Papers by Zirui Wang
Accelerating LLM Reasoning via Early Rejection with Partial Reward Modeling (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly relied upon for solving complex reasoning tasks. |
| Approach: | They propose to use Process Reward Models to scale inference time compute by generating in parallel . they propose to provide early signals that enable the rejection of suboptimal candidates before full generation of step is complete. |
| Outcome: | The proposed method achieves 1.4 – 9 reduction in inference FLOPs without degrading final performance. |
KIA: Knowledge-Guided Implicit Vision-Language Alignment for Chest X-Ray Report Generation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing reports on medical images and reports lack fine-grained cross-modal interaction, leading to insufficient understanding of detailed information. |
| Approach: | They propose a framework for establishing cross-modal semantic alignment in radiology report pairs using knowledge-guided implicit vision-language alignment. |
| Outcome: | KIA improves understanding of medical images and reports by incorporating medical knowledge to enhance pathological observation and anatomical landm. |
ClusterAttn: KV Cache Compression under Intrinsic Attention Clustering (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for sparse attention apply the same pattern across different attention heads and inputs, but fail to capture the intrinsic attention clustering in large language models. |
| Approach: | They propose a training-free sparse attention method that provides an efficient prompt cache compression scheme under intrinsic attention clustering for efficient LLM inference. |
| Outcome: | The proposed method reduces memory usage by 10%–65% and increases throughput by 2.6–4.8 times with no accuracy loss. |
DocMMIR: A Framework for Document Multi-modal Information Retrieval (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing multi-modal information retrieval models lack a comprehensive exploration of document-level retrieval . existing models suffer from the absence of cross-domain datasets at this granularity. |
| Approach: | They propose a multi-modal document retrieval framework to unify diverse document formats and domains with a comprehensive retrieval scenario. |
| Outcome: | The proposed framework improves document retrieval performance on a large multimodal dataset. |
Towards Human-aligned Evaluation for Linear Programming Word Problems (2024.lrec-main)
Copied to clipboard
Linzi Xing, Xinglu Wang, Yuxi Feng, Zhenan Fan, Jing Xiong, Zhijiang Guo, Xiaojin Fu, Rindra Ramamonjison, Mahdi Mostajabdaveh, Xiongwei Han, Zirui Zhou, Yong Zhang
| Challenge: | Existing evaluation methodologies for MWPs diverge from human judgment and face challenges in recognizing mathematically equivalent answers. |
| Approach: | They propose an evaluation metric rooted in graph edit distance that features benefits such as permutation invariance and more accurate program equivalence identification. |
| Outcome: | The proposed evaluation metric features benefits such as permutation invariance and more accurate program equivalence identification. |
MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing (2026.acl-industry)
Copied to clipboard
Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Zirui Tang, Boyu Niu, Yuanhong Zheng, Dongsheng Ma, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Fangdong Wang, Jingzhou Chen, Xiaomeng Zhao, Liqun Wei, Wei Li, Shasha Wang, RuiLiang Xu, Yuanyuan Cao, Lu Chen, Qianqian Wu, Huaiyu Gu, Lindong Lu, Dechen Lin, null Shenguanlin, Xuanhe Zhou, Linfeng Zhang, Yuhang Zang, Xiaoyi Dong, Jiaqi Wang, Bo Zhang, Lei Bai, Pei Chu, Weijia Li, Jiang Wu, Lijun Wu, Zhenxiang Li, Guangyu Wang, Zhongying Tu, Chao Xu, Kai Chen, Bowen Zhou, Dahua Lin, Wentao Zhang, Conghui He
| Challenge: | Document images are characterized by higher resolutions, denser content, and more complex structural layouts. |
| Approach: | They propose a 1.2B-parameter document parsing vision-language model that decouples layout analysis from local content recognition. |
