Papers by Yunfang Wu

Ungrammatical-syntax-based In-context Example Selection for Grammatical Error Correction (2024.naacl-long)

Challenge:	Existing question generation models treat input passage as a sequence-to-sequence generative task, but they are not aware of text structure.
Approach:	They propose to model text structure as answer position and syntactic dependency and propose a mask attention mechanism to make syntaktic structure of input passage accessible.
Outcome:	The proposed model outperforms the strong pre-trained model ProphetNet on a SQuAD dataset and achieves competitive results with the state-of-the-art model.

Copied to clipboard

Chenming Tang, Fanyi Qu, Yunfang Wu

Syntactic Text Processing

Challenge:	In-context learning (ICL) has shown impressive results on many tasks, but applying LLMs to grammatical error correction (GEC) is still a challenging task.
Approach:	They propose an ungrammatical-syntax-based in-context example selection strategy that measures similarity of sentences based on their syntactic structures and identify optimal ICL examples sharing the most similar ill-formed syntax to the test input.
Outcome:	The proposed strategy outperforms word-matching and semantics-based methods on a syntax-oriented task like GEC on benchmark English datasets.

Learning Universal Sentence Representations with Mean-Max Attention Autoencoder (D18-1)

Copied to clipboard

Minghua Zhang, Yunfang Wu, Weikang Li, Wei Li

How to Ask Good Questions? Try to Leverage Paraphrases (2020.acl-main)

Challenge:	Existing methods to learn universal sentence representations focus on supervised learning.
Approach:	They propose a mean-max attention autoencoder that uses a multi-head mechanism to reconstruct the input sequence.
Outcome:	The proposed model outperforms state-of-the-art unsupervised single methods on a wide range of 10 transfer tasks.

Copied to clipboard

Xin Jia, Wenjie Zhou, Xu Sun, Yunfang Wu

Text Generation

Challenge:	Existing methods to generate human-like questions rely on paraphrases to generate good questions.
Approach:	They propose to integrate paraphrase knowledge into question generation to generate human-like questions by combining paraphrases with a back-translation method.
Outcome:	The proposed model achieves obvious performance gain over several strong baselines and human evaluation validates that it can ask questions of high quality by leveraging paraphrase knowledge.

Unleashing Large Language Models’ Proficiency in Zero-shot Essay Scoring (2024.findings-emnlp)

Copied to clipboard

Sanwoo Lee, Yida Cai, Desong Meng, Ziyang Wang, Yunfang Wu

3DM: Distill, Dynamic Drop, and Merge for Debiasing Multi-modal Large Language Models (2025.findings-acl)

Challenge:	Recent advances in automated essay scoring (AES) have relied on labeled essays, requiring tremendous cost and expertise for their acquisition.
Approach:	They propose a zero-shot prompting framework that automatically decomposes writing proficiency into distinct traits and generates scoring criteria for each trait.
Outcome:	The proposed framework outperforms straightforward prompting (Vanilla) on TOEFL11 and ASAP, while the small-sized Llama2-13b-chat significantly outperformed ChatGPT.

Copied to clipboard

Zhaoxi Zhang, Sanwoo Lee, Zhixiang Wang, Yunfang Wu

Position Offset Label Prediction for Grammatical Error Correction (2022.coling-1)

Challenge:	Recent advances in Multi-modal Language Models have shown remarkable performance in multimodal tasks . however, these models often exhibit inherent biases that compromise their reliability and fairness.
Approach:	They propose a framework that integrates Distill, Dynamic Drop, and Merge to address these challenges.
Outcome:	The proposed framework outperforms existing methods in balancing debiasing and improving performance on the MMSD2.0 sarcasm detection dataset.

Copied to clipboard

Xiuyu Wu, Jingsong Yu, Xu Sun, Yunfang Wu

Syntactic Text Processing

Challenge:	Experimental results show that our proposed POL-Pc framework improves baseline models and yields consistent performance gain over various data augmentation methods.
Approach:	They propose a position offset label prediction subtask to integrate correction editing operations into a unified framework.
Outcome:	The proposed model outperforms baseline models on Chinese, English and Japanese datasets by a wide margin.

Beyond Spurious Signals: Debiasing Multimodal Large Language Models via Counterfactual Inference and Adaptive Expert Routing (2025.findings-emnlp)

Copied to clipboard

Zichen Wu, Hsiu-Yuan Huang, Yunfang Wu

Robustness in NLP

Challenge:	Multimodal Large Language Models (MLLMs) often rely on spurious correlations, undermining their robustness and generalization.
Approach:	They propose a causal mediation-based debiasing framework to address correlation bias in MLLMs . they distinguish core semantics from spurious textual and visual contexts using counterfactual examples .
Outcome:	The proposed framework surpasses existing state-of-the-art models on sarcasm detection and sentiment analysis tasks.

