Papers by Song Yuan

Robustness via Referencing: Defending against Prompt Injection Attacks by Referencing the Executed Instruction (2026.findings-acl)

Challenge:	Existing methods to perform multimodal information extraction only investigated entity-based tasks under supervised learning with adequate labeled data.
Approach:	They propose to investigate the entity-based MIE tasks under the low-resource settings by decomposing the features into image, entity, and context factors.
Outcome:	The proposed method is able to perform on two public MIE benchmark datasets and the experimental results confirm it.

Copied to clipboard

Yulin Chen, Haoran Li, Yuan Sui, Yue Liu, Yufei He, Xiaoling Bai, Chi Fei, Li Yabo, Haozhe Ma, Yangqiu Song, Bryan Hooi

PM2F2N: Patient Multi-view Multi-modal Feature Fusion Networks for Clinical Outcome Prediction (2022.findings-emnlp)

Challenge:	Prompt injection attacks manipulate large language models (LLMs) by misleading them to deviate from the original input instructions and execute maliciously injected instructions.
Approach:	They propose a prompt injection defense method that suppresses the model's instruction-following tendencies rather than suppressing them.
Outcome:	The proposed method outperforms prompt-engineering-based approaches and fine-tuning methods and reduces the ASR to nearly 0% in some scenarios.

Copied to clipboard

Ying Zhang, Baohang Zhou, Kehui Song, Xuhui Sui, Guoqing Zhao, Ning Jiang, Xiaojie Yuan

Can Indirect Prompt Injection Attacks Be Detected and Removed? (2025.acl-long)

Challenge:	Existing methods focused on time series data but ignored clinical notes . fusion of multi-modal features of patients from different views is not feasible due to the time series and clinical notes data being stored as time series.
Approach:	They propose to combine time series and clinical notes to fuse multi-modal features of patients from different perspectives using graph neural networks.
Outcome:	The proposed method is superior to existing models on MIMIC-III benchmark.

Copied to clipboard

Yulin Chen, Haoran Li, Yuan Sui, Yufei He, Yue Liu, Yangqiu Song, Bryan Hooi

K-PLUG: Knowledge-injected Pre-trained Language Model for Natural Language Understanding and Generation in E-Commerce (2021.findings-emnlp)

Challenge:	Recent studies have developed various detection mechanisms to protect against prompt injection attacks.
Approach:	They investigate the feasibility of detecting and removing indirect prompt injection attacks . they use two methods to evaluate their performance and train detection models .
Outcome:	The proposed method is based on a benchmark dataset and is available on github . it evaluates the performance of existing models and open-source detection models .

Copied to clipboard

Song Xu, Haoran Li, Peng Yuan, Yujia Wang, Youzheng Wu, Xiaodong He, Ying Liu, Bowen Zhou

SOLAR: Serendipity Optimized Language Model Aligned for Recommendation (2025.findings-emnlp)

Challenge:	Existing pre-trained language models are not explicitly aware of domain-specific knowledge, which is essential for downstream tasks in many domains, such as tasks in e-commerce scenarios.
Approach:	They propose a knowledge-injected pre-trained language model that can be transferred to both natural language understanding and generation tasks.
Outcome:	The proposed model significantly outperforms baselines across the board in e-commerce scenarios.

Copied to clipboard

Zichen Yuan, Lifan Sun, Yucen Zhuang, Yue Wang, Xinyuan Song, Tianqi Xu, Siyuan Li, Junchen Fu, Youhua Li, Sirui Hong, Jiaqi Chen, Joemon M. Jose, Yongxin Ni

PALT: Parameter-Lite Transfer of Language Models for Knowledge Graph Completion (2022.findings-emnlp)

Challenge:	Large Language Models have shown strong potential in recommendation tasks . however, their application to serendipity-oriented recommendations remains challenging .
Approach:	They propose a domain-adaptive instruction tuning method that aligns Large Language Models with recommendation tasks.
Outcome:	The proposed framework bridges the domain gap between LLMs and recommendation tasks.

