Papers by Yifan Zhu
TreeBoN: Enhancing Inference-Time Alignment with Speculative Tree-Search and Best-of-N Sampling (2025.findings-emnlp)
Copied to clipboard
Jiahao Qiu, Yifu Lu, Yifan Zeng, Jiacheng Guo, Jiayi Geng, Chenhao Zhu, Xinzhe Juan, Ling Yang, Huazheng Wang, Kaixuan Huang, Yue Wu, Mengdi Wang
| Challenge: | Best-of-N (BoN) sampling generates multiple responses and selects the best one, achieving improved performance but with a high computational cost. |
| Approach: | They propose a framework that integrates a speculative tree-search strategy into Best-of-N (BoN) Sampling. |
| Outcome: | The proposed framework outperforms Best-of-N (BoN) sampling but has high computational cost . tree-search strategy reduces computational overhead while maintaining high output quality . |
HiCoLoRA: Addressing Context-Prompt Misalignment via Hierarchical Collaborative LoRA for Zero-Shot DST (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to zero-shot Dialog State Tracking (zs-DST) are inadequate to generalize to new domains without extensive training. |
| Approach: | They propose a framework that enhances zero-shot slot inference through robust prompt alignment. |
| Outcome: | Experiments on multi-domain datasets show that HiCoLoRA outperforms baselines, achieving SOTA in zs-DST. |
Hierarchical Memory Organization for Wikipedia Generation (2025.acl-long)
Copied to clipboard
Eugene J. Yu, Dawei Zhu, Yifan Song, Xiangyu Wong, Jiebin Zhang, Wenxuan Shi, Xiaoguang Li, Qun Liu, Sujian Li
| Challenge: | Existing methods for generating Wikipedia articles do not utilize memory directly for outline generation. |
| Approach: | They propose a method to generate Wikipedia articles autonomously by leveraging a hierarchical memory architecture. |
| Outcome: | The proposed framework outperforms baseline methods in producing informative and reliable articles. |
WhitenedCSE: Whitening-based Contrastive Learning of Sentence Embeddings (2023.acl-long)
Copied to clipboard
| Challenge: | Extensive experiments on seven semantic textual similarity tasks show our method achieves consistent improvement over the contrastive learning baseline and sets new states of the art. |
| Approach: | They propose a whitening-based contrastive learning method for sentence embedding learning which combines contrastive and shuffled group whitening. |
| Outcome: | The proposed method achieves better alignment and uniformity on seven semantic textual similarity tasks. |
CoCA: Fusing Position Embedding with Collinear Constrained Attention in Transformers for Long Context Window Extending (2024.acl-long)
Copied to clipboard
| Challenge: | Existing models that use self-attention and position embedding have anomalous behavior that hinder long context window extrapolation. |
| Approach: | They propose a collinear constraint between Q and K to integrate RoPE and self-attention. |
| Outcome: | The proposed model integrates self-attention and position embedding into LLMs without fine-tuning. |
RAGEval: Scenario Specific RAG Evaluation Dataset Generation Framework (2025.acl-long)
Copied to clipboard
Kunlun Zhu, Yifan Luo, Dingling Xu, Yukun Yan, Zhenghao Liu, Shi Yu, Ruobing Wang, Shuo Wang, Yishan Li, Nan Zhang, Xu Han, Zhiyuan Liu, Maosong Sun
| Challenge: | Existing evaluation metrics for RAG systems are lacking due to high costs of data construction and lack of factual accuracy. |
| Approach: | They propose a framework to evaluate RAG systems in specialized scenarios . they propose three new metrics to evaluate LLM-generated responses . |
| Outcome: | The proposed framework outperforms zero-shot and one-shot methods in terms of clarity, safety, conformity, and richness of generated samples. |
Complex Numerical Reasoning with Numerical Semantic Pre-training Framework (2025.emnlp-main)
Copied to clipboard
| Challenge: | Numerical knowledge graphs (NKGs) are not limited to discrete entity-relation knowledge. |
| Approach: | They propose to combine numerical values and entities to solve multi-hop complex reasoning over incomplete knowledge graphs. |
| Outcome: | The proposed approach handles up to 102 types of complex numerical reasoning queries on three public datasets. |
AgentBank: Towards Generalized LLM Agents via Fine-Tuning on 50000+ Interaction Trajectories (2024.findings-emnlp)
Copied to clipboard
Yifan Song, Weimin Xiong, Xiutian Zhao, Dawei Zhu, Wenhao Wu, Ke Wang, Cheng Li, Wei Peng, Sujian Li
| Challenge: | Existing studies focus on specialized agents designed for particular tasks. |
| Approach: | They propose to scale annotated interaction trajectories and fine-tune LLMs on AgentBank to get a series of agent models, Samoyed. |
| Outcome: | The proposed model can scale to get generalized agent capabilities. |
Knowledge-Selective Pretraining for Attribute Value Extraction (2023.findings-emnlp)
Copied to clipboard
Hui Liu, Qingyu Yin, Zhengyang Wang, Chenwei Zhang, Haoming Jiang, Yifan Gao, Zheng Li, Xian Li, Chao Zhang, Bing Yin, William Wang, Xiaodan Zhu
| Challenge: | Existing methods for AVE are limited on rare attributes due to poor generalization ability. |
