Papers by Zhendong Mao

Zero-Shot Detection of LLM-Generated Text using Temperature Sensitivity (2026.acl-long)

Challenge:	Existing methods to detect AI-generated text rely on internal evidences, but external evidences are not considered.
Approach:	They propose a hierarchical graph network that utilizes internal and external factual structures to detect AI-generated text.
Outcome:	The proposed network outperforms current state-of-the-art methods on four datasets.

Copied to clipboard

Shixuan Ma, Jiahao Li, Zhendong Mao, Quan Wang

Text Style Transfer with Contrastive Transfer Pattern Mining (2023.acl-long)

Challenge:	Existing methods for detecting LLM-generated text rely on statistical features that are insufficient for reliable detection.
Approach:	They propose a temperature-sensitive detector that modulates decoding temperature and monitors how probability distributions respond to temperature.
Outcome:	The proposed method is based on a temperature sensitivity feature and a simple zero-shot detector built upon normalized temperature sensitivity.

Copied to clipboard

Jingxuan Han, Quan Wang, Licheng Zhang, Weidong Chen, Yan Song, Zhendong Mao

LIRE: listwise reward enhancement for preference alignment (2024.findings-acl)

Challenge:	Existing methods for text style transfer only focus on the transformation between styles, yet they do not take into account that this transformation can be achieved via different hidden transfer patterns.
Approach:	They propose a novel approach which automatically mines hidden transfer patterns to improve TST . they use a clustering module to automatically discover hidden transfer pattern from the data .
Outcome:	The proposed method can be applied in a plug-and-play manner to enhance other methods to further improve their performance.

Copied to clipboard

Mingye Zhu, Yi Liu, Lei Zhang, Junbo Guo, Zhendong Mao

Challenge:	prevailing approaches to preference alignment focus on pairwise comparisons, with limited exploration into multi-response scenarios.
Approach:	They propose a listwise reward enhancement approach that integrates offline rewards of multiple responses into a streamlined listwise framework.
Outcome:	The proposed approach outperforms existing methods on dialogue and summarization tasks with good transferability to out-of-distribution data.

Automated Creativity Evaluation for Large Language Models: A Reference-Based Approach (2025.findings-emnlp)

Copied to clipboard

Ruizhe Li, Chiwei Zhu, Benfeng Xu, Xiaorui Wang, Zhendong Mao

Curriculum Learning for Natural Language Understanding (2020.acl-main)

Challenge:	Existing methods for evaluating creativity of machine-generated texts rely on costly manual annotations or fail to align closely with human assessments.
Approach:	They propose an automated method based on the Torrance Test of Creative Writing (TTCW) .
Outcome:	The proposed method improves the alignment between LLM evaluations and human assessments.

Copied to clipboard

Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, Yongdong Zhang

S2ynRE: Two-stage Self-training with Synthetic data for Low-resource Relation Extraction (2023.acl-long)

Challenge:	Pre-trained language models can be fine tuned to perform NLU tasks in a straightforward manner.
Approach:	They propose a pretrain-finetune paradigm for natural language understanding (NLU) they propose 'a cross-trainset' approach that allows users to distinguish easy from difficult examples .
Outcome:	The proposed approach achieves significant performance improvements on a wide range of NLU tasks.

Copied to clipboard

Benfeng Xu, Quan Wang, Yajuan Lyu, Dai Dai, Yongdong Zhang, Zhendong Mao

Knowledge Context Modeling with Pre-trained Language Models for Contrastive Knowledge Graph Completion (2024.findings-acl)

Challenge:	Existing methods for relation extraction suffer from the inadequacy of large-scale annotated data.
Approach:	They propose a framework for two-stage self-training with synthetic data for relation extraction .
Outcome:	The proposed framework is based on two-stage self-training with synthetic data . it is able to synthesize large quantities of training data and iteratively and alternately learn from synthetic and golden data together.

Copied to clipboard

Guangqian Yang, Yi Liu, Lei Zhang, Licheng Zhang, Hongtao Xie, Zhendong Mao

Fine-grained Knowledge Enhancement for Retrieval-Augmented Generation (2025.findings-acl)

Challenge:	Text-based knowledge graph completion methods neglect knowledge contexts in inferring process.
Approach:	They propose a framework which models the knowledge context as additional prompts with pre-trained language models for knowledge graph completion.
Outcome:	The proposed framework achieves state-of-the-art on FB15k-237, WN18RR and Wikidata5M datasets.

