Papers by Wenxuan Zhou

A Causal View of Entity Bias in (Large) Language Models (2023.findings-emnlp)

Challenge:	Existing methods for detecting hallucination in long-form tasks focus on limited domains or rely heavily on external fact-checking tools, which may not always be available.
Approach:	They propose a new paradigm that augments fine-tuning with an auxiliary task for the model to jointly learn with the main task of hallucination detection.
Outcome:	The proposed method outperforms existing methods for detecting hallucination in open-domain long-form generation and is more accurate than random guessing.

Copied to clipboard

Fei Wang, Wenjie Mo, Yiwei Wang, Wenxuan Zhou, Muhao Chen

FocalPO: Enhancing Preference Optimizing by Focusing on Correct Preference Rankings (2025.acl-short)

Challenge:	Entity bias affects pretrained (large) language models, causing them to rely on (biased) parametric knowledge to make unfaithful predictions.
Approach:	They propose a structured causal model whose parameters are easier to estimate . they propose to perturb the original entity with neighboring entities .
Outcome:	The proposed model reduces biasing information pertaining to the original entity while still preserving sufficient semantic information from similar entities.

Copied to clipboard

Tong Liu, Xiao Yu, Wenxuan Zhou, Jindong Gu, Volker Tresp

Challenge:	Efficient preference optimization algorithms such as Direct Preference Optimization (DPO) have become a popular approach in aligning large language models with human preferences.
Approach:	They propose a preference optimization variant that instead down-weighs misranked preference pairs and prioritizes enhancing the model’s understanding of pairs that it can already rank correctly.
Outcome:	The proposed model outperforms DPO on benchmarks like Alpaca Eval 2.0 and Arena-Hard using mistral-base-7B and Llama-3-Instruct-8B with the introduced hyperparameter fixed.

Seeking Patterns, Not just Memorizing Procedures: Contrastive Learning for Solving Math Word Problems (2022.findings-acl)

Copied to clipboard

Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, Yunbo Cao

Semantic Text Processing

Challenge:	Existing models memorize procedures from context and rely on shallow heuristics to solve MWPs.
Approach:	They propose a contrastive learning approach where the neural network perceives the divergence of patterns.
Outcome:	The proposed method greatly improves performance in monolingual and multilingual settings.

An Improved Baseline for Sentence-level Relation Extraction (2022.aacl-short)

Copied to clipboard

Wenxuan Zhou, Muhao Chen

ThinkGuard: Deliberative Slow Thinking Leads to Cautious Guardrails (2025.findings-acl)

Challenge:	Sentence-level relation extraction (RE) aims at identifying the relationship between two entities in a sentence.
Approach:	They propose to improve sentence-level relation extraction by adding entity representations with typed markers to the model.
Outcome:	The proposed model outperforms existing methods on entity representation and noisy labels on TACRED dataset.

Copied to clipboard

Xiaofei Wen, Wenxuan Zhou, Wenjie Jacky Mo, Muhao Chen

Text Classification

Challenge:	Existing guardrails rely on rule-based filtering or single-pass classification, limiting their ability to handle nuanced safety violations.
Approach:	They propose a critique-augmented guardrail model that distills knowledge from high-capacity LLMs by generating structured critiques alongside safety labels.
Outcome:	The proposed model outperforms existing guardrail models on multiple safety benchmarks and achieves the highest average F1 and AUPRC.

MetaScale: Test-Time Scaling with Evolving Meta-Thoughts (2026.findings-acl)

Copied to clipboard

Qin Liu, Wenxuan Zhou, Nan Xu, James Y. Huang, Fei Wang, Sheng Zhang, Hoifung Poon, Muhao Chen

CrowdSelect: SyntheticInstruction Data Selection with Multi-LLM Wisdom (2026.findings-eacl)

Challenge:	Existing approaches impose fixed cognitive structures that enhance performance in specific tasks but lack adaptability across diverse scenarios.
Approach:	They propose a test-time scaling framework based on meta-thoughts to improve performance . meta-thinkts are adaptive thinking strategies tailored to a given task .
Outcome:	Experimental results show that MetaScale outperforms standard inference approaches . it can scale more effectively with increasing sampling budgets and produces more structured responses .