Copied to clipboard

Jianhao Shen, Chenguang Wang, Ye Yuan, Jiawei Han, Heng Ji, Koushik Sen, Ming Zhang, Dawn Song

See the World, Discover Knowledge: A Chinese Factuality Evaluation for Large Vision Language Models (2025.findings-acl)

Challenge:	Pretrained language models (LMs) are a powerful transfer learning approach for knowledge graph (KG) completion.
Approach:	They propose a parameter-lite transfer learning approach for pretrained language models for knowledge graph (KG) completion.
Outcome:	The proposed model outperforms the state-of-the-art models on a knowledge graph completion benchmark by tuning 1% of the parameters.

Copied to clipboard

Jihao Gu, Yingyao Wang, Pi Bu, Chen Wang, Ziming Wang, Tengtao Song, Donglai Wei, Jiale Yuan, Yingxiu Zhao, Yancheng He, Shilong Li, Jiaheng Liu, Meng Cao, Jun Song, Yingshui Tan, Xiang Li, Wenbo Su, Xiaoyong Zhu, Bo Zheng

EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms (2025.naacl-long)

Challenge:	Existing models for large vision language models do not fully reflect their knowledge capacity and reliability, resulting in erroneous outputs that do not align with the image content or provide answers lacking knowledge evidence.
Approach:	They propose a Chinese-based benchmark for visual factuality across 8 major topics and 56 subtopics and a multi-hop question construction.
Outcome:	The proposed model decouples visual factuality into two parts: seeing the world and discovering knowledge.

Copied to clipboard

Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, Deqing Yang

DP3: Differentially Private Prompt Perturbation for Multi-turn LLM Inference (2026.findings-acl)

Challenge:	Existing work on extending specialized agents to multi-agent systems is dependent on human-designed frameworks, limiting the functional scope and scalability of agent systems.
Approach:	They propose a generic method to automatically extend specialized agents to multi-agent systems via evolutionary algorithm . they consider existing agent frameworks as the initial individual and apply evolutionary operators to generate multiple agents with diverse settings.
Outcome:	The proposed method can extend specialized agents to multi-agent systems . it can generate multiple agents with diverse settings, and improves performance across tasks .

Copied to clipboard

Lele Zheng, Chao Zhang, Feiyang Yuan, Ke Cheng, Tao Zhang, Anxiao Song, Yulong Shen

Selecting Key Views for Zero-Shot Entity Linking (2023.findings-emnlp)

Challenge:	Large language models (LLMs) are widely used for text understanding and generation . existing methods that assume single-turn interactions break down in multi-turn settings .
Approach:	They propose a differentially private prompt perturbation framework for multi-turn LLM inference . DP3 constructs a perturbation mapping table to reuse perturbations for recurring tokens .
Outcome:	The proposed framework reduces privacy costs and degrades cross-turn semantic coherence . it also provides a context-aware utility function to maintain semantic consistency across turns .

Copied to clipboard

Xuhui Sui, Ying Zhang, Kehui Song, Baohang Zhou, Xiaojie Yuan, Wensheng Zhang

Bring Invariant to Variant: A Contrastive Prompt-based Framework for Temporal Knowledge Graph Forecasting (2024.lrec-main)

Challenge:	Entity linking is a task of assigning ambiguous mentions in textual input to entities in knowledge bases.
Approach:	They propose a framework to align mentions in text to entities in knowledge bases . they use unsupervised clustering to select key views from descriptions .
Outcome:	The proposed framework achieves state-of-the-art on the zero-shot entity linking dataset.

Copied to clipboard

Ying Zhang, Xinying Qian, Yu Zhao, Baohang Zhou, Kehui Song, Xiaojie Yuan

REAL: REtrieval-reAsoning and Logic-constructed Attention Behaviors for Long-Context KV Cache Compression (2026.acl-long)

Challenge:	Existing methods for temporal knowledge graph forecasting are insufficient structural contexts to learn effective representations.
Approach:	They propose a Contrastive Prompt-based framework with Entity background information for TKG forecasting that brings time-invariant entity background information to time-variant structural information.
Outcome:	The proposed framework is effective and stays competitive in inference with limited structural information.