| Outcome: | The proposed model surpasses general-purpose and domain-specific models on multiple benchmarks while maintaining significantly lower computational overhead. |
AssistedDS: Benchmarking How External Domain Knowledge Assists LLMs in Automated Data Science (2025.findings-emnlp)
Copied to clipboard
An Luo, Xun Xian, Jin Du, Fangqiao Tian, Ganghua Wang, Ming Zhong, Shengchun Zhao, Xuan Bi, Zirui Liu, Jiawei Zhou, Jayanth Srinivasa, Ashish Kundu, Charles Fleming, Mingyi Hong, Jie Ding
| Challenge: | Large language models (LLMs) have advanced the automation of data science workflows, yet it remains unclear whether they can critically leverage external domain knowledge as human data scientists do in practice. |
| Approach: | They propose a benchmark to evaluate how large language models handle external domain knowledge in tabular prediction tasks. |
| Outcome: | The proposed model evaluates whether it can critically leverage external domain knowledge as human data scientists do in practice. |
Unveiling Internal Reasoning Modes in LLMs: A Deep Dive into Latent Reasoning vs. Factual Shortcuts with Attribute Rate Ratio (2025.emnlp-main)
Copied to clipboard
Yiran Yang, Haifeng Sun, Jingyu Wang, Qi Qi, Zirui Zhuang, Huazheng Wang, Pengfei Ren, Jing Wang, Jianxin Liao
| Challenge: | Existing research in multi-hop questions has identified two reasoning modes, but has not investigated how these modes differ during inference. |
| Approach: | They propose a classification metric that compares latent reasoning and factual shortcuts in multi-hop questions. |
| Outcome: | The proposed metric achieves 90% accuracy on the proposed datasets and demonstrates effectiveness in RAG conflict scenarios. |
Taylor Unswift: Secured Weight Release for Large Language Models via Taylor Expansion (2024.emnlp-main)
Copied to clipboard
Guanchu Wang, Yu-Neng Chuang, Ruixiang Tang, Shaochen Zhong, Jiayi Yuan, Hongye Jin, Zirui Liu, Vipin Chaudhary, Shuai Xu, James Caverlee, Xia Hu
| Challenge: | Existing mechanisms compromise ownership rights or raise data privacy concerns . existing mechanisms compromise security of released large language models . |
| Approach: | They propose a TaylorMLP to preserve the ownership of large language models by transforming the weights of LLMs into Taylor-series parameters instead of releasing original weights . |
| Outcome: | The proposed model preserves ownership of large language models and prevents their abuse by adjusting the generation speed and causing low-speed token generation. |
Under the Shadow of Babel: How Language Shapes Reasoning in LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | linguistic relativity suggests that the structure of language shapes cognitive patterns . large language models internalize the habitual logical structures embedded in different languages, authors say . |
| Approach: | a study introduces a bilingual dataset for causal reasoning in Chinese and English. |
| Outcome: | a new study shows that large language models internalize reasoning biases shaped by language . the model internalizes language-specific preferences and rigidly applies them to atypical inputs, the study shows . |
Audio Jailbreak: An Open Comprehensive Benchmark for Jailbreaking Large Audio-Language Models (2026.acl-long)
Copied to clipboard
Zirui Song, Qian Jiang, Mingxuan Cui, Mingzhe Li, Lang Gao, Zeyu Zhang, Zixiang Xu, Yanbo Wang, Guangxian Ouyang, Zhenhao Chen, Xiuying Chen
| Challenge: | a recent study evaluated large audio-language models against jailbreak attacks . a new benchmark is being developed to evaluate LAM safety against jailbreaking attacks based on temporal and semantic nature of speech . |
| Approach: | They propose a benchmark to evaluate LAM jailbreak vulnerabilities in adversarial audio prompts . they use a dataset of 1,495 adversarials to evaluate their performance . |
| Outcome: | The proposed benchmark evaluates state-of-the-art LAMs against jailbreak attacks . it demonstrates that even small, semantically preserved perturbations can reduce safety . |