Copied to clipboard

Yisen Li, Lingfeng Yang, Wenxuan Shen, Pan Zhou, Yao Wan, Weiwei Lin, Dongping Chen

Machine Translation

Challenge:	Existing methods for capturing instruction-following complexity rely on single-dimensional signals, but they fail to capture complexity across diverse fields.
Approach:	They propose three foundational metrics that leverage Multi-LLMs wisdom to capture instruction-response pair characteristics and propose CrowdSelect, an integrated metric incorporating a clustering-based approach to maintain response diversity.
Outcome:	The proposed metrics outperform existing models on MT-bench and Arena-hard and show improvements of 4.81% on full and LoRA fine-tuning.

Pardon? Evaluating Conversational Repair in Large Audio-Language Models (2026.findings-acl)

Copied to clipboard

Shuanghong Huang, Jinlei Xu, Youchao Zhou, Yanghao Zhou, Xuan Zhao, Chong Feng, Wenxuan Zhang

Natural Language Interfaces

Challenge:	Existing evaluations of large audio-language models focus on answer accuracy and robustness to acoustic perturbations, but they assume that inputs remain semantically answerable.
Approach:	They propose a repair-aware evaluation setting that explicitly distinguishes between answerable and unanswerable audio inputs.
Outcome:	The proposed evaluation setting distinguishes between answerable and unanswerable audio inputs.

Multi-hop Evidence Retrieval for Cross-document Relation Extraction (2023.findings-acl)

Copied to clipboard

Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, Muhao Chen

Should We Rely on Entity Mentions for Relation Extraction? Debiasing Relation Extraction with Counterfactual Analysis (2022.naacl-main)

Challenge:	Relation Extraction (RE) is a task that seeks to identify the relation of entities described according to some context.
Approach:	They propose a multi-hop evidence retrieval method based on evidence path mining and ranking to support cross-document relation extraction.
Outcome:	The proposed method acquires cross-document evidence and boosts performance in both closed and open environments.

Copied to clipboard

Yiwei Wang, Muhao Chen, Wenxuan Zhou, Yujun Cai, Yuxuan Liang, Dayiheng Liu, Baosong Yang, Juncheng Liu, Bryan Hooi

Continual Contrastive Finetuning Improves Low-Resource Relation Extraction (2023.acl-long)

Challenge:	Existing studies rely on entity information for sentence-level relation extraction (RE) but this can leak superficial and spurious clues of relations.
Approach:	They propose to use entity mentions to extract relations from textual context . they use a causal graph to model dependencies between variables in RE models .
Outcome:	The proposed method yields significant gains on both effectiveness and generalization for RE.

Copied to clipboard

Wenxuan Zhou, Sheng Zhang, Tristan Naumann, Muhao Chen, Hoifung Poon

Prix-LM: Pretraining for Multilingual Knowledge Base Construction (2022.acl-long)

Challenge:	Relation extraction (RE) has been challenging in low-resource domains and with limited resources.
Approach:	They propose to pretrain and finetune the RE model using consistent objectives of contrastive learning.
Outcome:	The proposed method outperforms PLM-based RE classifier on two document-level RE datasets.

Copied to clipboard

Wenxuan Zhou, Fangyu Liu, Ivan Vulić, Nigel Collier, Muhao Chen

Semantic Text Processing

Challenge:	Existing methods to build and enrich multilingual knowledge bases have not been successful . knowledge expressed in different languages may be complementary and unequally distributed .
Approach:	They propose a model that integrates useful multilingual and KB-based factual knowledge into a single model.
Outcome:	The proposed model can provide richer combined knowledge than monolingual KBs.

Code Execution as Grounded Supervision for LLM Reasoning (2025.emnlp-main)

Copied to clipboard

Dongwon Jung, Wenxuan Zhou, Muhao Chen

mDPO: Conditional Preference Optimization for Multimodal Large Language Models (2024.emnlp-main)

Challenge:	Existing methods for generating high-quality CoT data rely on costly human annotations and error-prone CoT.
Approach:	They propose a method that extracts verifiable, step-by-step reasoning traces from code execution and transforms them into a natural language CoT reasoning.
Outcome:	The proposed method produces highly accurate reasoning data and reduces overall token length during inference by reducing meaningless repetition and overthinking.