Copied to clipboard

Jingxuan Han, Zhendong Mao, Yi Liu, Yexuan Che, Zheren Fu, Quan Wang

Information Retrieval

Challenge:	Existing studies rely on semantic similarity to retrieve knowledge but ignore fine-grained information within documents.
Approach:	They propose a fine-grained knowledge enhancement method to fill knowledge gaps with retrieved external information by a Chain-of-Thought prompting procedure and a decoding enhancement strategy to constrain the document-based decoding process.
Outcome:	The proposed method can be applied in a plug-and-play manner to enhance its performance with no additional modules or training process.

Random Entity Quantization for Parameter-Efficient Compositional Knowledge Graph Representation (2023.emnlp-main)

Copied to clipboard

Jiaang Li, Quan Wang, Yi Liu, Licheng Zhang, Zhendong Mao

Visual-Linguistic Dependency Encoding for Image-Text Retrieval (2024.lrec-main)

Challenge:	Existing approaches to learning on Knowledge Graphs (KGs) are not critical for learning on KGs.
Approach:	They propose an alternative approach to represent entities by composing entity-corresponding codewords matched from predefined small-scale codebooks.
Outcome:	The proposed approach achieves similar results to existing methods.

Copied to clipboard

Wenxin Guo, Lei Zhang, Kun Zhang, Yi Liu, Zhendong Mao

Multimodality

Challenge:	Existing approaches to image-text retrieval ignore semantic discrepancies caused by syntactic structure in natural language expressions and relationships among visual entities.
Approach:	They propose a visual-linguistic dependency encoder framework which explicitly models the dependency information among textual words and interaction patterns between image regions.
Outcome:	The proposed framework outperforms existing methods on a vision-linguistic compositional structure reasoning dataset.

UniRel: Unified Representation and Interaction for Joint Relational Triple Extraction (2022.emnlp-main)

Copied to clipboard

Wei Tang, Benfeng Xu, Yuyue Zhao, Zhendong Mao, Yifeng Liu, Yong Liao, Haiyong Xie

Align Documents to Questions: Question-Oriented Document Rewriting for Retrieval-Augmented Generation (2026.findings-acl)

Challenge:	Existing approaches to extract rich correlations between entities and relations are not fully exploited by existing methods.
Approach:	They propose to unify entities and relations by jointly encoding them within a concatenated natural language sequence and unify the modeling of interactions with a proposed Interaction Map.
Outcome:	The proposed method is more efficient and efficient than existing methods and can be scaled up to 2021.

Copied to clipboard

Jiaang Li, Zhendong Mao, Quan Wang, Yuning Wan, Yongdong Zhang

WildGraphBench: Benchmarking GraphRAG with Wild-Source Corpora (2026.findings-acl)

Challenge:	Retrieval-Augmented Generation (RAG) enhances the factuality of Large Language Models (LLMs) however, LLMs exhibit a stylistic bias when presented with mixed contexts, revealing a bottleneck in their utility.
Approach:	They propose a style-controlled rewriter that aligns retrieved documents with a question-oriented style while preserving facts.
Outcome:	The proposed model improves RAG pipelines by 8% with negligible latency overhead.

Copied to clipboard

Pengyu Wang, Benfeng Xu, Licheng Zhang, Shaohan Wang, Mingxuan Du, Chiwei Zhu, Zhendong Mao

Information Retrieval

Challenge:	Existing benchmarks for Graph-based Retrieval-Augmented Generation (GraphRAG) rely on short, curated passages as external knowledge, failing to adequately evaluate systems in realistic settings involving long contexts and large-scale heterogeneous documents.
Approach:	They propose a benchmark to assess GraphRAG performance in the wild using Wikipedia's unique structure where cohesive narratives are grounded in long and heterogeneous external reference documents.
Outcome:	Experiments with articles across 12 top-level topics show that GraphRAG performs better in the wild than existing methods.

From Real to Synthetic: Synthesizing Millions of Diversified and Complicated User Instructions with Attributed Grounding (2025.acl-long)

Copied to clipboard

Chiwei Zhu, Benfeng Xu, Xiaorui Wang, Zhendong Mao

Challenge:	a pursuit of diverse, complex, and large-scale instruction data is crucial for automatically aligning large language models . authors: methods that generate synthetic instructions at scale suffer from limited grounding sources . attributed grounding is a technique that can be used to align language models with human .
Approach:	They synthesize 1 million instructions using attributed grounding and a bottom-up synthesis process that leverages web documents to generate a situation, then a meaningful instruction.
Outcome:	The proposed framework achieves leading performance on benchmarks and scales with more web corpora.

FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents (2026.acl-long)

Copied to clipboard

Chiwei Zhu, Benfeng Xu, Mingxuan Du, Shaohan Wang, Xiaorui Wang, Zhendong Mao, Yongdong Zhang

EmRel: Joint Representation of Entities and Embedded Relations for Multi-triple Extraction (2022.naacl-main)

Challenge:	Long trajectories in deep research often exceed model context limits, compressing token budgets for both evidence collection and report writing.
Approach:	They propose a file-system-based framework that scales deep research beyond context window . a Context Builder agent acts as a librarian and a Report Writer agent composes the final report .
Outcome:	Experiments on two open-ended benchmarks show that FS-Researcher achieves state-of-the-art report quality across different backbone models.

Copied to clipboard

Benfeng Xu, Quan Wang, Yajuan Lyu, Yabing Shi, Yong Zhu, Jie Gao, Zhendong Mao

RESEMO: A Benchmark Chinese Dataset for Studying Responsive Emotion from Social Media Content (2024.findings-acl)

Challenge:	Existing studies only explore entity representations, but propose a novel triple perspective for relation extraction.
Approach:	They propose to explicitly introduce relation representation and jointly represent it with entities to identify valid triples.
Outcome:	The proposed method is based on ablations and document-level relation extraction and joint entity and relation extraction.

Copied to clipboard

Bo Hu, Meng Zhang, Chenfei Xie, Yuanhe Tian, Yan Song, Zhendong Mao

M-RangeDetector: Enhancing Generalization in Machine-Generated Text Detection through Multi-Range Attention Masks (2025.findings-acl)

Challenge:	Existing studies on social media text processing do not focus on responsive emotion analysis.
Approach:	They propose a Chinese dataset named ResEmo for responsive emotion analysis, including 3813 posts with 68,781 comments collected from Weibo, the largest social media platform in China.
Outcome:	The proposed dataset includes 3813 posts with 68,781 comments collected from weibo, the largest social media platform in China.

Copied to clipboard

Kaijie Jiao, Quan Wang, Licheng Zhang, Zikang Guo, Zhendong Mao

Mitigating Biases in Language Models via Bias Unlearning (2025.emnlp-main)

Challenge:	Existing supervised methods for text detection are overfitting within their training domains.
Approach:	They propose a method that integrates four distinct attention masking strategies into a Multi-Range Attention module to learn various writing strategies for machine-generated text detection.
Outcome:	The proposed method improves the generalization capability of existing detectors on three datasets.

Copied to clipboard

Dianqing Liu, Yi Liu, Guoqing Jin, Zhendong Mao

CodeRipple: Wavelet-Based Detection of LLM-Generated Code (2026.acl-long)

Challenge:	Recent debiasing approaches target different demographic groups, harming fairness and discrimination.
Approach:	They propose a model debiasing framework which targets stereotypes by unlearning stereotype forgetting and anti-stereotype retention.
Outcome:	The proposed framework outperforms existing methods in mitigating bias while retaining language modeling capabilities.

Copied to clipboard

Xingyu Yao, Zhendong Mao, Quan Wang

Programming Languages in NLP

Challenge:	Existing training-free detectors rely on global statistics of the Token Perplexity Sequence (TPS) and struggle with code.
Approach:	They propose a training-free detection framework that characterizes TPS morphology across scales.
Outcome:	The proposed framework outperforms existing training-free detectors on three challenging benchmarks spanning programming languages, multiple generating LLMs, and various evasion strategies.

KNN-Instruct: Automatic Instruction Construction with K Nearest Neighbor Deduction (2024.emnlp-main)

Copied to clipboard

Jianshang Kou, Benfeng Xu, Chiwei Zhu, Zhendong Mao

Reasoning

Challenge:	Existing methods for generating synthetic instructions for large language models suffer from stale distribution and scalability.
Approach:	They propose a method which incorporates KNN deduction to produce meaningful new instructions by summarizing and learning from existing ones.
Outcome:	The proposed method outperforms all 7B models on the LMSYS leaderboard.

Rationales Are Not Silver Bullets: Measuring the Impact of Rationales on Model Performance and Reliability (2025.findings-acl)

Copied to clipboard

Chiwei Zhu, Benfeng Xu, An Yang, Junyang Lin, Quan Wang, Chang Zhou, Zhendong Mao

Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation (2024.acl-long)

Challenge:	Existing studies have shown that training language models with rationales augmentation is beneficial, but this view does not hold consistently.
Approach:	They conduct comprehensive investigations to thoroughly inspect the impact of rationales on model performance and a novel perspective of model reliability.
Outcome:	The proposed method outperforms untrained models in several areas and provides informative regulations on the broad utilization of rationales.

