Papers by Wen Luo
AMANDA: Agentic Medical Knowledge Augmentation for Data-Efficient Medical Visual Question Answering (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing Med-MLLMs fail when deployed in low-resource settings where abundant labeled data is unavailable. |
| Approach: | They propose a training-free agentic framework that performs medical knowledge augmentation via LLM agents. |
| Outcome: | The proposed framework performs medical knowledge augmentation via LLM agents. |
Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations (2026.acl-long)
Copied to clipboard
Wen Luo, Guangyue Peng, Wei Li, Shaohang Wei, Feifan Song, Liang Wang, Nan Yang, Xingxing Zhang, Jing Jin, Furu Wei, Houfeng Wang
| Challenge: | Previous work shows that large language models generate hallucinations, yet the origins and mechanisms of these signals remain unclear. |
| Approach: | They propose to validate and disentangle two different pathways for truthfulness cues . they also propose to use the same mechanism to derive self-contained evidence from the generated answer . |
| Outcome: | The proposed applications improve hallucination detection performance by integrating two different inputs. |
Encode Errors: Representational Retrieval of In-Context Demonstrations for Multilingual Grammatical Error Correction (2025.findings-acl)
Copied to clipboard
| Challenge: | a novel method for encoding fine-grained error patterns improves performance on GEC. |
| Approach: | They propose a method for encoding grammatical errors from LLMs' internal states using a GER method. |
| Outcome: | The proposed method significantly boosts performance in ICL settings on multilingual GEC datasets. |
Explanation based In-Context Demonstrations Retrieval for Multilingual Grammatical Error Correction (2025.naacl-long)
Copied to clipboard
| Challenge: | Grammatical error correction (GEC) aims to correct grammatical, spelling, and semantic errors in natural language text. |
| Approach: | They propose a retrieval method based on natural language grammatical error explanations to match inputs with pre-constructed databases where explanations for erroneous samples are generated by LLMs. |
| Outcome: | The proposed method outperforms existing semantic and BM25-based retrieval techniques without additional training or language adaptation. |
IterAlign: Iterative Constitutional Alignment of Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Empirical results show that iterAlign improves truthfulness, helpfulness, harmlessness and honesty, improving the LLM alignment by up to 13.5% in harmlessness. |
| Approach: | They propose a data-driven constitution discovery and self-alignment framework called IterAlign to overcome these drawbacks by leveraging red teaming to uncover weaknesses of an LLM. |
| Outcome: | Empirical results show that iterAlign improves truthfulness, helpfulness, harmlessness and honesty by up to 13.5%. |
DisCal: Distribution-Aware Calibration for Mathematical Reasoning Under Character-Level Noisy Inputs (2026.acl-long)
Copied to clipboard
Bo Zhang, Jiawei Zhang, Cong Gao, Bingxu Han, Minghao Hu, Jun Zhang, Yunbo Cao, Zhunchen Luo, Wen Yao, Guotong Geng, Zhong Wang
| Challenge: | Existing methods for calibration of large reasoning models (LRMs) focus on clean inputs, leaving noise unexplored. |
| Approach: | They propose a confidence calibration framework for character-level noisy inputs that extracts uncertainty signals from both the empirical answer distribution and the model’s predictive distribution and integrates them via a learned calibrator. |
| Outcome: | Experiments on multiple mathematical reasoning benchmarks show that DisCal outperforms existing calibration methods under noisy inputs, reducing expected calibration error (ECE) by up to 39.21% and improving Area Under the Receiver Operating Characteristic Curve (AUROC) by 31.44%. |
M3: A Multi-View Fusion and Multi-Decoding Network for Multi-Document Reading Comprehension (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for multi-document reading comprehension cannot make full of the advantages of both approaches. |
| Approach: | They propose a multi-view fusion and multi-decoding method that integrates multiple documents for answering questions. |
| Outcome: | The proposed method improves on two mainstream multi-document reading comprehension datasets. |
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. |
Lost in Overlap: Exploring Logit-based Watermark Collision in LLMs (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing watermarking methods embed imperceptible identifiers into text to address copyright concerns. |
| Approach: | They propose a new philosophy for watermark attacks that addresses watermark collision . they demonstrate that collision poses a threat to all logit-based watermark algorithms . |
| Outcome: | The proposed method improves watermark collision performance on top of other methods. |
