Papers by Ziyi Zhang
DentalGPT: Incentivizing Multimodal Reasoning in Dentistry (2026.findings-acl)
Copied to clipboard
Zhenyang Cai, Jiaming Zhang, Junjie Zhao, Ziyi Zeng, Yanchao Li, Liang Jingyi, Junying Chen, Yunjin Yang, Jiajun You, Shuzhi Deng, null Xieruiqiii, Yuanting Chen, Xiangyi Feng, Jianquan Li, Liangyi Chen, Junwen Wang, Shan Jiang, Benyou Wang
| Challenge: | Current multimodal large language models (MLLMs) show limited understanding of dental images. |
| Approach: | They propose a dental-specialized multimodal large language model trained via staged multimodal alignment and reinforcement learning. |
| Outcome: | The proposed model outperforms state-of-the-art models on disease classification and dental VQA tasks. |
FC-Attack: Jailbreaking Multimodal Large Language Models via Auto-Generated Flowcharts (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent research shows that multimodal large language models are vulnerable to jailbreak attacks . |
| Approach: | They propose a jailbreak attack method based on auto-generated flowcharts . the flowchartings are then combined with a benign textual prompt to execute the attack . |
| Outcome: | The proposed method achieves an attack success rate of up to 96% via images and 78% via videos across multiple MLLMs. |
TriEx: A Game-based Tri-View Framework for Explaining Internal Reasoning in Multi-Agent LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing explainability methods for large language models have been limited in capturing interaction-dependent belief dynamics and multi-agent reasoning. |
| Approach: | They propose a tri-view explainability framework that instruments sequential decision making with aligned artifacts. |
| Outcome: | The proposed framework enables analysis of explanation faithfulness, belief dynamics, and evaluator reliability, revealing systematic mismatches between what agents say, what they believe, and what they do. |
Self-Training with Pseudo-Label Scorer for Aspect Sentiment Quad Prediction (2024.acl-long)
Copied to clipboard
| Challenge: | Aspect Sentiment Quad Prediction (ASQP) aims to predict all quads (aspect term, aspect category, opinion term, sentiment polarity) for a given review. |
| Approach: | They propose a self-training framework with a pseudo-label scorer to assess the match between reviews and their pseudo-labels and train a generative model on it. |
| Outcome: | The proposed framework can predict all quads (aspect term, aspect category, opinion term, sentiment polarity) for a given review, and it can significantly improve self-training. |
LongVideoAgent: Multi-Agent Reasoning with Long Videos (2026.acl-long)
Copied to clipboard
| Challenge: | a key emerging challenge is robust long video understanding, authors say . current methods compress content into lossy summaries or rely on limited toolsets . |
| Approach: | They propose a multi-agent framework where a master LLM coordinates a grounding agent and a vision agent to extract targeted textual observations. |
| Outcome: | The proposed model outperforms strong non-agent baselines on episode-level datasets . the proposed model significantly outperformed existing models on other datasets. |
A Unified Supervised and Unsupervised Dialogue Topic Segmentation Framework Based on Utterance Pair Modeling (2025.naacl-long)
Copied to clipboard
| Challenge: | Unsupervised methods for dialogue topic segmentation are difficult to surpass due to short sentences, serious references and non-standard language. |
| Approach: | They propose a method to divide a dialogue into different topic paragraphs to better understand its structure and content. |
| Outcome: | The proposed method achieves the best results on multiple benchmark datasets across different scenarios. |
Are We in the AI-Generated Text World Already? Quantifying and Monitoring AIGT on Social Media (2025.acl-long)
Copied to clipboard
| Challenge: | Social media platforms are experiencing a growing presence of AI-Generated Texts (AIGTs) however, the misuse of AIGTs could have profound implications for public opinion . |
| Approach: | They collect a dataset with 2.4M posts from 3 major social media platforms . they then construct a diverse dataset to train and evaluate AIGT detectors . |
| Outcome: | The proposed dataset analyzes 2.4M posts from 3 major social media platforms from 2022 to 2024 . it finds that Medium and Quora show marked increases in AAR . |
