Papers by Yutong Zhang
AgentInit: Initializing LLM-based Multi-Agent Systems via Diversity and Expertise Orchestration for Effective and Efficient Collaboration (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing MAS initialization methods do not fully account for the collaborative needs of the generated agents in subsequent stages. |
| Approach: | They propose to use a Natural Language to Format mechanism to optimize the structure of agent teams and incorporate a natural language to format mechanism to ensure consistency and standardization. |
| Outcome: | The proposed method outperforms state-of-the-art initialization methods and pre-defined strategies across various frameworks and tasks while reducing token consumption. |
MMRC: A Large-Scale Benchmark for Understanding Multimodal Large Language Model in Real-World Conversation (2025.acl-long)
Copied to clipboard
Haochen Xue, Feilong Tang, Ming Hu, Yexin Liu, Qidong Huang, Yulong Li, Chengzhi Liu, Zhongxing Xu, Chong Zhang, Chun-Mei Feng, Yutong Xie, Imran Razzak, Zongyuan Ge, Jionglong Su, Junjun He, Yu Qiao
| Challenge: | Existing multimodal large language models lack the ability to memorize, recall, and reason in sustained interactions. |
| Approach: | They propose a multimodal real-world conversation benchmark for evaluating open-ended abilities of multimodal large language models. |
| Outcome: | The proposed benchmarks show that the models perform better in open-ended conversations. |
PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations (2026.acl-long)
Copied to clipboard
Yuhe Wu, Guangyu Wang, Yuran Chen, Jiatong Zhang, Yutong Zhang, Yujie Chen, Jiaming Shang, Guang Zhang, Zhuang Liu
| Challenge: | Existing benchmarks for hallucination evaluation rely on mixed queries and posterior evaluation, which quantifies hallucinosity severity but offers limited insight into where and why they occur. |
| Approach: | They propose a controlled benchmark that disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors. |
| Outcome: | The proposed model disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors. |
“I’ve Decided to Leak”: Probing Internals Behind Prompt Leakage Intents (2025.emnlp-main)
Copied to clipboard
Jianshuo Dong, Yutong Zhang, Liu Yan, Zhenyu Zhong, Tao Wei, Ke Xu, Minlie Huang, Chao Zhang, Han Qiu
| Challenge: | Large language models (LLMs) exhibit prompt leakage vulnerabilities, raising intellectual property and confidentiality concerns. |
| Approach: | They use probing techniques to capture LLMs’ intent-related internal representations and show that they internalize prompt leakage intents in their hidden states before generating tokens. |
| Outcome: | The proposed probes achieve 90%+ AUROC across all tested models, even when applied to new system prompts and attacks. |
DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation (2026.findings-acl)
Copied to clipboard
Yutong Song, Jiang Wu, Kazi Shaharair Sharif, Pengfei Zhang, Wenjun Huang, Honghui Xu, Nikil Dutt, Amir M. Rahmani
| Challenge: | Simulating dementia patients with large language models is challenging due to the need to model cognitive impairment, emotional dynamics, and nonverbal behaviors over long conversations. |
| Approach: | They propose an expert-guided dementia dialogue agent for multi-turn patient simulation . they introduce a framework that trains a single LLM to jointly generate reasoning traces, patient utterances, and aligned behavioral actions . |
| Outcome: | The proposed model outperforms baselines in persona fidelity, clinical validity, and educational effectiveness. |
AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for MAS suffer from high token consumption and inefficiency due to frequent generation and communication among multiple agents. |
| Approach: | They propose a multi-agent system based on large language models that identifies redundant agents and communication across different communication rounds by optimizing the adjacency matrices of the communication graphs and eliminates them to enhance both token efficiency and task performance. |
| Outcome: | The proposed method reduces prompt token consumption and completion token consumption by 18.4% and improves task performance by 1.14. |
DeReA: Improving Idiom Translation with Detect-Retrieve-Arbitrate Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to idiom translation are limited by the constraints of static parametric memory and retrieval noise . idiomatic expressions are non-compositional units where figurative meanings diverge from literal interpretations . |
| Approach: | They propose a detect-retrieve-arbitrate framework that detects idiomatic spans by reasoning over semantic conflicts between literal and contextual meanings. |
| Outcome: | The proposed framework improves GPT-5-mini and Emerging Slang datasets on various model scales. |
Reframe Your Life Story: Interactive Narrative Therapist and Innovative Moment Assessment with Large Language Models (2025.emnlp-main)
Copied to clipboard
Yi Feng, Jiaqi Wang, Wenxuan Zhang, Zhuang Chen, Shen Yutong, Xiyao Xiao, Minlie Huang, Liping Jing, Jian Yu
| Challenge: | Existing approaches to mental health support lack realism and capture therapeutic progression over time. |