Copied to clipboard

Mengjie Li, Yuan Feng, Xike Xie, William J. Song

Reasoning

Challenge:	Existing methods to evict keyvalue caches ignore diverse behavior in failure cases, such as bias and distraction.
Approach:	They propose a method to analyze attention head behaviors in success and failure scenarios by maximizing signal-to-noise ratio and minimizing noise from bias and distraction.
Outcome:	The proposed method achieves comparable accuracy to the strongest baseline, HeadKV-R2 on LongBench v2 while requiring 32x less space.

Self-Attention Guided Copy Mechanism for Abstractive Summarization (2020.acl-main)

Copied to clipboard

Song Xu, Haoran Li, Peng Yuan, Youzheng Wu, Xiaodong He, Bowen Zhou

PRINCE: Prefix-Masked Decoding for Knowledge Enhanced Sequence-to-Sequence Pre-Training (2022.emnlp-main)

Challenge:	Abstractive summarization models have been widely used to extract words from source into summary, but how to ensure that important words in source are copied remains a challenge.
Approach:	They propose a Transformer-based model to enhance copy mechanism by identifying the importance of each source word based on the degree centrality.
Outcome:	The proposed model outperforms baseline methods on CNN/Daily Mail and Gigaword datasets.

Copied to clipboard

Song Xu, Haoran Li, Peng Yuan, Youzheng Wu, Xiaodong He

Backdoor-Powered Prompt Injection Attacks Nullify Defense Methods (2025.findings-emnlp)

Challenge:	Existing studies focus on injecting noises into the input sequence, but feasibility of injecting them into the decoding sequence remains an open question.
Approach:	They propose a pre-training paradigm that integrates knowledge-enhanced decoding with noises in the prefix to strengthen the representation learning of entities that span over multiple input tokens.
Outcome:	The proposed model achieves state-of-the-art results on two knowledge-driven data-to-text generation tasks with up to 2% BLEU gains.

Copied to clipboard

Yulin Chen, Haoran Li, Yuan Sui, Yangqiu Song, Bryan Hooi

Can Large Language Models Grasp Legal Theories? Enhance Legal Reasoning with Insights from Multi-Agent Collaboration (2024.findings-emnlp)

Challenge:	Recent studies have shown that LLMs are vulnerable to prompt injection attacks because of their instruction-following abilities and inability to distinguish the instructions in the data content.
Approach:	They propose backdoor-powered prompt injection attacks that trick LLMs into deviating from the original input instruction and executing the attackers’ target instruction.
Outcome:	The proposed attacks trick the LLMs into deviating from the input instruction and executing the attackers’ target instruction.

Copied to clipboard

Weikang Yuan, Junjie Cao, Zhuoren Jiang, Yangyang Kang, Jun Lin, Kaisong Song, Tianqianjin Lin, Pengwei Yan, Changlong Sun, Xiaozhong Liu

Beyond Transcription: Unified Audio Schema for Perception-Aware AudioLLMs (2026.findings-acl)

Challenge:	Existing studies have found that when LLMs are given criminal facts and legal rules, then asked whether cases constitute a certain charge, they struggle to understand legal theories and perform basic legal reasoning tasks.
Approach:	They propose a task to assess LLMs' understanding of legal theories and reasoning capabilities by using a novel framework: Multi-Agent framework for improving complex legal reasoning capability.
Outcome:	The proposed framework improves LLMs' understanding of legal theories and reasoning abilities in real-world scenarios.

Copied to clipboard

Linhao Zhang, Yuhan Song, Aiwei Liu, Chuhan Wu, Sijun Zhang, Wei Jia, Yuan Liu, Houfeng Wang, Zhou Xiao

Act as you think: Reinforcing Consistent Reasoning in Medical Visual Question Answering (2026.acl-long)

Challenge:	Recent Audio Large Language Models (AudioLLMs) excel at reasoning tasks, but struggle at elementary auditory perception.
Approach:	They propose a framework that organizes audio information into three explicit components in a unified JSON format.
Outcome:	The proposed framework boosts fine-grained perception by 10.9% on MMSU over state-of-the-art models while preserving robust reasoning capabilities.