A Framework of Knowledge Graph-Enhanced Large Language Model Based on Question Decomposition and Atomic Retrieval (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to enhance LLMs with knowledge graphs have limited results . knowledge graph question answering (KGQA) provides interpretable reasoning for large language models . |
| Approach: | They propose a framework for KG-enhanced LLM based on question decomposition and atomic retrieval . they propose question decomposing tree as framework for LLM reasoning . |
| Outcome: | The proposed framework outperforms existing reasoning-based baselines on KGQA datasets. |
Trajectory2Task: Training Robust Tool-Calling Agents with Synthesized Yet Verifiable Data for Complex User Intents (2026.acl-long)
Copied to clipboard
Ziyi Wang, Yuxuan Lu, Yimeng Zhang, Pei Chen, Ziwei Dong, Jing Huang, Jiri Gesi, Xianfeng Tang, Chen Luo, Qun Liu, Yisi Sang, Hanqing Lu, Manling Li, Jin Lai, Dakuo Wang
| Challenge: | Tool-calling agents are increasingly deployed in real-world customer-facing workflows . but most studies on tool-callers focus on idealized settings with general, fixed, and well-specified tasks. |
| Approach: | They propose a tool-calling agent-based data pipeline that converts trajectories into user-facing tasks with controlled intent adaptations. |
| Outcome: | The proposed pipeline can be used to study tool use under three scenarios. |
Attribution-Based Analysis and Optimization of Modular Agentic Workflows (2026.findings-acl)
Copied to clipboard
Yingxuan Yang, Bo Huang, Siyuan Qi, Chao Feng, Haoyi Hu, Yuxuan Zhu, Jinbo Hu, Haoran Zhao, Ziyi He, Xiao Liu, ZongYu Wang, Muning Wen, Lin Qiu, Xuezhi Cao, Xunliang Cai, Yong Yu, Weinan Zhang
| Challenge: | Large Language Models (LLMs) have driven the rise of agentic workflows . yet, how can we attribute performance gains to individual upgrades and their interactions? |
| Approach: | They propose a game-theoretic framework that models component upgrades as players and evaluates component coalitions to compute Shapley values. |
| Outcome: | The proposed framework provides interaction-aware attribution and recommendation for model allocation under a fixed workflow structure. |
Beyond Static Persona Consistency: Dynamic Persona Coherence in LLM Role-Playing (2026.acl-long)
Copied to clipboard
Yirui QI, Xiaoming Zhang, Ruilin Zeng, Mengyao Liu, Ziyi Zhou, Dezhuang Miao, Bingyu Yan, Zhenyu Guan
| Challenge: | Existing LLMs conflate identity consistency with emotional rigidity . Existing models exhibit either robotic repetition or persona drift . |
| Approach: | They propose a framework that decouples Identity-Layer Stability from Adaptive-Layer Appropriateness to achieve persona coherence repair. |
| Outcome: | Experiments on GPT-4o, Claude-3.5-Sonnet, and DeepSeek-V3.2 show consistent improvements (+16–84% gains) |
Are Your LLMs Capable of Stable Reasoning? (2025.findings-acl)
Copied to clipboard
Junnan Liu, Hongwei Liu, Linchen Xiao, Ziyi Wang, Kuikun Liu, Songyang Gao, Wenwei Zhang, Songyang Zhang, Kai Chen
| Challenge: | Existing evaluation protocols and metrics do not capture the full spectrum of LLM capabilities, especially in complex reasoning tasks. |
| Approach: | They propose a new evaluation metric that continuously assesses model performance across multiple sampling attempts, quantifying both the model’s potential capabilities and operational consistency. |
| Outcome: | The proposed evaluation metric measures model performance across multiple sampling attempts and provides comprehensive insights into their potential capabilities and operational consistency. |
UniSumm and SummZoo: Unified Model and Diverse Benchmark for Few-Shot Summarization (2023.acl-long)
Copied to clipboard
| Challenge: | a new benchmark summarization model is being developed to train few-shot summarizers . a large number of summarizing tasks are required to perform well in heterogeneous datasets. |
| Approach: | They propose a few-shot summarization model pre-trained with multiple summarizing tasks . they propose 'uniSumm' to be prefix-tuned to excel at any few-shot summarisation task . |
| Outcome: | The proposed model outperforms baseline models under automatic and human evaluations and achieves comparable results in human evaluation. |
TwinVoice: A Multi-dimensional Benchmark Towards Digital Twins via LLM Persona Simulation (2026.findings-acl)
Copied to clipboard
Bangde Du, Minghao Guo, Songming He, Ziyi Ye, Xi Zhu, Weihang Su, Shuqi Zhu, Yujia Zhou, Yongfeng Zhang, Qingyao Ai, Yiqun Liu
| Challenge: | Existing studies show that advanced LLMs produce text indistinguishable from human writing. |