SAFER: A Controllable Safeguard for LLMs against Backdoor Attacks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing inference-time defenses lack explicit control over false acceptance rate (FAR) existing inference time defenses aim to mitigate poisoned inputs but lack explicit FAR control . |
| Approach: | They propose a framework that provides explicit control over false acceptance rate without prior knowledge of backdoor samples. |
| Outcome: | The proposed framework outperforms existing inference-time defenses on three benchmark datasets . it provides explicit and provable control over false acceptance rate without prior knowledge of backdoor samples . |
KV Cache Compression, But What Must We Give in Return? A Comprehensive Benchmark of Long Context Capable Approaches (2024.findings-emnlp)
Copied to clipboard
Jiayi Yuan, Hongyi Liu, Shaochen Zhong, Yu-Neng Chuang, Songchen Li, Guanchu Wang, Duy Le, Hongye Jin, Vipin Chaudhary, Zhaozhuo Xu, Zirui Liu, Xia Hu
| Challenge: | Long context capability is a crucial competency for large language models as it mitigates the human struggle to digest long-form texts. |
| Approach: | They propose to evaluate 10+ state-of-the-art approaches for long context-capable LLMs. |
| Outcome: | The proposed methods are compared against 10+ state-of-the-art approaches across seven categories of long context tasks. |
Efficient Meta Lifelong-Learning with Limited Memory (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing natural language learning models fail to continuously learn new tasks as they are re-trained throughout their lifetime. |
| Approach: | They propose a meta-lifelong framework that combines three common lifelong learning principles . they propose to store past examples in episodic memory and replay them at training and inference time . |
| Outcome: | The proposed framework achieves state-of-the-art performance using 1% memory size and narrows the gap with multi-task learning. |
MMAU: A Holistic Benchmark of Agent Capabilities Across Diverse Domains (2025.findings-naacl)
Copied to clipboard
Guoli Yin, Haoping Bai, Shuang Ma, Feng Nan, Yanchao Sun, Zhaoyang Xu, Shen Ma, Jiarui Lu, Xiang Kong, Aonan Zhang, Dian Ang Yap, Yizhe Zhang, Karsten Ahnert, Vik Kamath, Mathias Berglund, Dominic Walsh, Tobias Gindele, Juergen Wiest, Zhengfeng Lai, Xiaoming Simon Wang, Jiulong Shan, Meng Cao, Ruoming Pang, Zirui Wang
| Challenge: | Existing benchmarks focus on specific application scenarios, emphasizing task completion but failing to dissect the underlying skills that drive these outcomes. |
| Approach: | They propose a Massive Multitask Agent Understanding benchmark that evaluates LLMs across five domains and offline tasks. |
| Outcome: | The Massive Multitask Agent Understanding (MMAU) benchmark evaluates models across five domains including Tool-use, Directed Acyclic Graph (DAG) QA, Data Science and Machine Learning coding, Contest-level programming and Mathematics. |
DSQG-Syn: Synthesizing High-quality Data for Text-to-SQL Parsing by Domain Specific Question Generation (2025.findings-naacl)
Copied to clipboard
Shaoming Duan, Youxuan Wu, Chuanyi Liu, Yuhao Zhang, Zirui Wang, Peiyi Han, Shengyuan Yu, Liang Yan, Yingwei Liang
| Challenge: | Existing methods for generating SQL queries using natural language questions produce inconsistent NLQ-SQL pairs. |
| Approach: | They propose a text-to-SQL data synthesis framework that generates domain-relevant questions . they synthesize NLQ-SqL pairs that are domain-specific and intent-consistent . |
| Outcome: | The proposed method outperforms closed-source LLMs on the Text-to-SQL task. |
ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities (2025.findings-naacl)
Copied to clipboard
Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Haoping Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, Zirui Wang, Ruoming Pang
| Challenge: | Recent advances in large language models have led to a growing interest in tool assisted LLMs . toolSandbox includes stateful tool execution, implicit state dependencies between tools . |