Odysseus Navigates the Sirens’ Song: Dynamic Focus Decoding for Factual and Diverse Open-Ended Text Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing decoding methods struggle to balance factuality and diversity . Deterministic decoding approaches suffer from degeneration and lack of diversity - a problem that is not addressed by the current literature. |
| Approach: | They propose a plug-and-play stochastic approach that adjusts decoding focus based on distributional differences across layers, leveraging the modular nature of factual knowledge within LLMs. |
| Outcome: | Extensive experiments on seven datasets show that DFD significantly improves performance. |
Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory (2026.acl-long)
Copied to clipboard
Derong Xu, Shuochen Liu, Pengfei Luo, Pengyue Jia, Yingyi Zhang, Yi Wen, Yimin Deng, Wenlin Zhang, Enhong Chen, Xiangyu Zhao, Tong Xu
| Challenge: | Existing memory systems rely on static, hand-crafted update rules for personalization, but sparse outcome rewards provide weak supervision, resulting in unstable long-horizon optimization. |
| Approach: | They propose a memory guideline optimization framework that learns how memory should be organized and what information to update. |
| Outcome: | The proposed framework learns how memory should be organized and what information to update. |
MARS-Bench: A Multi-turn Athletic Real-world Scenario Benchmark for Dialogue Evaluation (2025.findings-emnlp)
Copied to clipboard
Chenghao Yang, Yinbo Luo, Zhoufutu Wen, Qi Chu, Tao Gong, Longxiang Liu, Kaiyuan Zhang, Jianpeng Jiao, Ge Zhang, Wenhao Huang, Nenghai Yu
| Challenge: | Large Language Models (LLMs) have been widely adopted in real-world dialogue applications, but their robustness is criticized all along. |
| Approach: | They propose to use play-by-play text commentary to build a multi-turn athletic real-world scenario dialogue benchmark to evaluate three critical aspects of multi-turned conversations: ultra multi- turn, interactive multi-twist, and cross-turn tasks. |
| Outcome: | The proposed benchmarks outperform open-source LLMs on three critical aspects of multi-turn conversations: ultra multi-turned, interactive multi- turn, and cross-turn tasks. |
From Unaligned to Aligned: Scaling Multilingual LLMs with Multi-Way Parallel Corpora (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experiments show that models trained on multi-way parallel data outperform those trained on unaligned data. |
| Approach: | They propose a large-scale, high-quality multi-way parallel corpus based on TED Talks that spans 113 languages with up to 50 languages aligned in parallel. |
| Outcome: | The proposed model outperforms models trained on unaligned multilingual data on six multilingual benchmarks. |
Visual Prompt Tuning for Few-Shot Text Classification (2022.coling-1)
Copied to clipboard
| Challenge: | Existing work on pretraining models for text classification uses image encoders instead of visual prompts. |
| Approach: | They propose a method to deploy large-scale pre-trained models in the prompt-tuning paradigm in few-shot learning. |
| Outcome: | The proposed method outperforms the most recent prompt-tuning methods on five public text classification datasets. |
Original Content Is All You Need! an Empirical Study on Leveraging Answer Summary for WikiHowQA Answer Selection Task (2022.coling-1)
Copied to clipboard
Liang Wen, Juan Li, Houfeng Wang, Yingwei Luo, Xiaolin Wang, Xiaodong Zhang, Zhicong Cheng, Dawei Yin
| Challenge: | Existing answer selection approaches for community question answering lack additional answer summaries due to redundancy and lengthiness issues of crowdsourced answers. |
| Approach: | They constructed a dataset which contains a corresponding reference summary for each original lengthy answer. |
| Outcome: | The proposed model improves the performance of a question and candidate answer on a WikiHowQA dataset. |
Embracing Large Language Models in Traffic Flow Forecasting (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to predict future traffic flows capture spatio-temporal dependencies, but they fail to adapt to test-time environmental changes. |
| Approach: | They propose to use large language models to help traffic flow forecasting by capturing spatio-temporal dependencies and using a large language model to select the most likely result. |
| Outcome: | The proposed method is based on large language models (LLMs) and an LLM-based selector. |
MMATH: A Multilingual Benchmark for Mathematical Reasoning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a benchmark for multilingual complex reasoning spans 374 high-quality math problems across 10 typologically diverse languages. |
| Approach: | They propose a benchmark for multilingual complex reasoning across 10 languages . they show reasoning in English and answering in target languages can enhance performance . |