Copied to clipboard

Songtao Jiang, Yuan Wang, Ruizhe Chen, Yan Zhang, Ruilin Luo, Bohan Lei, Yeying Jin, Sibo Song, ZhiBo Yang, Jimeng Sun, Jian Wu, Zuozhu Liu

Robustness in NLP

Challenge:	Recent advances have improved the accuracy of medical visual question answering (Med-VQA) however, the high stakes nature of the medical domain has precipitated a shift towards interpretability and transparency of reasoning processes.
Approach:	They propose a reinforcement learning from verifiable rewards framework that rewards internal consistency and logical coherence.
Outcome:	The proposed framework rewards internal consistency and logical coherence, and is highly versatile, the authors show.

Question Directed Graph Attention Network for Numerical Reasoning over Text (2020.emnlp-main)

Copied to clipboard

Kunlong Chen, Weidi Xu, Xingyi Cheng, Zou Xiaochuan, Yuyu Zhang, Le Song, Taifeng Wang, Yuan Qi, Wei Chu

Reasoning

Challenge:	Numerical reasoning requires both natural language understanding and arithmetic computation.
Approach:	They propose a graph representation for the context of the passage and question needed for numerical reasoning.
Outcome:	The proposed model achieves remarkable results in benchmark datasets such as DROP.

MELOV: Multimodal Entity Linking with Optimized Visual Features in Latent Space (2024.findings-acl)

Copied to clipboard

Xuhui Sui, Ying Zhang, Yu Zhao, Kehui Song, Baohang Zhou, Xiaojie Yuan

On the Faithfulness for E-commerce Product Summarization (2020.coling-main)

Challenge:	Existing approaches to multimodal entity linking focus on textual contexts but lack in social media vision modality.
Approach:	They propose a latent space vision feature optimization framework MELOV to address these challenges . they exploit variational autoencoder to mine shared information and generate text-based visual features .
Outcome:	The proposed framework is superior to existing methods on three benchmark datasets.

Copied to clipboard

Peng Yuan, Haoran Li, Song Xu, Youzheng Wu, Xiaodong He, Bowen Zhou

Text Generation

Challenge:	e-commerce product summarization requires consistency between product attributes and summary . inconsistent product summaries can mislead users and decrease public credibility .
Approach:	They propose a model to generate e-commerce product summaries with product attributes . they encode product attribute table and constrain attribute words to be presented only through copying .
Outcome:	The proposed model significantly improves the faithfulness of e-commerce product summarization tasks.

LeCoDe: A Benchmark Dataset for Interactive Legal Consultation Dialogue Evaluation (2026.acl-long)

Copied to clipboard

Weikang Yuan, Kaisong Song, Zhuoren Jiang, Junjie Cao, Yujie Zhang, Jun Lin, Kun Kuang, Ji Zhang, Xiaozhong Liu

Dialogue Systems & Conversational Agents

Challenge:	Current systems for legal consultation are insufficient to handle the knowledge-intensive nature of real-world consultations.
Approach:	They propose a multi-turn benchmark dataset to evaluate LLMs in legal consultation settings.
Outcome:	The proposed framework assesses LLMs’ consultation capabilities in terms of (1) clarification capability and (2) professional advice quality.

Train in Vain: Functionality-Preserving Poisoning to Prevent Unauthorized Use of Code Datasets (2026.findings-acl)

Copied to clipboard

Yuan Xiao, Jiaming Wang, Yuchen Chen, Wei Song, Jun Sun, Shiqing Ma, Yanzhou Mu, Juan Zhai, Chunrong Fang, Jin Song Dong, Zhenyu Chen

Programming Languages in NLP

Challenge:	Existing methods for dataset poisoning require full-dataset poison, which breaks code compilability.
Approach:	They propose a functionality-preserving poisoning approach that injects short, compilable weak-use fragments into executed code paths.
Outcome:	The proposed method contaminates 10% of the dataset while maintaining 100% compilability and functional correctness.

EASYTOOL: Enhancing LLM-based Agents with Concise Tool Instruction (2025.naacl-long)

Copied to clipboard

Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Kan Ren, Dongsheng Li, Deqing Yang

BioFEG: Generate Latent Features for Biomedical Entity Linking (2023.emnlp-main)

Challenge:	EASYTOOL combines tools from diverse tool documentation into a single tool instruction.
Approach:	They propose a framework that transforms tool documentation into a unified tool instruction.
Outcome:	EASYTOOL combines extensive tool documentation into a concise tool instruction . it reduces token consumption and improves performance of LLM-based agents .