Think Outside the Policy: In-Context Steered Policy Optimization (2026.findings-acl)

Copied to clipboard

Hsiu-Yuan Huang, Chenming Tang, Weijie Liu, Clive Bai, Saiyong Yang, Yunfang Wu

Reasoning

Challenge:	Existing Reinforcement Learning from Verifiable Rewards (RLVR) methods exhibit limited exploration due to reliance on on-policy rollouts which are limited to the current policy’s distribution, resulting in narrow trajectory diversity.
Approach:	They propose a framework that leverages the in-context learning capability of Large Reasoning Models to provide expert guidance using existing datasets.
Outcome:	The proposed framework improves RLVR performance and training stability on mathematical reasoning benchmarks.

Question-type Driven Question Generation (D19-1)

Copied to clipboard

Wenjie Zhou, Minghua Zhang, Yunfang Wu

Text Generation

Challenge:	Existing work suffers from mismatching between question type and answer . existing work fails to generate questions with type how while answer is personal name .
Approach:	They propose to automatically predict the question type based on the input answer and context.
Outcome:	The proposed model improves on both SQuAD and MARCO datasets and improves accuracy on the input answer and context.

Focus-Driven Contrastive Learning for Medical Question Summarization (2022.coling-1)

Copied to clipboard

Ming Zhang, Shuai Dou, Ziyang Wang, Yunfang Wu

Mixture-of-Prompt-Experts for Multi-modal Semantic Understanding (2024.lrec-main)

Challenge:	Existing methods to summarize health questions are not able to capture well question focus and lack the ability to understand sentence-level semantics.
Approach:	They propose a question focus-driven contrastive learning framework to capture question focus and exploit contrastive training at both encoder and decoder to obtain better sentence representations.
Outcome:	The proposed model achieves 5.33, 12.85 and 3.81 points over the baseline model on three medical benchmark datasets.

Copied to clipboard

Zichen Wu, Hsiu-Yuan Huang, Fanyi Qu, Yunfang Wu

From Spelling to Grammar: A New Framework for Chinese Grammatical Error Correction (2022.findings-emnlp)

Challenge:	Multimodal semantic understanding is crucial for developing machines capable of interpreting complex interplay of text and visual information.
Approach:	They propose a multi-modal soft prompt framework that integrates three experts of soft prompts . they propose sarcasm detection and sentiment analysis tasks that are critical for few-shot learning .
Outcome:	The proposed model outperforms the 8.2B model InstructBLIP with 2% parameters . it significantly outperformed other prompt methods on VLMs or task-specific methods .

Copied to clipboard

Xiuyu Wu, Yunfang Wu

Syntactic Text Processing

Challenge:	Chinese Grammatical Error Correction (CGEC) aims to generate correct sentences from erroneous sequences.
Approach:	They propose a zero-shot approach for spelling error correction that is simple but effective . they propose auxiliary task to predict POS sequence of target sentence .
Outcome:	The proposed framework achieves 42.11 F-0.5 on the English GEC dataset outperforms the previous state-of-the-art by a wide margin of 1.30 points.

Do Not Step Into the Same River Twice: Learning to Reason from Trial and Error (2026.acl-long)

Copied to clipboard

Chenming Tang, Hsiu-Yuan Huang, Weijie Liu, Clive Bai, Saiyong Yang, Yunfang Wu

Unsupervised Distractor Generation via Large Language Model Distilling and Counterfactual Contrastive Decoding (2024.findings-acl)

Challenge:	Existing approaches to RLVR train LMs based on their own on-policy responses and are constrained by the initial capability of LM.
Approach:	They propose an approach that hints LMs with their self-made mistakes without external guidance.
Outcome:	The proposed approach outperforms the normal group relative policy optimization and requires no external guidance.

Copied to clipboard

Fanyi Qu, Hao Sun, Yunfang Wu

Aligning Language Models with Real-time Knowledge Editing (2026.acl-long)

Challenge:	Recent studies show that large-scale models can generate unsupervised DG without expensive distractor annotations.
Approach:	They propose a dual task training framework that integrates pseudo distractors from LLMs and answer information as the objective target with a two-stage training process.
Outcome:	The proposed method surpasses GPT-3.5-turbo zero-shot performance with 200 fewer model parameters.

