Papers by Xu Wan
Enabling Stroke-Level Structural Analysis of Hieroglyphic Scripts without Language-Specific Priors (2026.findings-acl)
Copied to clipboard
Fuwen Luo, Zihao Wan, Ziyue Wang, Yaluo Liu, Pau Tong Lin Xu, Xuanjia Qiao, Xiaolong Wang, Peng Li, Yang Liu
| Challenge: | Existing structural analysis methods for hieroglyphic scripts are script-specific and labor-intensive. |
| Approach: | They propose a hieroglyphic Stroke Analyzer framework that captures character-internal structures and semantics without handcrafted data. |
| Outcome: | The proposed framework captures character-internal structures and semantics without priors . it can be used to generalize hieroglyphic scripts across languages . |
V-ALPHASOCIAL: Benchmark and Self-Reflective Chain-of-Thought Generation for Visual Social Commonsense Reasoning (2025.findings-acl)
Copied to clipboard
Zongyu Lin, Zhikun Xu, Xiaohan Song, Yixin Wan, Xingcheng Yao, Tsung-Han Lin, Selina Song, Pranav Subbaraman, Ben Zhou, Kai-Wei Chang, Yizhou Sun
| Challenge: | Social commonsense reasoning is a multimodal task that requires both textual and visual cues. |
| Approach: | They propose a method that integrates visual cues into social commonsense reasoning tasks. |
| Outcome: | The proposed method improves social commonsense reasoning on a multimodal foundation model. |
Tracing and Dissecting How LLMs Recall Factual Knowledge for Real World Questions (2025.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models have shown promising ability to perform commonsense reasoning. |
| Approach: | They propose a two-dimensional analysis framework that incorporates token back-tracing and token decoding to uncover how LLMs conduct factual knowledge recall. |
| Outcome: | The proposed framework shows that LLMs lack relevant knowledge but struggle to select the most accurate information based on context during the retrieval and rerank phase. |
Are LLM-based Evaluators Confusing NLG Quality Criteria? (2024.acl-long)
Copied to clipboard
| Challenge: | Existing studies show that LLMs confuse evaluation criteria, which reduces their reliability. |
| Approach: | They propose a hierarchical classification system for 11 common aspects with corresponding different evaluation criteria. |
| Outcome: | The proposed system is based on 11 common aspects with different evaluation criteria. |
kNN-ICL: Compositional Task-Oriented Parsing Generalization with Nearest Neighbor In-Context Learning (2024.naacl-long)
Copied to clipboard
Wenting Zhao, Ye Liu, Yao Wan, Yibo Wang, Qingyang Wu, Zhongfen Deng, Jiangshu Du, Shuaiqi Liu, Yunlong Xu, Philip Yu
| Challenge: | Recent advances in task-oriented parsing involve formulating the task as a sequence-to-sequence problem, relying on a wealth of labeled data. |
| Approach: | They propose a task-oriented parsing framework that integrates nearest-neighbor learning with a nearest-nearest approach. |
| Outcome: | The proposed model can be used to synthesize computer programs based on a natural-language prompt without additional data or specialized prompts. |
MC-MKE: A Fine-Grained Multimodal Knowledge Editing Benchmark Emphasizing Modality Consistency (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for knowledge editing in multimodal large language models focus on limited scenarios due to the lack of rigorous definition of multimodal knowledge. |
| Approach: | They propose a decomposed definition of multimodal knowledge and a benchmark to evaluate it. |
| Outcome: | The proposed method reveals that it is difficult to define multimodal knowledge editing in LLMs. |
TheoremQA: A Theorem-driven Question Answering Dataset (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent LLMs like GPT-4 and PaLM-2 have made tremendous progress in solving fundamental math problems like GSM8K by achieving over 90% accuracy. |
| Approach: | They propose to use theorem-driven question-answering dataset to evaluate AI models' ability to apply theoretic concepts to solving challenging science problems. |
| Outcome: | TheoremQA is curated by domain experts and contains 800 high-quality questions covering 350 theoremics from Math, Physics, EE&CS, and Finance. |
code-transformed: The Influence of Large Language Models on Code (2026.findings-eacl)
Copied to clipboard
| Challenge: | Using Large Language Models, code generation capabilities have transformed programming practices. |