| Outcome: | The proposed benchmark demonstrates that models with high-quality reasoning can perform in multiple languages. |
Spec-o3: A Tool-Augmented Vision-Language Agent for Rare Celestial Object Candidate Vetting via Automated Spectral Inspection (2026.acl-long)
Copied to clipboard
| Challenge: | Spec-o3 is a tool-augmented vision-language agent that performs astronomer-aligned spectral inspection. |
| Approach: | They propose a tool-augmented vision-language agent that performs astronomer-aligned spectral inspection via interleaved multimodal chain-of-thought reasoning. |
| Outcome: | Spec-o3 outperforms traditional visual inspection methods on rare-object inspection tasks. |
Attention Weights as an Indicator: Analyzing and Improving Document Utilization in Retrieval-Augmented Generation (2026.acl-long)
Copied to clipboard
| Challenge: | In traditional RAG models, documents are grouped into categories based on their quality and order, and the quality of inputs is variable due to ineffective retrievers or misalignment between the retriever and generator. |
| Approach: | They propose to use attention weights to enhance document utilization from three perspectives: document ranking, placement, and filtering. |
| Outcome: | The proposed method outperforms baselines and improves document utilization effectiveness in a training-free manner. |
TemplateRL: Structured Template-Guided Reinforcement Learning for LLM Reasoning (2026.findings-acl)
Copied to clipboard
Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao
| Challenge: | Existing RL methods rely on unstructured self-sampling to fit scalar rewards, resulting in inefficient rollouts. |
| Approach: | They propose a structured template-guided RL framework that augments policy optimization with explicit template guidance. |
| Outcome: | Experiments show that TemplateRL outperforms GRPO and GRPI by 99% on AIME and 41% on AMC with superior stability on weak models and remarkable cross-domain generalization. |
FEA-Bench: A Benchmark for Evaluating Repository-Level Code Generation for Feature Implementation (2025.acl-long)
Copied to clipboard
Wei Li, Xin Zhang, Zhongxin Guo, Shaoguang Mao, Wen Luo, Guangyue Peng, Yangyu Huang, Houfeng Wang, Scarlett Li
| Challenge: | Existing benchmarks focus on standalone programming problems, such as HumanEval, MBPP, and LiveCodeBench. |
| Approach: | They propose to use large language models to evaluate their ability to perform incremental development within code repositories by collecting pull requests from 83 GitHub repositorias and using rule-based and intent-based filtering to construct task instances focused on new feature development. |
| Outcome: | The proposed benchmarks show that large language models perform significantly worse in the FEA-Bench, highlighting considerable challenges in repository-level incremental code development. |
Learn to Memorize: Scalable Continual Learning in Semiparametric Models with Mixture-of-Neighbors Induction Memory (2025.acl-long)
Copied to clipboard
| Challenge: | Semiparametric language models (LMs) use static storage, which lacks learning capability and is disconnected from the internal information flow of the parametric models. |
| Approach: | They reconceptualize the non-parametric memory represented by kNN-LM as a learnable Mixture-of-Neighbors Induction Memory (MoNIM) this synergizes the induction capabilities of attention heads with the memorization strength of feed-forward networks . |
| Outcome: | The proposed model is a learnable Mixture-of-neighbors induction memory (MoNIM) it synergizes the induction capabilities of attention heads with the memorization strength of feed-forward networks (FFNs). |
Beyond Examples: Towards Automated Thought-level In-Context Reasoning for Large Language Models (2026.acl-long)
Copied to clipboard
Jinyang Wu, Mingkuan Feng, Shuai Zhang, Feihu Che, Zhengqi Wen, Chonghua Liao, Ling Yang, Haoran Luo, Zheng Lian, Jianhua Tao
| Challenge: | In-context learning (ICL) struggles with complex reasoning due to superficial, example-level implicit imitation. |
| Approach: | They propose an automated method that shifts from surface-level examples to more guidance-oriented thought patterns. |
| Outcome: | The proposed method achieves 80.6% accuracy on MATH and 62.5% on AMC, surpassing GPT-4o’s 77.2% and 57.5% accuracy. |
DynamicNER: A Dynamic, Multilingual, and Fine-Grained Dataset for LLM-based Named Entity Recognition (2025.emnlp-main)
Copied to clipboard
Hanjun Luo, Yingbin Jin, Yiran Wang, Xinfeng Li, Tong Shang, Xuecheng Liu, Ruizhe Chen, Kun Wang, Hanan Salam, Qingsong Wen, Zuozhu Liu
| Challenge: | Existing datasets designed for Named Entity Recognition methods are inadequate for LLMs. |
| Approach: | They propose a dataset that is multilingual and multi-granular and enables LLMs to be applied to Named Entity Recognition methods. |