Copied to clipboard

Chenming Tang, Yutong Yang, Kexue Wang, Yunfang Wu

FPT: Feature Prompt Tuning for Few-shot Readability Assessment (2024.naacl-long)

Challenge:	Mainstream knowledge editing methods are static and fail to keep pace with the evolving real-world knowledge.
Approach:	They propose a new paradigm for knowledge editing that integrates edit augmentation and self-adaptive post-alignment inference into CRAFT to improve edit success.
Outcome:	The proposed method shows significant performance gain on CRAFT and traditional datasets compared to existing methods.

Copied to clipboard

Ziyang Wang, Sanwoo Lee, Hsiu-Yuan Huang, Yunfang Wu

Are Pre-trained Language Models Useful for Model Ensemble in Chinese Grammatical Error Correction? (2023.acl-short)

Challenge:	Prompt-based methods lack crucial linguistic knowledge for readability assessment tasks such as word length, sentence length, and usage of different difficulty-level words.
Approach:	They propose a new prompt-based tuning framework that incorporates linguistic knowledge and a loss function to calibrate the similarity ranking order between categories.
Outcome:	The proposed framework outperforms the large language model gpt-3.5-turbo-16k in most cases.

Copied to clipboard

Chenming Tang, Xiuyu Wu, Yunfang Wu

Composable Cross-prompt Essay Scoring by Merging Models (2025.emnlp-main)

Challenge:	Model ensemble has been in widespread use for Grammatical Error Correction (GEC), boosting model performance.
Approach:	They propose to use model ensembles computed by pre-trained language models to improve model performance.
Outcome:	The proposed ensembles do not improve but get worse after the PLM-based ensemble.

Copied to clipboard

Sanwoo Lee, Kun Liang, Yunfang Wu

CTR-Guided Generative Query Suggestion in Conversational Search (2025.emnlp-industry)

Challenge:	Existing approaches to cross-prompt automated essay scoring use all available sources . however, using multiple sources for continual adaptation raises privacy concerns .
Approach:	They propose a source-free adaptation approach that selectively merges the parameters of individual models without further access to the source datasets.
Outcome:	The proposed method outperforms joint-training methods on all sources while maintaining computational efficiency.

Copied to clipboard

Erxue Min, Hsiu-Yuan Huang, Xihong Yang, Min Yang, Xin Jia, Yunfang Wu, Hengyi Cai, Junfeng Wang, Shuaiqiang Wang, Dawei Yin

Natural Language Interfaces

Challenge:	Generating effective query suggestions requires aligning model outputs with user click preferences.
Approach:	They propose a generative framework that leverages click modeling to denoise implicit feedback and enables reliable preference optimization for improving real-world user engagement.
Outcome:	The proposed framework outperforms strong baselines in CTR, relevance, diversity and diversity.

Dynamic Fisher-weighted Model Merging via Bayesian Optimization (2025.naacl-long)

Copied to clipboard

Sanwoo Lee, Jiahao Liu, Qifan Wang, Jingang Wang, Xunliang Cai, Yunfang Wu

Asking Questions Like Educational Experts: Automatically Generating Question-Answer Pairs on Real-World Examination Data (2021.emnlp-main)

Challenge:	Existing merging approaches involve scaling the parameters model-wise or integrating parameter importance parameter-wise.
Approach:	They propose a method for merging model-based models at the parameter level without training data or joint training.
Outcome:	The proposed model merging framework outperforms baseline models on validation sets.

Copied to clipboard

Fanyi Qu, Xin Jia, Yunfang Wu

Challenge:	Existing approaches to generate high quality question-answer pairs are limited . a new framework is proposed for the question-answer generation task on real-world examination data.
Approach:	They propose a multi-agent communication model to generate and optimize the question and keyphrases iteratively and then apply the generated question and keys to guide the generation of answers.
Outcome:	The proposed framework makes great breakthroughs in the question-answer pair generation task.

SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation (2026.findings-acl)

Copied to clipboard

Gengyang Li, Wang Cai, Yifeng Gao, Yunfang Wu

Reasoning

Challenge:	Large language models (LLMs) achieve strong reasoning with Chain-of-Thought prompting, but long and redundant traces substantially increase inference cost.
Approach:	They propose a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights.
Outcome:	Experiments on GSM8K, MMLU, GPQA, and BBH show that SyncThink achieves 62.00% average Top@1 accuracy using 656 generated tokens and 28.68s latency, compared to 61.22%, 2141 tokens, and 92.01s for full CoT decoding.