| Approach: | They analyze 20,000 GitHub repositories linked to arXiv papers published between 2020 and 2025 . they identify measurable trends in the evolution of coding style that align with LLM-generated code . |
| Outcome: | The proposed study examines 20,000 GitHub repositories linked to arXiv papers . it finds that LLMs influence code style, and that they can be observed in real-world code . |
RST-Guarder: Enhancing Long-Context Robustness for Safeguards via RST Parsing and Probabilistic Inference (2026.acl-long)
Copied to clipboard
| Challenge: | Existing guardrail models for harmful-content detection degrade on long-form inputs . Existing models are vulnerable to policy-violating responses, causing false positives based on benign content . |
| Approach: | They propose an inference-time method that improves harmful-content detection for long-form inputs without additional data curation or model training. |
| Outcome: | The proposed method improves harmful-content detection for long-form inputs without additional data curation or model training. |
Evaluating Factuality in Cross-lingual Summarization (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluation metrics for monolingual summarization require translation to evaluate the factuality of cross-lingual summmarization. |
| Approach: | They propose to analyze cross-lingual factuality by collecting annotations and generated summaries from models at summary level and sentence level. |
| Outcome: | The proposed dataset shows that over 50% of generated summaries contain factual errors with different characteristics from monolingual summarization. |
JointCQ: Improving Factual Hallucination Detection with Joint Claim and Query Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for detecting factual hallucinations in generated content exhibit limitations in the first two stages of the halluciation detection pipeline. |
| Approach: | They propose a joint claim-and-query generation framework that can detect factual hallucinations in generated content. |
| Outcome: | The proposed method outperforms existing methods on open-domain QA hallucination detection benchmarks. |
MIL-Decoding: Detoxifying Language Models at Token-Level via Multiple Instance Learning (2023.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models generate toxic language which can cause security risks to their applications. |
| Approach: | They propose a method which detoxifies language models at token-level by interpolating it with a trained multiple instance learning network. |
| Outcome: | The proposed model outperforms baseline models in detoxification while hurting generation fluency a little bit. |
OEE-CFC: A Dataset for Open Event Extraction from Chinese Financial Commentary (2024.findings-emnlp)
Copied to clipboard
Qizhi Wan, Changxuan Wan, Rong Hu, Dexi Liu, Xu Wenwu, Kang Xu, Zou Meihua, Liu Tao, Jie Yang, Zhenwei Xiong
| Challenge: | Existing corpora with unconventional entities serving as event arguments lack rich multi-events and shared arguments. |
| Approach: | They develop an open event template that includes 21 event argument roles and an open corpus supporting open event extraction. |
| Outcome: | The proposed corpus includes 17,469 events, 44,221 arguments, 3,644 complex arguments, and 5,898 shared arguments. |
Diversifying Neural Text Generation with Part-of-Speech Guided Softmax and Sampling (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods to generate text using contextual features do not consider syntactic structure clues. |
| Approach: | They propose using linguistic annotation, i.e., part-of-speech (POS), to guide the text generation. |
| Outcome: | The proposed method can generate more diverse text while maintaining comparable quality. |
AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models (2025.findings-acl)
Copied to clipboard
Qi Liu, Jingqing Ruan, Hao Li, Haodong Zhao, Desheng Wang, Jiansong Chen, Wan Guanglu, Xunliang Cai, Zhi Zheng, Tong Xu
| Challenge: | Existing multi-objective preference alignment methods for large language models face limitations such as auxiliary reward/reference models and computational complexity. |
| Approach: | They propose a framework that achieves dynamic balance across preference dimensions by using dimension-aware generation metrics as implicit rewards. |
| Outcome: | Empirical results show that AMoPO outperforms state-of-the-art methods by 28.5% . |
DecoCal: Decoding with Calibration in Diffusion Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Diffusion Large Language Models (DLLMs) generate text via iterative token denoising . but decoding is challenging, with many tokens appearing predictable early . |