| Approach: | a new tool-based evaluation tool is released to help LLMs evaluate their tool-use capabilities. a tool-driven evaluation tool includes stateful tool execution, implicit state dependencies between tools and a built-in user simulator. |
| Outcome: | the toolSandbox evaluation benchmark shows that open source and proprietary models have a performance gap . the benchmarks show that even the most capable LLMs are challenged by state dependent tasks . |
FineState-Bench: Benchmarking State-Conditioned Grounding for Fine-grained GUI State Setting (2026.findings-acl)
Copied to clipboard
| Challenge: | FineState-Bench evaluates whether an agent can correctly ground an instruction to the intended UI control and reach the exact target state. |
| Approach: | They propose a benchmark that evaluates whether an agent can correctly ground an instruction to the intended UI control and reach the exact target state. |
| Outcome: | The proposed benchmark evaluates whether an agent can ground an instruction to the intended UI control and reach the exact target state. |
Distantly Supervised Contrastive Learning for Low-Resource Scripting Language Summarization (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for code summarization are limited in resources and require atomic commands and category constraints to enhance code representations. |
| Approach: | They propose a framework that leverages limited atomic commands and category constraints to enhance code representations. |
| Outcome: | The proposed framework outperforms baseline methods in a number of domains and demonstrates superiority over competing frameworks. |
When Personalization Tricks Detectors: The Feature-Inversion Trap in Machine-Generated Text Detection (2026.acl-long)
Copied to clipboard
Lang Gao, Xuhui Li, Chenxi Wang, Mingzhe Li, Wei Liu, Zirui Song, Jinghui Zhang, Rui Yan, Preslav Nakov, Xiuying Chen
| Challenge: | Personalized MGT detection remains largely underexplored due to personalization challenges . large language models (LLMs) can imitate personal writing styles, but they can generate fake news and misinformation. |
| Approach: | They propose a benchmark to evaluate detector robustness under personalization . they attribute this limitation to a feature-inversion trap that flips the effect in personalized contexts . |
| Outcome: | The proposed framework predicts detector robustness under personalization with an 85% correlation to actual results. |
On Negative Interference in Multilingual Models: Findings and A Meta-Learning Treatment (2020.emnlp-main)
Copied to clipboard
| Challenge: | Modern multilingual models are trained on concatenated text from multiple languages in hopes of conferring benefits to each (positive transfer) however, recent work has shown that this approach can degrade performance on high-resource languages, a phenomenon known as negative interference. |
| Approach: | They propose a meta-learning algorithm that adds language-specific parameters as meta-parameters and trains them in a manner that explicitly improves shared layers’ generalization on all languages. |
| Outcome: | The proposed model improves cross-lingual transferability and generalization on all languages, and improves on the language-specific parameters. |
AEGIS: A Holistic Benchmark for Evaluating Forensic Analysis of AI-Generated Academic Images (2026.acl-long)
Copied to clipboard
Bo Zhang, Tzu-Yen Ma, Zichen Tang, Junpeng Ding, Zirui Wang, Yizhuo Zhao, Peilin Gao, Zijie Xi, Zixin Ding, Haiyang Sun, Haocheng Gao, Yuan Liu, Liangjia Wang, Yiling Huang, Yujie Wang, Yuyue Zhang, Ronghui Xi, Yuanze Li, Jiacheng Liu, Zhongjun Yang, Haihong E
| Challenge: | AEGIS examines whether current models can effectively audit AI-generated images in academic papers. |
| Approach: | They propose a holistic benchmark for forensic analysis of AI-Generated academic ImageS that reveals limitations in academic image forensics. |
| Outcome: | AEGIS compared with existing benchmarks on seven academic categories and features key advances in forensic analysis. |