Copied to clipboard

Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, Muhao Chen

Multimodality

Challenge:	Recent studies have attempted to apply DPO to multimodal scenarios but have found it challenging to achieve consistent improvement.
Approach:	They propose a multimodal DPO objective that prevents the over-prioritization of language-only preferences by also optimizing image preference.
Outcome:	The proposed method significantly improves performance on two multimodal LLMs of different sizes and three widely used benchmarks.

Improving Model Factuality with Fine-grained Critique-based Evaluator (2025.acl-long)

Copied to clipboard

Yiqing Xie, Wenxuan Zhou, Pradyot Prakash, Di Jin, Yuning Mao, Quintin Fettes, Arya Talebzadeh, Sinong Wang, Han Fang, Carolyn Rose, Daniel Fried, Hejia Zhang

SFTMix: Elevating Language Model Instruction Tuning with Mixup Recipe (2026.acl-long)

Challenge:	Factuality evaluation aims to detect factual errors produced by language models and guide the development of more factual models.
Approach:	They propose a framework that leverages FenCE to improve the factuality of LM generators by constructing training data.
Outcome:	The proposed framework improves the factuality of LM generators by enhancing their training data.

Copied to clipboard

Yuxin Xiao, Shujian Zhang, Marzyeh Ghassemi, Wenxuan Zhou

Summarization as Indirect Supervision for Relation Extraction (2022.findings-emnlp)

Challenge:	Efforts to improve instruction tuning often focus on higher-quality supervised fine-tuning datasets, typically requiring data filtering with proprietary LLMs or human annotation.
Approach:	They propose a Mixup-based recipe that elevates LLM instruction tuning without relying on well-curated datasets.
Outcome:	The proposed model improves instruction-following and healthcare-specific tasks with consistent improvements across LLM families and SFT datasets.

Copied to clipboard

Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, Muhao Chen

WPO: Enhancing RLHF with Weighted Preference Optimization (2024.emnlp-main)

Challenge:	Relation extraction (RE) models rely on training data with expensive annotations . et al., 2018; Zhao e.t al, 2018) .
Approach:	They propose a method that converts RE into a summarization formulation by using constraint decoding techniques.
Outcome:	The proposed method improves relation extraction models with high-resource and high-contrast inferences.

Copied to clipboard

Wenxuan Zhou, Ravi Agrawal, Shujian Zhang, Sathish Reddy Indurthi, Sanqiang Zhao, Kaiqiang Song, Silei Xu, Chenguang Zhu

Self-regulation: Employing a Generative Adversarial Network to Improve Event Detection (P18-1)

Challenge:	Off-policy preference optimization suffers from a distributional gap between the policy used for data collection and the target policy, leading to suboptimal optimization.
Approach:	They propose a method to simulate on-policy learning with off-police preference data.
Outcome:	The proposed method outperforms Direct Preference Optimization (DPO) by up to 5.6% on Alpaca Eval 2 and MT-bench.

Copied to clipboard

Yu Hong, Wenxuan Zhou, Jingli Zhang, Guodong Zhou, Qiaoming Zhu

Robustness in NLP

Challenge:	Recent studies show that neural networks can be used for event detection but can be contaminated by spurious features.
Approach:	They propose a self-regulated learning approach by utilizing a generative adversarial network to generate spurious features.
Outcome:	The proposed method is highly effective and adaptable on the ACE 2005 and TAC-KBP 2015 corpora.

Contrastive Instruction Tuning (2024.findings-acl)

Copied to clipboard

Tianyi Yan, Fei Wang, James Y. Huang, Wenxuan Zhou, Fan Yin, Aram Galstyan, Wenpeng Yin, Muhao Chen

Answer Consolidation: Formulation and Benchmarking (2022.naacl-main)

Challenge:	Current LLMs exhibit limited robustness to unseen instructions, generating inconsistent outputs when the same instruction is phrased with slightly varied forms or language styles.
Approach:	They propose a method which maximizes the similarity between the hidden representations of semantically equivalent instruction-instance pairs while minimizing the similarities between semantically different ones.
Outcome:	Experiments on the PromptBench benchmark show that Contrastive Instruction Tuning improves LLMs’ robustness to unseen instructions with variations across character, word, sentence, and semantic levels by +2.5% in accuracy.