| Outcome: | The proposed dataset is multilingual and multi-granular, covering 8 languages and 155 entity types, with corpora spanning a diverse range of domains. |
HearSay Benchmark: Do Audio LLMs Leak What They Hear? (2026.findings-acl)
Copied to clipboard
Jin Wang, Kaiwen Luo, Liang Lin, Weiliu Wang, Yitian Chen, Moayad Aloqaily, Xuehai Tang, Zhenhong Zhou, Kun Wang, Li Sun, Qingsong Wen
| Challenge: | Recent advances in audio large language models have led to their potential privacy implications unexplored. |
| Approach: | They propose a benchmark to examine whether ALLMs leak user privacy through acoustic voiceprints. |
| Outcome: | The proposed benchmark is constructed from over 22,000 real-world audio clips. |
Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding (2025.findings-acl)
Copied to clipboard
| Challenge: | Low-resource methods for LLM alignment have been popular, but still face challenges in obtaining high-quality and aligned content. |
| Approach: | They propose a framework to enhance alignment ability of base models by the guidance of a small aligned model. |
| Outcome: | The proposed framework outperforms baseline methods while avoiding degradation on downstream tasks. |
Naive Bayes-based Context Extension for Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have shown promising in-context learning abilities, but are often impeded by length limitations of transformer architecture. |
| Approach: | They propose a framework to enable existing Large Language Models to perform ICL with an increased number of demonstration examples by significantly expanding their context size. |
| Outcome: | The proposed framework outperforms existing methods in in-context learning with an increased number of demonstration examples while preserving linear efficiency. |
Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models (2024.acl-long)
Copied to clipboard
Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, Ji-Rong Wen
| Challenge: | Despite the impressive multilingual capabilities demonstrated by LLMs, the understanding of how these abilities develop and function remains nascent. |
| Approach: | They propose a novel detection method to pinpoint language-specific neurons within LLMs by selectively activating or deactivating these neurons. |
| Outcome: | The proposed method can “steer” the output language of LLMs by selectively activating or deactivating language-specific neurons. |
SafeConf: A Confidence-Calibrated Safety Self-Evaluation Method for Large Language Models (2025.findings-emnlp)
Copied to clipboard
Bo Zhang, Cong Gao, Linkang Yang, Bingxu Han, Minghao Hu, Zhunchen Luo, Guotong Geng, Xiaoying Bai, Jun Zhang, Wen Yao, Zhong Wang
| Challenge: | Large language models (LLMs) have many advantages but they also pose significant safety risks. |
| Approach: | They propose a method to enhance the safety self-evaluation capability of LLMs . they perform semantic mutations on the original safety evaluation questions . |
| Outcome: | The proposed method improves safety self-evaluation accuracy by 5.86% and 7.79% over baseline methods on Chinese and English datasets. |
LLMBox: A Comprehensive Library for Large Language Models (2024.acl-demos)
Copied to clipboard
Tianyi Tang, Hu Yiwen, Bingqian Li, Wenyang Luo, ZiJing Qin, Haoxiang Sun, Jiapeng Wang, Shiyi Xu, Xiaoxue Cheng, Geyang Guo, Han Peng, Bowen Zheng, Yiru Tang, Yingqian Min, Yushuo Chen, Jie Chen, Ranchi Zhao, Luran Ding, Yuhao Wang, Zican Dong, Xia Chunxuan, Junyi Li, Kun Zhou, Xin Zhao, Ji-Rong Wen
| Challenge: | a library to facilitate the development, use, and evaluation of large language models (LLMs) is presented. |
| Approach: | They propose a unified library to facilitate the development, use and evaluation of large language models (LLMs). |
| Outcome: | The proposed library is based on extensive experiments in a variety of evaluation settings. |
An Empirical Study on the Overlapping Problem of Open-Domain Dialogue Datasets (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing benchmark datasets for open-domain dialogue generation are advancing the field . overlapping between training and test sets can cause fake performance . |
| Approach: | They analyze dailyDialog and OpenSubtitles to find out how overlapping can be exploited to obtain fake state-of-the-art performance. |
| Outcome: | The proposed datasets are cleaned and set up for future research. |
LogToP: Logic Tree-of-Program with Table Instruction-tuned LLMs for Controlled Logical Table-to-Text Generation (2026.findings-eacl)
Copied to clipboard
Yupian Lin, Guangya Yu, Cheng Yuan, Huan Du, Hui Luo, Yuang Bian, Jingping Liu, Zhidong He, Wen Du, Tong Ruan
| Challenge: | Existing LLMs are difficult to achieve satisfactory results in table-related tasks. |
| Approach: | They propose to develop a specialized logical table-to-text generation model that can be used for table-related tasks. |
| Outcome: | The proposed model achieves state-of-the-art on a Logic2Text dataset. |