| Approach: | They propose a benchmark to assess persona simulation across diverse contexts by decomposing the evaluation into six fundamental capabilities including opinion consistency, memory recall, logical reasoning, persona tone, and syntactic style. |
| Outcome: | The proposed model achieves moderate accuracy but falls short of the basic capabilities needed to simulate personas in real-world contexts. |
MACSum: Controllable Summarization with Mixed Attributes (2023.tacl-1)
Copied to clipboard
Yusen Zhang, Yang Liu, Ziyi Yang, Yuwei Fang, Yulong Chen, Dragomir Radev, Chenguang Zhu, Michael Zeng, Rui Zhang
| Challenge: | Existing work on controllable summarization with mixed attributes lacks designated annotations. |
| Approach: | They propose a human-annotated summarization benchmark for controllable summarizing with mixed attributes based on news and dialogue sources . |
| Outcome: | The proposed dataset contains human-annotated summarization datasets with mixed attributes . hard prompt models yield the best performance on most metrics and human evaluations . mixed-attribute control is still challenging for summarizing tasks . |
Language Model as Planner and Formalizer under Constraints (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have been widely used in planning but lack interpretability and control. |
| Approach: | They propose to augment widely used planning benchmarks with manually annotated, fine-grained, and rich natural language constraints spanning four formally defined categories. |
| Outcome: | The proposed model outperforms existing models in 4 state-of-the-art reasoning LLMs, 4 formal languages, and 4 datasets. |
KG-RAG: Enhancing GUI Agent Decision-Making via Knowledge Graph-Driven Retrieval-Augmented Generation (2025.emnlp-main)
Copied to clipboard
Ziyi Guan, Jason Chun Lok Li, Zhijian Hou, Pingping Zhang, Donglai Xu, Yuzhi Zhao, Mengyang Wu, Jinpeng Chen, Thanh-Toan Nguyen, Pengfei Xian, Wenao Ma, Shengchao Qin, Graziano Chesi, Ngai Wong
| Challenge: | Recent advances in GUI agents have limited app-specific knowledge of complex mobile tasks. |
| Approach: | They propose a Knowledge Graph-driven Retrieval-Augmented Generation framework that transforms fragmented UTGs into structured vector databases for efficient real-time retrieval. |
| Outcome: | The proposed framework outperforms existing methods in a 75.8% success rate and 84.6% decision accuracy test across mobile apps. |
Utility-Oriented Visual Evidence Selection for Multimodal Retrieval-Augmented Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for multimodal retrieval-augmented generation rely on semantic relevance or surface-level similarity, which are often misaligned with the actual utility of visual evidence for downstream reasoning. |
| Approach: | They propose a latent notion of evidence usefulness and propose 'surrogate-accelerated' framework that efficiently estimates evidence utility using lightweight multimodal models. |
| Outcome: | The proposed framework outperforms state-of-the-art models while achieving substantial reductions in computational cost. |
RAt: Injecting Implicit Bias for Text-To-Image Prompt Refinement Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Text-to-image prompt refinement (T2I-Refine) aims to rephrase or extend an input prompt with more descriptive details that can be leveraged to generate images with higher quality. |
| Approach: | They develop an adversarial prompt attacking framework that implicitly attacks input prompts with intentional adversarials to generate images with higher quality. |
| Outcome: | The proposed framework can implicitly attack input prompts with implicit concept biases to generate images with higher quality and explicit visual bias towards the target group. |
AiM: Taking Answers in Mind to Correct Chinese Cloze Tests in Educational Applications (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods to correct handwritten assignments are to use OCR to recognize characters and compare them to answers. |
| Approach: | They propose a multimodal approach to correct handwritten Chinese characters by combining the visual information of students' handwriting with the encoded representations of answers. |
| Outcome: | The proposed model outperforms OCR-based methods by a large margin. |
SearchGym: Bootstrapping Real-World Search Agents via Cost-Effective and High-Fidelity Environment Simulation (2026.acl-long)
Copied to clipboard
Xichen Zhang, Ziyi He, Yinghao Zhu, Sitong Wu, Shaozuo Yu, Meng Chu, Wenhu Zhang, Haoru Tan, Jiaya Jia
| Challenge: | Search agents are a pivotal paradigm for solving open-ended, knowledge-intensive reasoning tasks. |