Copied to clipboard

Xuhui Sui, Ying Zhang, Xiangrui Cai, Kehui Song, Baohang Zhou, Xiaojie Yuan, Wensheng Zhang

Knowledge Representation

Challenge:	Existing approaches to biomedical entity linking suffer from multiple types of errors due to the rarity of many biomedically relevant entities in real-world scenarios.
Approach:	They propose a latent feature generation framework to generate latent semantic features for unseen entities to capture fine-grained coherence information of unseened entities.
Outcome:	The proposed framework is superior to existing models on two benchmark datasets.

OpenCoder: The Open Cookbook for Top-Tier Code Large Language Models (2025.acl-long)

Copied to clipboard

Siming Huang, Tianhao Cheng, Jason Klein Liu, Weidi Xu, Jiaran Hao, Liuyihan Song, Yang Xu, Jian Yang, Jiaheng Liu, Chenchen Zhang, Linzheng Chai, Ruifeng Yuan, Xianzhen Luo, Qiufeng Wang, YuanTao Fan, Qingfu Zhu, Zhaoxiang Zhang, Yang Gao, Jie Fu, Qian Liu, Houyi Li, Ge Zhang, Yuan Qi, Xu Yinghui, Wei Chu, Zili Wang

LongDocURL: a Comprehensive Multimodal Long Document Benchmark Integrating Understanding, Reasoning, and Locating (2025.acl-long)

Challenge:	Code LLMs lack reproducible data pipelines and training protocols for reproducible advancements in code intelligence.
Approach:	They propose a top-tier code LLM that releases model weights and inference code . reproducible data pipelines, rigorous experimental ablation results and training protocols are included .
Outcome:	The proposed model achieves comparable performance to leading models and serves as an "open cookbook" reproducible training data, rigorous experimental ablation results, and detailed training protocols are also included in the model.

Copied to clipboard

Chao Deng, Jiale Yuan, Pi Bu, Peijie Wang, Zhong-Zhi Li, Jian Xu, Xiao-Hui Li, Yuan Gao, Jun Song, Bo Zheng, Cheng-Lin Liu

Reasoning

Challenge:	Existing document understanding benchmarks only handle a small number of pages . existing models are limited to handling only a limited number of documents .
Approach:	They propose a long document understanding benchmark that integrates three primary tasks and 20 sub-tasks based on different primary tasks.
Outcome:	The proposed model outperforms existing benchmarks on open-source and closed-source models . the model outpersforms other models on more than 33,000 pages of documents .

Learn to Copy from the Copying History: Correlational Copy Network for Abstractive Summarization (2021.emnlp-main)

Copied to clipboard

Haoran Li, Song Xu, Peng Yuan, Yujia Wang, Youzheng Wu, Xiaodong He, Bowen Zhou

Text Generation

Challenge:	Existing methods for abstractive summarization use encoder-decoder attention, but this leads to incomplete copying.
Approach:	They propose a copying scheme that takes advantage of prior copying distributions and explicitly encourages the model to copy the input word that is relevant to the previously copied one.
Outcome:	The proposed scheme achieves state-of-the-art on summarization benchmarks . it takes advantage of prior copying distributions and explicitly encourages copying .

Sens-Merging: Sensitivity-Guided Parameter Balancing for Merging Large Language Models (2025.findings-acl)

Copied to clipboard

Shuqi Liu, Han Wu, Bowei He, Xiongwei Han, Mingxuan Yuan, Linqi Song

LAVa: Layer-wise KV Cache Eviction with Dynamic Budget Allocation (2025.findings-emnlp)

Challenge:	Existing task vector-based model merging methods apply uniform coefficients across all parameters, overlooking varying parameter importance both within and across tasks.
Approach:	They propose a sensitivity-guided coefficient adjustment method that optimizes existing model merging techniques by operating at both task-specific and cross-task levels.
Outcome:	The proposed method outperforms existing model merging techniques on mistral 7B and LLaMA2 7B/13B models and enables them to outperformed specialized models.