| Approach: | They propose a Decoding framework that performs Calibration of token-level confidence across diffusion steps and leverages the calibrated results to guide decoding decisions. |
| Outcome: | Experiments on multiple DLLMs and benchmarks show that DecoCal improves generation accuracy compared to existing strategies. |
Huatuo-26M, a Large-scale Chinese Medical QA Dataset (2025.findings-naacl)
Copied to clipboard
Xidong Wang, Jianquan Li, Shunian Chen, Yuxuan Zhu, Xiangbo Wu, Zhiyi Zhang, Xiaolong Xu, Junying Chen, Jie Fu, Xiang Wan, Anningzhe Gao, Benyou Wang
| Challenge: | Large Language Models are a powerful tool for medical research, but the data is a bottleneck. |
| Approach: | They propose to use the largest ever medical Question Answering dataset with 26 Million QA pairs as a fine-tuning data for training large language models. |
| Outcome: | The proposed dataset demonstrates that it can be used to train large language models and improves zero-shot performance on other datasets. |
AgentOCR: Reimagining Agent History via Optical Self-Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have enabled agentic systems trained with reinforcement learning over multi-turn interaction, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token and memory costs. |
| Approach: | They propose a framework that represents the accumulated observation-action history as a compact rendered image. |
| Outcome: | The proposed framework preserves over 95% of text-based agent performance while significantly reducing token consumption (>50%), yielding consistent token and memory efficiency. |
FaStFact: Faster, Stronger Long-Form Factuality Evaluations in LLMs (2025.findings-emnlp)
Copied to clipboard
Yingjia Wan, Haochen Tan, Xiao Zhu, Xinyu Zhou, Zhiwei Li, Qingsong Lv, Changxuan Sun, Jiaqi Zeng, Yi Xu, Jianqiao Lu, Yinhong Liu, Zhijiang Guo
| Challenge: | Prior evaluation pipelines fail to evaluate factuality of long-form LLMs due to inefficiency and costly human assessment. |
| Approach: | They propose a fast and strong evaluation pipeline that can evaluate factuality of long-form LLMs . they propose 'faStFact' to reduce cost of web searching and inference calling . |
| Outcome: | The proposed evaluation pipeline achieves highest alignment with human evaluation and efficiency among existing baselines. |
RICO: Improving Accuracy and Completeness in Image Recaptioning via Visual Reconstruction (2025.emnlp-main)
Copied to clipboard
Yuchi Wang, Yishuo Cai, Shuhuai Ren, Sihan Yang, Linli Yao, Yuanxin Liu, Yuanxing Zhang, Pengfei Wan, Xu Sun
| Challenge: | Existing recaptioning methods suffer from inaccuracies due to missing fine-grained details. |
| Approach: | They propose a framework that refines captions through visual reconstruction using a text-to-image model and a visual reconstruction framework. |
| Outcome: | The proposed framework outperforms baselines on CapsBench and CompreCap by 10%. |
Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards (2026.acl-long)
Copied to clipboard
Youliang Yuan, Qiuyang Mang, Jingbang Chen, Hong Wan, Xiaoyuan Liu, Junjielong Xu, Jen-tse Huang, Wenxuan Wang, Wenxiang Jiao, Pinjia He
| Challenge: | Existing models are susceptible to reward hacking, leading to a substantial overestimation of a model's reasoning ability. |
| Approach: | They propose a Rubric Reward Model that rewards the entire reasoning trajectory against problem-specific rubrics. |
| Outcome: | The proposed model outperforms outcome-only supervision on four math benchmarks and boosts Verified Pass@1024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. |
Interpretable User Satisfaction Estimation for Conversational Systems with Large Language Models (2024.acl-long)
Copied to clipboard
Ying-Chun Lin, Jennifer Neville, Jack Stokes, Longqi Yang, Tara Safavi, Mengting Wan, Scott Counts, Siddharth Suri, Reid Andersen, Xiaofeng Xu, Deepak Gupta, Sujay Kumar Jauhar, Xia Song, Georg Buscher, Saurabh Tiwary, Brent Hecht, Jaime Teevan
| Challenge: | Existing approaches to user satisfaction estimation are hard to interpret and lack generalizable patterns. |
| Approach: | They propose to use supervised prompting to extract interpretable user satisfaction signals from natural language utterances to tailor an LLM to USE using labeled examples. |