| Approach: | They propose a search agent simulation environment that bootstraps robust search agents using Reinforcement Learning. |
| Outcome: | The proposed model outperforms the web-enhanced ASearcher model by 10.6%. |
SR-LLM: Rethinking the Structured Representation in Large Language Model (2025.acl-long)
Copied to clipboard
Jiahuan Zhang, Tianheng Wang, Ziyi Huang, Yulong Wu, Hanqing Wu, DongbaiChen DongbaiChen, Linfeng Song, Yue Zhang, Guozheng Rao, Kaicheng Yu
| Challenge: | Structured representations have long been pivotal in computational linguistics, but their role remains ambiguous in the Large Language Models (LLMs) era. |
| Approach: | They propose a framework that integrates structured representations into LLMs from training-free and training-dependent perspectives. |
| Outcome: | The proposed framework integrates structured representations through natural language descriptions in LLM prompts while augmenting the model’s inference capability through fine-tuning on linguistically described structured representation. |
SACTOR: LLM-Driven Correct and Idiomatic C to Rust Translation with Static Analysis and FFI-Based Verification (2026.acl-long)
Copied to clipboard
Tianyang Zhou, Ziyi Zhang, Haowen Lin, Somesh Jha, Mihai Christodorescu, Kirill Levchenko, Varun Chandrasekaran
| Challenge: | Large language models (LLMs) have shown promise in producing idiomatic translations, but offer no correctness guarantees. |
| Approach: | They propose a C-to-Rust translation tool that uses an initial "unidiomatic" translation followed by an "idiomatic refinement" they evaluate SACTOR on 200 programs from two datasets and two more complex scenarios . |
| Outcome: | The proposed tool delivers high end-to-end correctness and produces safe, idiomatic Rust with up to 7 fewer Clippy warnings. |
Evo-Attacker: Memory-Augmented Reinforcement Learning for Long-Horizon Tool Attacks on LLM-MAS (2026.acl-long)
Copied to clipboard
| Challenge: | Existing tool attacks are limited by domain specificity or fixed and static templates. |
| Approach: | They propose an attack-based memory-augmented reinforcement learning process that constructs a dynamic attack memory and employs deliberative reasoning to retrieve adversarial patterns. |
| Outcome: | Evo-Attacker outperforms baselines in the long-horizon credit assignment challenge. |
P-QuASAR: A Unified Probabilistic Framework for Holistic Patent Quality Assessment and Refinement (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for assessing patent quality rely on modular pipelines or generic detectors, resulting in fragmented decisions and limited integration across quality dimensions. |
| Approach: | They propose a probabilistic framework that represents patent specifications as Quality Graphs. |
| Outcome: | The proposed framework outperforms existing methods on 500 patents against seven baselines. |
Unraveling the Mystery of Artifacts in Machine Generated Text (2022.lrec-1)
Copied to clipboard
| Challenge: | Recent studies show that human-written text is not distinguishable from synthetic text because of semantic errors or logical contradictions. |
| Approach: | They propose to analyze the forms of artifacts left by neural Text Generation Models by corrupting texts and replacing them with linguistic or statistical features. |
| Outcome: | The proposed method replaces text with linguistic or statistical features and improves the accuracy of the model. |
TiMem: Temporal-Hierarchical Memory Consolidation for Long-Horizon Conversational Agents (2026.findings-acl)
Copied to clipboard
Kai Li, Xuanqing Yu, Ziyi Ni, Yi Zeng, Yao Xu, Zheqing Zhang, Xin Li, Jitao Sang, Xiaogang Duan, Xuelei Wang, Chengbao Liu, Jie Tan
| Challenge: | Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, leading to fragmented memories and unstable long-horizon personalization. |
| Approach: | They propose a temporal–hierarchical memory framework that organizes conversations through a Temporal Memory Tree. |
| Outcome: | The proposed framework outperforms baselines while reducing the recalled memory length by 52.20%. |
Mitigating Training Imbalance in LLM Fine-Tuning via Selective Parameter Merging (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies suggest that the order of training samples can affect model performance, but this is not the case. |
| Approach: | They propose to merge supervised fine-tuning models with different data orders to mitigate this imbalance by parameter merging. |
| Outcome: | The proposed method outperforms the weighted-average method on five datasets. |