Word Form Matters: LLMs’ Semantic Reconstruction under Typoglycemia (2025.findings-acl)
Copied to clipboard
| Challenge: | Typoglycemia is a phenomenon where people can read words even when the middle letters of the words are scrambled. |
| Approach: | They propose a reliable metric to quantify the degree of semantic reconstruction and validate its effectiveness. |
| Outcome: | The proposed metric quantifies the degree of semantic reconstruction and validates its effectiveness. |
Evaluating and Mitigating Object Hallucination in Large Vision-Language Models: Can They Still See Removed Objects? (2025.naacl-long)
Copied to clipboard
| Challenge: | LVLMs often mistakenly determine objects as present in images where they do not exist . authors propose a new benchmark to evaluate object hallucinations by removing objects from images and asking the model whether it can still see the removed objects. |
| Approach: | They propose a benchmark to evaluate object hallucinations by removing objects from images . they propose oDPO, a direct preference optimization objective based on visual objects . |
| Outcome: | The proposed benchmark reduces the likelihood of object hallucinations by removing objects from images and asking the model whether it can still see the removed objects. |
Seek-and-Solve: Benchmarking MLLMs for Visual Clue-Driven Reasoning in Daily Scenarios (2026.findings-acl)
Copied to clipboard
Xiaomin Li, Tala Wang, Zichen Zhong, Ying Zhang, Zirui Zheng, Takashi Isobe, Dezhuang Li, Huchuan Lu, You He, Xu Jia
| Challenge: | Existing benchmarks focus on evaluating MLLMs’ pre-existing knowledge or perceptual understanding, often neglecting the critical capability of reasoning. |
| Approach: | They propose a benchmark designed for visual clue-driven reasoning in daily scenarios that combines rigorous grounding in authentic daily activities and challenging query design that necessitates more than surface-level perception. |
| Outcome: | The proposed benchmark identifies visual clues and their ability to provide robust reasoning in daily scenarios. |
Re3: Relevance & Recency Retrieval for Mitigating Temporal Hallucination (2026.acl-long)
Copied to clipboard
| Challenge: | Existing retrievers suffer from temporal-semantic misalignment and outdated-document interference . Existing frameworks suffer from both temporal validity and outdated factual versions . |
| Approach: | They propose a framework that mitigates temporal hallucinations by embedding heterogeneous temporal signals into the semantic space to ensure retrieval fidelity. |
| Outcome: | Experiments show that Re3 outperforms baselines by 9.7% in generation accuracy . the framework outperformed strongest baselines on challenging dynamic tasks . |
Example Quality Matters: Multi-Aspects Example Augmentation for Private Library Programming (2026.acl-long)
Copied to clipboard
Yuhao Li, Haifeng Sun, Xuesong Zhang, Shu Yao, Haoyu Zheng, Yvchuan Wang, Huazheng Wang, Zirui Zhuang, Qi Qi, Jianxin Liao, Jingyu Wang
| Challenge: | Existing approaches to code generation fail to consider the quality of retrieved examples. |
| Approach: | They propose a retrieval-augmented generation method that combines existing API examples to improve complexity and readability. |
| Outcome: | The proposed method achieves up to 22% accuracy improvement over baseline methods. |
ModularMoE: Fast LLM Customization with Parameter-Sharing Mixture-of-Experts for Low-Resource Settings (2026.findings-acl)
Copied to clipboard
Jiaxing Liu, Qi Qi, Haifeng Sun, Dunjun Li, Zirui Zhuang, Bo He, Xiang Yang, Cong Liu, Jianxin Liao, Jingyu Wang
| Challenge: | Large Language Models impose significant computational and storage burdens on personal devices . existing customization approaches incur excessive computational costs or lead to suboptimal performance . |
| Approach: | They propose a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment. |
| Outcome: | The proposed training framework outperforms state-of-the-art training frameworks at the same sparsity level while delivering up to 2.71 inference speedup. |