Copied to clipboard

Yiqun Shen, Song Yuan, Zhengze Zhang, Xiaoliang Wang, Daxin Jiang, Nguyen Cam-Tu

Challenge:	Existing methods for cache compression are heuristic and lack dynamic budget allocation . cnn's john mccartney and johnny mccain present a new approach for cache eviction and dynamic budgets .
Approach:	They propose a unified framework for cache compression that minimizes information loss in transformer residual streams.
Outcome:	The proposed method consistently maintains top performance across task types.

Incorporating Object-Level Visual Context for Multimodal Fine-Grained Entity Typing (2023.findings-emnlp)

Copied to clipboard

Ying Zhang, Wenbo Fan, Kehui Song, Yu Zhao, Xuhui Sui, Xiaojie Yuan

Improving Zero-Shot Entity Linking Candidate Generation with Ultra-Fine Entity Type Information (2022.coling-1)

Challenge:	Experimental results show that fine-grained entity typing is superior to text-based methods.
Approach:	They propose a task called fine-grained entity typing to classify entities . they propose combining textual and visual contexts to capture fine-granular semantic information .
Outcome:	The proposed approach achieves superior classification performance compared to previous text-based approaches.

Copied to clipboard

Xuhui Sui, Ying Zhang, Kehui Song, Baohang Zhou, Guoqing Zhao, Xin Wei, Xiaojie Yuan

PrivLM-Bench: A Multi-level Privacy Evaluation Benchmark for Language Models (2024.acl-long)

Challenge:	Entity linking is a task of assigning entity mentions to referent entities in a knowledge base.
Approach:	They propose to use ultra-fine-grained type information to improve the generalization ability of EL models by utilizing a low-level task to extract ultra-finish entity type information.
Outcome:	The proposed model achieves state-of-the-art in the zero-shot entity linking task .

Copied to clipboard

Haoran Li, Dadi Guo, Donghao Li, Wei Fan, Qi Hu, Xin Liu, Chunkit Chan, Duanyi Yao, Yuan Yao, Yangqiu Song

A Span-based Multimodal Variational Autoencoder for Semi-supervised Multimodal Named Entity Recognition (2022.emnlp-main)

Challenge:	generative large language models (LLMs) exhibit surprising capability and integrate previous tasks into a unified text generation formulation.
Approach:	They propose a privacy evaluation benchmark to quantify the privacy leakage of language models.
Outcome:	The proposed benchmark compares PPLMs with different privacy implementations to find out how privacy leakage is handled.

Copied to clipboard

Baohang Zhou, Ying Zhang, Kehui Song, Wenya Guo, Guoqing Zhao, Hongbin Wang, Xiaojie Yuan

Explaining Length Bias in LLM-Based Preference Evaluations (2025.findings-emnlp)

Challenge:	Existing methods for named entity recognition on social media are not efficient for semi-supervised MNER because of the mismatch between the posted text and image.
Approach:	They propose a novel method to fuse the text and image features for multimodal named entity recognition under semi-supervised setting by exploiting modal-specific VAEs.
Outcome:	The proposed method outperforms baselines under supervised setting and improves performance with less labeled data than existing semi-supervised methods.

Copied to clipboard

Zhengyu Hu, Linxin Song, Jieyu Zhang, Zheyuan Xiao, Tianfu Wang, Zhengyu Chen, Nicholas Jing Yuan, Jianxun Lian, Kaize Ding, Hui Xiong

Is Parameter Collision Hindering Continual Learning in LLMs? (2025.coling-main)

Challenge:	a preference evaluation metric is often biased towards longer responses, revealing a reliability problem . a decomposition of the preference evaluation into two components is needed to understand this bias.
Approach:	They propose to decompose the preference evaluation metric into two key components . the first component is length-dependent and related to trustworthiness .
Outcome:	The proposed evaluation metric is based on two components: desirability and information mass.

Copied to clipboard

Shuo Yang, Kun-Peng Ning, Yu-Yang Liu, Jia-Yu Yao, Yong-Hong Tian, Yi-Bing Song, Li Yuan