| Approach: | They propose a framework that simulates expert narrative therapists by planning therapeutic stages, guiding reflection levels, and generating contextually appropriate responses through retrieval-augmentation. |
| Outcome: | The proposed framework outperforms standard methods in quality and depth on 260 simulated clients and 230 human participants. |
Generative Interfaces for Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models are increasingly seen as assistants, copilots, and consultants . however, their linear request-response format often makes interactions inefficient in multi-turn tasks . |
| Approach: | They propose a paradigm in which large language models respond to user queries by generating user interfaces that enable more adaptive and interactive engagement. |
| Outcome: | The proposed paradigm outperforms traditional chat-based interfaces in many tasks and interaction patterns. |
AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time (2025.emnlp-main)
Copied to clipboard
Junyu Zhang, Runpei Dong, Han Wang, Xuying Ning, Haoran Geng, Peihao Li, Xialin He, Yutong Bai, Jitendra Malik, Saurabh Gupta, Huan Zhang
| Challenge: | Existing monotonic scaling methods for large reasoning models are not reliable. |
| Approach: | They propose a universal framework for modulating reasoning progress in large reasoning models at test time. |
| Outcome: | The proposed framework unifies and generalizes existing monotonic scaling methods and enables flexible and dense slow-to-fast reasoning modulation. |
WorldCuisines: A Massive-Scale Benchmark for Multilingual and Multicultural Visual Question Answering on Global Cuisines (2025.naacl-long)
Copied to clipboard
Genta Indra Winata, Frederikus Hudi, Patrick Amadeus Irawan, David Anugraha, Rifki Afina Putri, Wang Yutong, Adam Nohejl, Ubaidillah Ariq Prathama, Nedjma Ousidhoum, Afifa Amriani, Anar Rzayev, Anirban Das, Ashmari Pramodya, Aulia Adila, Bryan Wilie, Candy Olivia Mawalim, Cheng Ching Lam, Daud Abolade, Emmanuele Chersoni, Enrico Santus, Fariz Ikhwantri, Garry Kuwanto, Hanyang Zhao, Haryo Akbarianto Wibowo, Holy Lovenia, Jan Christian Blaise Cruz, Jan Wira Gotama Putra, Junho Myung, Lucky Susanto, Maria Angelica Riera Machin, Marina Zhukova, Michael Anugraha, Muhammad Farid Adilazuarda, Natasha Christabelle Santosa, Peerat Limkonchotiwat, Raj Dabre, Rio Alexander Audino, Samuel Cahyawijaya, Shi-Xiong Zhang, Stephanie Yulia Salim, Yi Zhou, Yinxuan Gui, David Ifeoluwa Adelani, En-Shiun Annie Lee, Shogo Okada, Ayu Purwarianti, Alham Fikri Aji, Taro Watanabe, Derry Tanti Wijaya, Alice Oh, Chong-Wah Ngo
| Challenge: | Vision Language Models struggle with cultural-specific knowledge, especially in languages other than English and in underrepresented cultural contexts. |
| Approach: | They propose a visual question answering (VQA) dataset with text-image pairs across 30 languages and dialects and a training dataset. |
| Outcome: | The proposed model performs better with correct location context, but struggles with adversarial contexts and predicting specific regional cuisines and languages. |
MemCoRL: Alternating Co-Optimization of Memory Retrieval and Utilization via Collaborative Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing research has proposed external memory modules for Large Language Models (LLMs) to overcome the limitations of finite input length and obtain contextual memory beyond the current input. |
| Approach: | They propose a two-stage alternating co-optimization reinforcement learning method that optimizes evidence retrieval and utilization using semantic feedback and rewards. |
| Outcome: | The proposed method outperforms baselines on lexical overlap and semantic similarity metrics, confirming the co-optimization in memory retrieval and memory utilization. |
Revisiting Commonsense Reasoning in Machine Translation: Training, Evaluation and Challenge (2023.acl-long)
Copied to clipboard
| Challenge: | CR is the ability to understand and navigate the world using basic knowledge and understanding shared by most people. |
| Approach: | They propose to incorporate pretrained knowledge into NMT models and use them as robust testbeds for investigating CR in NMT. |
| Outcome: | The proposed method improves the training of NMT models with high CR abilities and provides accurate evaluation metrics. |
MMCLIP: Cross-Modal Attention Masked Modelling for Medical Language-Image Pre-Training (2026.acl-long)
Copied to clipboard
| Challenge: | Existing vision-and-language pretraining methods face challenges in reconstructing pathological features due to limited data. |
| Approach: | They propose a method that uses masked modeling to enhance visual and linguistic learning. |
| Outcome: | MMCLIP integrates unpaired data through disease-kind prompts to achieve state-of-the-art performance in zero-shot and fine-tuning across five benchmarks. |
MASSW: A New Dataset and Benchmark Tasks for AI-Assisted Scientific Workflows (2025.findings-naacl)
Copied to clipboard
Xingjian Zhang, Yutong Xie, Jin Huang, Jinge Ma, Zhaoying Pan, Qijia Liu, Ziyang Xiong, Tolga Ergen, Dongsub Shim, Honglak Lee, Qiaozhu Mei
| Challenge: | Scientific innovation is driven by detailed workflows, which include critical steps such as contextualizing literature, generating ideas, validating ideas, and planning new research. |