Beyond Demonstrations: Dynamic Vector Construction from Latent Representations (2025.emnlp-main)

Copied to clipboard

Wang Cai, Hsiu-Yuan Huang, Zhixiang Wang, Yunfang Wu

Large Language Models Might Not Care What You Are Saying: Prompt Format Beats Descriptions (2025.findings-emnlp)

Challenge:	Existing methods for In-Context Learning (ICL) are sensitive to ICL-specific factors and rely on heuristic-based injection positions.
Approach:	They propose a method that extracts task-relevant representations from large language models and reinjects them during inference.
Outcome:	The proposed method outperforms few-shot In-Context Learning (ICL) and LoRA methods without repeated demonstration processing.

Copied to clipboard

Chenming Tang, Zhixiang Wang, Hao Sun, Yunfang Wu

A Unified Neural Network Model for Readability Assessment with Feature Projection and Length-Balanced Loss (2022.emnlp-main)

Challenge:	In-context learning has improved performance of large language models, but descriptive instructions are still under-explored.
Approach:	They propose an ensemble prompt framework to describe selection criteria of multiple in-context examples. preliminary experiments on machine translation confirm that this framework boosts ICL performance.
Outcome:	The proposed framework improves on commonsense, math, logical reasoning and hallucination tasks with three LLMs.

Copied to clipboard

Wenbiao Li, Wang Ziyang, Yunfang Wu

SCOI: Syntax-augmented Coverage-based In-context Example Selection for Machine Translation (2024.emnlp-main)

Challenge:	Traditional readability assessment models employ hundreds of linguistic features, but it is less explored for readability assessments.
Approach:	They propose a BERT-based model with feature projection and length-balanced loss to determine the difficulty level of a given text.
Outcome:	The proposed model achieves significant improvements over baseline models on three English benchmark datasets and one Chinese dataset.

Copied to clipboard

Chenming Tang, Zhixiang Wang, Yunfang Wu

Multilinguality

Challenge:	In-context learning improves performance of large language models (LLMs) performance of ICL highly depends on quality of demonstrations .
Approach:	They propose a syntactic-augmented COverage-based In-context example selection strategy that leverages syntastic knowledge beyond word matching to select better examples for machine translation.
Outcome:	The proposed strategy obtains the highest average COMET score among learning-free methods.

Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis (2026.acl-long)

Copied to clipboard

Wang Cai, Yilin Wen, Jinchang Hou, Du Su, Guoqiu Wang, Zhonghou Lv, Chenfu Bao, Yunfang Wu

An Error-Guided Correction Model for Chinese Spelling Error Correction (2022.findings-emnlp)

Challenge:	Existing mitigation strategies rely on global gradient geometry to resolve alignment conflicts . however, they overlook Modular Heterogeneity within Transformers, resulting in suboptimal trade-offs . Conflict-Aware Sparse Tuning (CAST) combines head-level diagnosis with sparse fine-tuning .
Approach:	They propose a framework that integrates head-level diagnosis with sparse fine-tuning to address this limitation.
Outcome:	The proposed framework integrates head-level diagnosis with sparse fine-tuning to reduce alignment conflicts in LLMs.

Copied to clipboard

Rui Sun, Xiuyu Wu, Yunfang Wu

Challenge:	Existing neural network approaches have achieved great progress on Chinese spelling correction, but there is still room for improvement.
Approach:	They propose an error-guided correction model that uses pre-trained BERT models to detect errors and integrate the error confusion set into the model.
Outcome:	The proposed model outperforms state-of-the-art models on widely used benchmarks and achieves superior performance on both quality and computation speed.

Multi-modal Semantic Understanding with Contrastive Cross-modal Feature Alignment (2024.lrec-main)

Copied to clipboard

Ming Zhang, Ke Chang, Yunfang Wu

Multimodality

Challenge:	Current work on multi-modal semantic understanding primarily exploits a dual-encoder structure to separate image and text, but fails to learn cross-modal feature alignment.
Approach:	They propose a CLIP-guided contrastive-learning-based architecture to perform multi-modal feature alignment by projecting features from different modalities into a unified deep space.
Outcome:	The proposed model outperforms baseline models on sarcasm detection and sentiment analysis tasks and is simple to implement without using task-specific external knowledge.

Multi-Task Learning with Language Modeling for Question Generation (D19-1)

Copied to clipboard

Wenjie Zhou, Minghua Zhang, Yunfang Wu