Copied to clipboard

Wenxuan Zhou, Qiang Ning, Heba Elfardy, Kevin Small, Muhao Chen

Natural Language Interfaces

Challenge:	Current question answering systems assume each question to have one correct answer.
Approach:	They propose a problem where answers are partitioned into multiple groups . they construct a comprehensive and non-redundant set of answers by picking one answer from each group .
Outcome:	The proposed model performs better than previous models, but it needs further improvements.

Parameter-Efficient Tuning with Special Token Adaptation (2023.eacl-main)

Copied to clipboard

Xiaocong Yang, James Y. Huang, Wenxuan Zhou, Muhao Chen

T-REG: Preference Optimization with Token-Level Reward Regularization (2025.acl-long)

Challenge:	a recent study shows that parameter-efficient tuning is a challenge for multitask deployments.
Approach:	They propose a parameter-efficient tuning technique that only updates a small subset of parameters when adapting a pretrained model to downstream tasks.
Outcome:	The proposed method achieves comparable performance to fine-tuning in natural language understanding tasks including text classification and NER with only 0.029% of parameters trained.

Copied to clipboard

Wenxuan Zhou, Shujian Zhang, Lingxiao Zhao, Tao Meng

GeoLM: Empowering Language Models for Geospatially Grounded Language Understanding (2023.emnlp-main)

Challenge:	Reinforcement learning from human feedback (RLHF) is a dominant approach for large language models to follow instructions and produce meaningful alignment.
Approach:	They propose a method that leverages human feedback to optimize large language models . they propose to use sequence-level and token-level rewards to optimize preference .
Outcome:	The proposed method outperforms baseline methods on Alpaca Eval 2 and Arena-Hard benchmarks.

Copied to clipboard

Zekun Li, Wenxuan Zhou, Yao-Yi Chiang, Muhao Chen

Getting Sick After Seeing a Doctor? Diagnosing and Mitigating Knowledge Conflicts in Event Temporal Reasoning (2024.findings-naacl)

Challenge:	Pretrained language models do not utilize valuable geospatial information in large databases, e.g., OpenStreetMap.
Approach:	They propose a geospatially grounded language model that connects linguistic and geospheric contexts.
Outcome:	The proposed model bridges the gap between natural language processing and geospatial sciences.

Copied to clipboard

Tianqing Fang, Zhaowei Wang, Wenxuan Zhou, Hongming Zhang, Yangqiu Song, Muhao Chen

Contrastive Out-of-Distribution Detection for Pretrained Transformers (2021.emnlp-main)

Challenge:	Event temporal reasoning aims at identifying the temporal relations between two or more events from narratives.
Approach:	They propose to detect knowledge conflicts in event temporal reasoning using bias indicators such as event relation prior bias, tense bias, narrative bias, and dependency bias.
Outcome:	The proposed method can be applied to Pre-trained Language Models and Large Language Model (LLMs) as additional training data or demonstrations for In- Context Learning.

Copied to clipboard

Wenxuan Zhou, Fangyu Liu, Muhao Chen

Sharpness-Aware Minimization with Dynamic Reweighting (2022.findings-emnlp)

Challenge:	Pretrained Transformers achieve remarkable performance when training and test data are from the same distribution, but in real-world scenarios, out-of-distribution instances can cause semantic shift problems.
Approach:	They propose to fine-tune the Transformers with a contrastive loss, which improves the compactness of representations, and to use the Mahalanobis distance in the model's penultimate layer to detect OOD instances.
Outcome:	The proposed method outperforms baselines in the real-world and achieves near-perfect OOD detection performance.

Copied to clipboard

Wenxuan Zhou, Fangyu Liu, Huan Zhang, Muhao Chen

Robust Natural Language Understanding with Residual Attention Debiasing (2023.findings-acl)

Challenge:	Deep neural networks are often overparameterized and can overfit training data.
Approach:	They propose an adversarial weight minimization algorithm that conducts adversarials and finds a common adversaria per-batch.
Outcome:	The proposed algorithm finds a common adversarial weight perturbation per-batch.

