . the proposed model is modular and can be used by a broad spectrum of practitioners . |
Q-Mamba: Towards more efficient Mamba models via post-training quantization (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing studies show that Mamba architectures have room for further optimization in linear projections and state caches. |
| Approach: | They propose a decoupled scale quantization scheme to mitigate outliers in states and channels by applying separate quantization scales. |
| Outcome: | The proposed method reduces memory consumption by 50% across various quantization settings, model sizes, and generation and zero-shot tasks. |
Recommend for a Reason: Unlocking the Power of Unsupervised Aspect-Sentiment Co-Extraction (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing review-based recommenders favor large and complex language encoders that can only learn latent and uninterpretable text representations. |
| Approach: | They propose a tightly coupled two-stage approach to extract latent user sentiments and item properties from reviews and an Attention-Property-aware Rating Estimator (APRE). |
| Outcome: | Extensive experiments on seven real-world Amazon review datasets show that the proposed approach extracts the latent user sentiments, item properties, and the complicated interactions between the two components. |
KoCo-Bench: Can Large Language Models Leverage Domain Knowledge in Software Development? (2026.acl-long)
Copied to clipboard
Xue Jiang, Ge Li, Jiaru Qian, Xianjie Shi, Chenjie Li, Hao Zhu, Ziyu Wang, Jielun Zhang, Zeyu Zhao, Kechi Zhang, Jia Li, Wenpin Jiao, Zhi Jin, Yihong Dong
| Challenge: | Existing domain-specific code benchmarks focus on assessing what knowledge LLMs possess rather than how they acquire and apply new knowledge. |
| Approach: | They propose a benchmark to evaluate domain specialization methods in real-world software development. |
| Outcome: | KOCO-bench is a new benchmark for evaluating domain specialization methods in real-world software development. |
WenetSpeech-Wu: Datasets, Benchmarks, and Models for a Unified Chinese Wu Dialect Speech Processing Ecosystem (2026.findings-acl)
Copied to clipboard
Chengyou Wang, Mingchen Shao, Jingbin Hu, Zeyu Zhu, Hongfei Xue, Bingshen Mu, Xin Xu, Xingyi Duan, Binbin Zhang, Zhu Pengcheng, Chuang Ding, Xiaojun Zhang, Hui Bu, Lei Xie
| Challenge: | despite its linguistic significance, the Wu dialect of Chinese has long been hindered by the lack of large-scale speech data, standardized evaluation benchmarks, and publicly available models. |
| Approach: | They propose to use WenetSpeech-Wu as a large-scale, multi-dimensionally annotated open-source speech corpus for the Wu dialect of Chinese. |
| Outcome: | The proposed dataset includes 8,000 hours of speech data and strong open-source models . the proposed dataset is competitive and empirically validated . |
From Completion to Editing: Unlocking Context-Aware Code Infilling via Search-and-Replace Instruction Tuning (2026.acl-long)
Copied to clipboard
Jiajun Zhang, Zeyu Cui, Jiaxi Yang, Lei Zhang, Yuheng Jing, Zeyao Ma, Tianyi Bai, Zilei Wang, Qiang Liu, Liang Wang, Binyuan Hui, Junyang Lin
| Challenge: | Fill-in-the-Middle (FIM) models suffer from performance degradation and prohibitive latency. |
| Approach: | They propose a search-and-replace infilling framework that integrates agentic verification and editing into a single-pass inference process. |
| Outcome: | The proposed framework harmonizes completion tasks with the instruction-following priors of Chat LLMs, extending the paradigm from static infilling to dynamic context-aware editing. |
Every Document Owns Its Structure: Inductive Text Classification via Graph Neural Networks (2020.acl-main)
Copied to clipboard
| Challenge: | Existing graph-based methods for text classification cannot capture contextual word relationships within each document nor can they produce inductive learning of new words. |
| Approach: | They propose to use Graph Neural Networks to learn the local word representations and then aggregate the word nodes as the document embeddings. |
| Outcome: | The proposed method outperforms state-of-the-art methods on four benchmark datasets. |