| Outcome: | The proposed method extracts interpretable signals of user satisfaction from natural language utterances more effectively than embedding-based approaches. |
HAD: HAllucination Detection Language Models Based on a Comprehensive Hallucination Taxonomy (2026.acl-industry)
Copied to clipboard
| Challenge: | relying on large language models for information has raised concerns about reliability and accuracy of outputs. |
| Approach: | They propose a hallucination taxonomy with 11 categories for various NLG tasks and propose HAllucination Detection models which integrate hallucinism detection, span-level identification, and correction into a single inference process. |
| Outcome: | The proposed models outperform baselines on HaluEval, FactCHD, and FaithBench, confirming their robustness and versatility. |
SegTune: Structured and Fine-Grained Control for Song Generation (2026.acl-long)
Copied to clipboard
Yuejiao Wang, Zihao Ji, Pengfei Cai, Xu Li, Haorui Zheng, Zewen Song, Zhongliang Liu, Chen Zhang, Pengfei Wan
| Challenge: | Recent advances in neural song generation have enabled high-quality synthesis from lyrics and global textual prompts. |
| Approach: | They propose a framework that allows users to specify local musical descriptions aligned to song segments. |
| Outcome: | The proposed framework outperforms baselines in musicality and controllability. |
Rethinking the Video Sampling and Reasoning Strategies for Temporal Sentence Grounding (2022.findings-emnlp)
Copied to clipboard
Jiahao Zhu, Daizong Liu, Pan Zhou, Xing Di, Yu Cheng, Song Yang, Wenzheng Xu, Zichuan Xu, Yao Wan, Lichao Sun, Zeyu Xiong
| Challenge: | Existing methods for temporal sentence grounding ignore two crucial issues . 1) Boundary-bias: the video downsampling process may lose these two frames . 2) Reasoning-biases: such incorrect new boundary frames lead to the reasoning bias . |
| Approach: | They propose a siamese sampling mechanism to generate additional contextual frames . they use a reasoning strategy to learn the inter-relationship among these frames a . |
| Outcome: | Extensive experiments demonstrate the effectiveness of a new siamese sampling network on three challenging datasets. |
Benchmarking Knowledge Boundary for Large Language Models: A Different Perspective on Model Evaluation (2024.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models have improved performance across tasks . however, the sensitivity of LLMs to prompt leads to unreliability of evaluation results . |
| Approach: | They propose a new concept to evaluate language models with a fixed question or limited paraphrases as the query. |
| Outcome: | The proposed method outperforms existing benchmarks on multiple language models . it avoids prompt sensitivity, rendering models more reliable and robust . |
AceGPT, Localizing Large Language Models in Arabic (2024.naacl-long)
Copied to clipboard
Huang Huang, Fei Yu, Jianqing Zhu, Xuening Sun, Hao Cheng, Song Dingjie, Zhihong Chen, Mosen Alharthi, Bang An, Juncai He, Ziche Liu, Junying Chen, Jianquan Li, Benyou Wang, Lian Zhang, Ruoyu Sun, Xiang Wan, Haizhou Li, Jinchao Xu
| Challenge: | Significant concerns emerge when addressing cultural sensitivity and local values. |
| Approach: | They propose a localized Large Language Model (LLM) specifically for Arabic, a language imbued with unique cultural characteristics inadequately addressed by current mainstream models. |
| Outcome: | The proposed model sets the state-of-the-art standard for open Arabic LLMs across various benchmarks. |
FaithBench: A Diverse Hallucination Benchmark for Summarization by Modern LLMs (2025.naacl-short)
Copied to clipboard
Forrest Sheng Bao, Miaoran Li, Renyi Qu, Ge Luo, Erana Wan, Yujia Tang, Weisi Fan, Manveer Singh Tamber, Suleman Kazi, Vivek Sourabh, Mike Qi, Ruixuan Tu, Chenyu Xu, Matthew Gonzales, Ofer Mendelevitch, Amin Ahmad
| Challenge: | Existing evaluations of hallucinations in large language models suffer from a lack of diversity and recency in the LLM and LLM families considered. |
| Approach: | They propose a summarization hallucination benchmark that challenges models to disagree on hallucines . they use models to generate answers or summaries from textual input . |
| Outcome: | The proposed model combines the best of 10 modern LLMs with ground truth annotations. |
DAMON: A Dialogue-Aware MCTS Framework for Jailbreaking Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for multi-turn attacks mainly utilize a predefined dialogue pattern, limiting their effectiveness in realistic situations. |