Copied to clipboard

Fei Wang, James Y. Huang, Tianyi Yan, Wenxuan Zhou, Muhao Chen

Robustness in NLP

Challenge:	Existing ensemble-based debiasing methods do not address unintended dataset biases . attention plays a crucial role in providing robust prediction in NLU models .
Approach:	They propose an end-to-end debiasing method that mitigates unintended biases from attention.
Outcome:	The proposed method improves the OOD performance of BERT-based models on three benchmarks.

EngiBench: A Benchmark for Evaluating Large Language Models on Engineering Problem Solving (2026.findings-acl)

Copied to clipboard

Xiyuan Zhou, Xinlei Wang, Yirui He, Ruixi Zou, Yang Wu, Yuheng Cheng, Yulu Xie, Wenxuan Liu, Huan Zhao, Yan Xu, Jinjin Gu, Junhua Zhao

False Sense of Security: Why Probing-based Malicious Input Detection Fails to Generalize (2026.findings-acl)

Challenge:	Existing benchmarks focus on well-defined or abstract reasoning and fail to capture real-world engineering problems.
Approach:	They propose a hierarchical benchmark to evaluate large language models on engineering problems.
Outcome:	The proposed model performs well under well-defined conditions and is based on three levels of difficulty and covers diverse engineering subfields.

Copied to clipboard

Cheng Wang, Zeming Wei, Qin Liu, Wenxuan Zhou, Muhao Chen

Improving Multilingual Instruction Finetuning via Linguistically Natural and Diverse Datasets (2024.findings-emnlp)

Challenge:	Recent work has leveraged probing-based approaches to study the separability of malicious and benign inputs in Large Language Models’ internal representations.
Approach:	They propose to use probing-based methods to study separability of malicious and benign inputs in LLMs' internal representations to detect harmful and benign content.
Outcome:	The proposed methods show that they learn superficial patterns rather than semantic harmfulness.

Copied to clipboard

Sathish Reddy Indurthi, Wenxuan Zhou, Shamil Chollampatt, Ravi Agrawal, Kaiqiang Song, Lingxiao Zhao, Chenguang Zhu

Multilinguality

Challenge:	Advancements in Large Language Models (LLMs) have significantly enhanced instruction-following capabilities, but most IFT datasets are predominantly in English, limiting model performance in other languages.
Approach:	They propose a method for collecting multilingual IFT datasets that preserves linguistic naturalness and ensures prompt diversity.
Outcome:	Experiments show that LLMs fine-tuned using this method show significant improvements in generative and discriminative tasks.

On-the-fly Denoising for Data Augmentation in Natural Language Understanding (2024.findings-eacl)

Copied to clipboard

Tianqing Fang, Wenxuan Zhou, Fangyu Liu, Hongming Zhang, Yangqiu Song, Muhao Chen

Low-Resource NLP

Challenge:	Existing methods to improve data augmentation performance may introduce noisy data that impairs training.
Approach:	They propose an on-the-fly denoising technique that learns from soft augmented labels provided by an organic teacher model trained on the cleaner original dataset.
Outcome:	The proposed method improves on text classification and question-answering tasks on general augmentation techniques and prevents overfitting on noisy labels.

Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting (2026.acl-short)

Copied to clipboard

Cheng Wang, Qin Liu, Wenxuan Zhou, Muhao Chen

Context-faithful Prompting for Large Language Models (2023.findings-emnlp)

Challenge:	Excessive exploitation can cause the model to become overconfident in its suboptimal solutions, thereby limiting its capabilities to explore novel reasoning strategies.
Approach:	They propose a method that dynamically down-weights extreme token-level updates via a Gaussian kernel and reduces the instability caused by the trade-off.
Outcome:	The proposed method improves downstream performance across reasoning benchmarks and stabilizes entropy as training progresses.

Copied to clipboard

Wenxuan Zhou, Sheng Zhang, Hoifung Poon, Muhao Chen

Learning from Noisy Labels for Entity-Centric Information Extraction (2021.emnlp-main)

Challenge:	Large language models encode parametric knowledge about world facts but overly rely on it can cause incorrect predictions in context-sensitive NLP tasks.
Approach:	They propose to use opinion-based prompts and counterfactual demonstrations to improve LLM faithfulness to contexts.
Outcome:	The proposed methods improve faithfulness to contexts using opinion-based prompts and counterfactual demonstrations.

Copied to clipboard

Wenxuan Zhou, Muhao Chen