| Approach: | They propose to use large language models to extract five key aspects from scientific publications to optimize scientific workflows. |
| Outcome: | The proposed dataset includes more than 152,000 peer-reviewed publications from 17 leading computer science conferences spanning the past 50 years. |
DiagnosisArena: Benchmarking Diagnostic Reasoning for Large Language Models (2026.findings-acl)
Copied to clipboard
Yakun Zhu, Zhongzhen Huang, Linjie Mu, Yutong Huang, Wei Nie, Jiaji Liu, Shaoting Zhang, Pengfei Liu, Xiaofan Zhang
| Challenge: | Existing medical benchmarks for diagnostic reasoning are limited in their ability to perform complex tasks. |
| Approach: | They propose to benchmark diagnostic capabilities of large language models to assess their accuracy and generalization bottlenecks. |
| Outcome: | The proposed model achieves 45.82%, 31.09%, and 17.79% accuracy, compared to current models, o3-mini, e1 and DeepSeek-R1 . |
TasTe: Teaching Large Language Models to Translate through Self-Reflection (2024.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to enhance LLMs' performance in machine translation are unable to fully exploit their instruction-following capabilities. |
| Approach: | They propose a framework for translating through self-reflection that involves two stages of inference . they propose to use the framework to refine LLMs' preliminary translations . |
| Outcome: | The proposed framework can produce translation outputs that match the quality of NMT systems. |
Revisiting the Reliability of Language Models in Instruction-Following (2026.acl-long)
Copied to clipboard
| Challenge: | Several benchmarks have been proposed to measure instruction-following accuracy, but these scores do not translate to reliable services in real-world use. |
| Approach: | They propose a new metric reliable@k and develop an automated pipeline to generate cousin prompts. |
| Outcome: | The proposed model can be instantiated with cousin prompts and generates high-quality cousin prompt data. |
KQA Pro: A Dataset with Explicit Compositional Programs for Complex Question Answering over Knowledge Base (2022.acl-long)
Copied to clipboard
Shulin Cao, Jiaxin Shi, Liangming Pan, Lunyiu Nie, Yutong Xiang, Lei Hou, Juanzi Li, Bin He, Hanwang Zhang
| Challenge: | Existing benchmarks for Complex KBQA lack compositional reasoning capabilities . Existing methods for Complex questions are poor in diversity or scale . |
| Approach: | They propose a compositional programming language to represent the reasoning process of complex questions. |
| Outcome: | The proposed dataset includes around 120K diverse natural language questions . it provides a compositional and interpretable programming language to represent the reasoning process of complex questions based on the proposed model . |
CultureBank: An Online Community-Driven Knowledge Base Towards Culturally Aware Language Technologies (2024.findings-emnlp)
Copied to clipboard
| Challenge: | CultureBank is a knowledge base built upon users’ self-narratives with 12K cultural descriptors sourced from TikTok and 11K from Reddit. |
| Approach: | They construct a pipeline to construct cultural knowledge bases from different online communities on a massive scale. |
| Outcome: | The proposed pipeline improves cultural awareness of language models by evaluating them on two cultural tasks in a zero-shot setting. |
Learning From the Source Document: Unsupervised Abstractive Summarization (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for abstractive summarization are under supervised training, but obtaining high-quality and large-scale datasets for supervised learning is laboriously difficult. |
| Approach: | They propose an unsupervised method that leverages contrastive learning to generate summaries by rewriting and paraphrasing the source documents to generate good summary. |
| Outcome: | The proposed method outperforms baseline methods on extensive experiments on source documents and fake documents. |
MedMCP-Calc: Benchmarking LLMs for Realistic Medical Calculator Scenarios via MCP Integration (2026.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks focus on static single-step calculations with explicit instructions. |
| Approach: | They propose a benchmark for evaluating medical calculators in realistic scenarios . they use 118 scenario tasks across 4 clinical domains to evaluate medical calculator performance . |
| Outcome: | The first benchmark for evaluating medical calculators in realistic scenarios is released . it features 118 scenario tasks across 4 clinical domains and is based on a model context protocol integration. |
Rhombus: Incentivizing Coordination in Parallel Thinking through Reinforcement Learning (2026.findings-acl)
Copied to clipboard
Ziyuan Nan, Qi Yi, Di Huang, Yutong Wu, Guanhua Huang, Xue Gong, Kejiao Li, Yuhao Jiang, Chenchen Zhang, Zenan Xu, Xing Hu, Bo Zhou
| Challenge: | Parallel thinking is a promising avenue for scaling test-time compute in Large Language Models . however, coordinating the exploration and aggregation stages remains challenging . |
| Approach: | They propose a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning. |
| Outcome: | The proposed framework improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets. |