| Approach: | They propose a multi-turn jailbreak attack method that leverages Monte Carlo Tree Search to explore multi-turned conversational spaces and identifies sub-instruction sequences that induce harmful responses. |
| Outcome: | The proposed method can induce undesired behaviors across five LLMs and three datasets. |
Constrained Regeneration for Cross-Lingual Query-Focused Extractive Summarization (2022.coling-1)
Copied to clipboard
Elsbeth Turcan, David Wan, Faisal Ladhak, Petra Galuscakova, Sukanta Sen, Svetlana Tchistiakova, Weijia Xu, Marine Carpuat, Kenneth Heafield, Douglas Oard, Kathleen McKeown
| Challenge: | Query-focused summarization of foreign-language documents can help a user understand whether a document is relevant to a query term. |
| Approach: | They propose to use machine translation and post-editing to improve human relevance judgments . they include a query term in a summary when its translation appears in the source document . |
| Outcome: | The proposed approach improves human relevance judgments by including a query term in a summary when its translation appears in the source document. |
Teaching the Pre-trained Model to Generate Simple Texts for Text Simplification (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing strategies to teach pre-trained models to generate simple texts are inadequate. |
| Approach: | They propose a continued pre-training strategy to teach pre-trained models to generate simple texts by randomly masking text spans in ordinary texts. |
| Outcome: | The proposed strategy improves on lexical simplification, sentence simplification and document-level simplification tasks over existing models. |
Group Preference Alignment: Customizing LLM Responses from In-Situ Conversations Only When Needed (2025.emnlp-industry)
Copied to clipboard
Ishani Mondal, Jack W. Stokes, Sujay Kumar Jauhar, Longqi Yang, Mengting Wan, Xiaofeng Xu, Xia Song, Jordan Lee Boyd-Graber, Jennifer Neville
| Challenge: | Existing methods for group-aware adaptation capture divergent preferences from real-world conversation logs into interpretable rubrics. |
| Approach: | They propose a group-aware personalization framework that captures context-specific preferences and steers LLMs accordingly. |
| Outcome: | The proposed framework improves group alignment without compromising perfomance on benchmarks. |
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback (2026.acl-long)
Copied to clipboard
Taiwei Shi, Zhuoer Wang, Longqi Yang, Ying-Chun Lin, Zexue He, Mengting Wan, Pei Zhou, Sujay Kumar Jauhar, Sihao Chen, Shan Xia, Hongfei Zhang, Jieyu Zhao, Xiaofeng Xu, Xia Song, Jennifer Neville
| Challenge: | Traditional alignment methods rely on human annotations and are subjective and misalignment with real-world user preferences. |
| Approach: | They propose a framework that leverages in-situ user feedback during conversations with LLMs to create preference datasets automatically. |
| Outcome: | The proposed framework identifies and classifies user feedback to LLM responses between conversation turns and creates examples of preferred and dispreferred responses according to user preferences. |
GenTool: Enhancing Tool Generalization in Language Models through Zero-to-One and Weak-to-Strong Simulation (2025.findings-acl)
Copied to clipboard
Jie He, Jennifer Neville, Mengting Wan, Longqi Yang, Hui Liu, Xiaofeng Xu, Xia Song, Jeff Z. Pan, Pei Zhou
| Challenge: | Large Language Models (LLMs) can expand their capabilities by integrating external tools. |
| Approach: | They propose a training framework that prepares LLMs for diverse generalization challenges in tool utilization. |
| Outcome: | The proposed framework improves the tool-usage capabilities of LLMs by up to 8B parameters, surpassing GPT-4o. |
AdapThink: Adaptive Thinking Preferences for Reasoning Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent research has highlighted a significant inefficiency associated with the slow thinking paradigm . models often overthink simple tasks while underthinking complex challenges . |
| Approach: | They propose a framework for adaptive reasoning preference control that dynamically adjusts reflection preferences based on group-level distributional statistics of reasoning length and reflection intensity. |
| Outcome: | The proposed framework reduces average response length by 17.1%-21.4% while improving performance by 6.12-6.59 points under 32K token budgets. |