Copied to clipboard

Tianqi Zhong, Zhaoyi Li, Quan Wang, Linqi Song, Ying Wei, Defu Lian, Zhendong Mao

On the Calibration of Large Language Models and Alignment (2023.findings-emnlp)

Challenge:	Existing MCTG methods face a noticeable performance drop in compositional testing.
Approach:	They propose a benchmark to evaluate compositional generalization of MCTG methods by combining multi-aspect labeled datasets and a crafted three-dimensional evaluation protocol.
Outcome:	The proposed framework improves compositional generalization performance by 3.64% and 94.4% in compositional testing.

Copied to clipboard

Chiwei Zhu, Benfeng Xu, Quan Wang, Yongdong Zhang, Zhendong Mao

Grammatical Error Correction via Mixed-Grained Weighted Training (2023.findings-emnlp)

Challenge:	Large language models are becoming more popular and are proving to be reliable . however, their reliability is often understudied due to their uncertainty and complex structure .
Approach:	They conduct a systematic examination of the calibration of aligned language models throughout the entire construction process including pretraining and alignment training.
Outcome:	The results shed light on whether popular large language models are well-calibrated and how the training process influences model calibration.

Copied to clipboard

Jiahao Li, Quan Wang, Chiwei Zhu, Zhendong Mao, Yongdong Zhang

Syntactic Text Processing

Challenge:	Empirical evaluation shows that MainGEC achieves consistent and significant performance improvements on two benchmark datasets.
Approach:	They propose to use mixed-grained weighted training to improve the training effect for GEC by analyzing the inherent discrepancies in annotated training data.
Outcome:	Empirical results show that the proposed method achieves significant performance improvements on two benchmark datasets.

Alleviating Hallucinations in Large Language Models via Truthfulness-driven Rank-adaptive LoRA (2025.findings-acl)

Copied to clipboard

Jiahao Li, Zhendong Mao, Quan Wang

Improving Chinese Spelling Check by Character Pronunciation Prediction: The Effects of Adaptivity and Granularity (2022.emnlp-main)

Challenge:	Existing methods to improve truthfulness are training-free without modifying the LLM itself.
Approach:	They propose a rank-adaptive LoRA method to improve LLM truthfulness that allocates ranks according to truthfulness correlations of LLM modules.
Outcome:	The proposed method outperforms state-of-the-art methods on the LLM family and makes the performance of 7B LLMs exceed GPT-4.

Copied to clipboard

Jiahao Li, Quan Wang, Zhendong Mao, Junbo Guo, Yanyan Yang, Yongdong Zhang

Syntactic Text Processing

Challenge:	Chinese spelling check (CSC) is a fundamental NLP task that detects and corrects spelling errors in Chinese texts.
Approach:	They propose an auxiliary task of Chinese pronunciation prediction to improve CSC . they propose adaptive weighting schemes and a delicate correction strategy .
Outcome:	The proposed auxiliary task improves Chinese pronunciation prediction on three benchmarks.

FlipGuard: Defending Preference Alignment against Update Regression with Constrained Optimization (2024.emnlp-main)

Copied to clipboard

Mingye Zhu, Yi Liu, Quan Wang, Junbo Guo, Zhendong Mao

Challenge:	Recent advances in preference alignment have significantly improved Large Language Models' ability to generate texts that align with human preferences and values.
Approach:	They propose a constrained optimization approach to detect and mitigate update regression with focal attention.
Outcome:	The proposed approach detects and mitigates update regression with focal attention while maintaining excellent overall performance.

Improve Safety Training of Large Language Models with Safety-Critical Singular Vectors Localization (2025.acl-long)

Copied to clipboard

Peijian Gu, Quan Wang, Zhendong Mao

Air-Decoding: Attribute Distribution Reconstruction for Decoding-Time Controllable Text Generation (2023.emnlp-main)

Challenge:	Recent work on safety training with modules such as low-rank adaptation (LoRA) to resist jailbreaks shows promise, but these approaches can inadvertently degrade a model’s general utility.
Approach:	They propose a plug-and-play method that locates safety-critical singular vectors within the model's parameter space and a dynamic rank number determination strategy to reduce parameter overhead.
Outcome:	The proposed method mitigates the impact of safety training on model utility by explicitly locating and leveraging safety-critical singular vectors within the model’s parameter space.

Copied to clipboard

Tianqi Zhong, Quan Wang, Jingxuan Han, Yongdong Zhang, Zhendong Mao