| Approach: | They propose to leverage pretraining and transfer learning to address weaknesses in existing methods. |
| Outcome: | The proposed method achieves new state-of-the-art performance without pretraining on rare attributes with limited training resources. |
Answering Ambiguous Questions through Generative Evidence Fusion and Round-Trip Prediction (2021.acl-long)
Copied to clipboard
Yifan Gao, Henghui Zhu, Patrick Ng, Cicero Nogueira dos Santos, Zhiguo Wang, Feng Nan, Dejiao Zhang, Ramesh Nallapati, Andrew O. Arnold, Bing Xiang
| Challenge: | Open-domain question answering is a task to answer questions using passages with diverse topics. |
| Approach: | They propose a model that aggregates evidence from multiple passages to adaptively predict a single answer or a set of question-answer pairs for ambiguous questions. |
| Outcome: | The proposed model achieves state-of-the-art performance on AmbigQA dataset and shows competitive performance on NQ-Open and TriviaQA. |
More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression (2025.findings-emnlp)
Copied to clipboard
Jiebin Zhang, Dawei Zhu, Yifan Song, Wenhao Wu, Chuqiao Kuang, Xiaoguang Li, Lifeng Shang, Qun Liu, Sujian Li
| Challenge: | storing more tokens in the KV cache at lower precision can enhance the long-context performance of large language models. |
| Approach: | They propose a token-precision trade-off strategy to optimize KV cache compression . they also propose storing more tokens in the KV at lower precision . |
| Outcome: | The proposed method achieves an optimal point within the Information Bottleneck compared to standalone KV pruning or KV quantization. |
TCPO: Thought-Centric Preference Optimization for Effective Embodied Decision-making (2025.emnlp-main)
Copied to clipboard
Kechen Jiao, Zhirui Fang, Jiahao Liu, Bei Li, Qifan Wang, Xinyu Liu, Junhao Ruan, Zhongjian Qiao, Yifan Zhu, Yaxin Xu, Jingang Wang, Xiu Li
| Challenge: | Existing post-SFT methods for embodied AI are constrained by sparse rewards and action-only optimization, resulting in low sample efficiency, poor consistency, and model degradation. |
| Approach: | They propose to integrate Thought-Centric Preference Optimization (TCPO) into embodied decision-making by transforming sparse reward signals into richer step sample pairs. |
| Outcome: | The proposed approach achieves an average success rate of 26.67% in the ALFWorld environment, and a 6% improvement over RL4VLM. |
SLAM-Omni: Timbre-Controllable Voice Interaction System with Single-Stage Training (2025.findings-acl)
Copied to clipboard
Wenxi Chen, Ziyang Ma, Ruiqi Yan, Yuzhe Liang, Xiquan Li, Ruiyang Xu, Zhikang Niu, Yanqiao Zhu, Yifan Yang, Zhanxun Liu, Kai Yu, Yuxuan Hu, Jinyu Li, Yan Lu, Shujie Liu, Xie Chen
| Challenge: | a new spoken dialogue system with single-stage training is demonstrating its low latency and high quality . SLAM-Omni achieves zero-shot timbre control by modeling spoken language with semantic tokens . |
| Approach: | They propose a timbre-controllable, end-to-end voice interaction system with single-stage training. |
| Outcome: | The proposed system outperforms previous models on 4 GPUs with limited data. |
InfoCL: Alleviating Catastrophic Forgetting in Continual Text Classification from An Information Theoretic Perspective (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have identified the severe performance decrease on analogous classes as a key factor for catastrophic forgetting. |
| Approach: | They propose a replay-based continual text classification method that uses fast-slow and current-past contrastive learning to perform mutual information maximization and better recover previously learned representations. |
| Outcome: | The proposed method achieves state-of-the-art on three text classification tasks. |
Video2Roleplay: A Multimodal Dataset and Framework for Video-Guided Role-playing Agents (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to RPAs focus on static role profiles, overlooking dynamic perceptual abilities inherent to humans. |
| Approach: | They propose a framework that combines adaptive temporal sampling with dynamic and static role profiles. |
| Outcome: | The proposed framework combines adaptive temporal sampling with dynamic and static role profiles. |
MASTER: Multi-Agent Security Through Exploration of Roles and Topological Structures - A Comprehensive Framework (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs)-based Multi-Agent Systems (MAS) exhibit remarkable problem-solving and task planning capabilities across diverse domains . |
| Approach: | They propose a security research framework for LLM-based multi-agent systems . they propose corresponding defense strategies to address MAS security risks . |
| Outcome: | The proposed framework amplifies the severity of security risks under MAS attacks . it offers an automated construction process for different MAS setups and an interaction paradigm . |
PersLEARN: Research Training through the Lens of Perspective Cultivation (2023.acl-demo)
Copied to clipboard
Yu-Zhe Shi, Shiqian Li, Xinyi Niu, Qiao Xu, Jiawen Liu, Yifan Xu, Shiyu Gu, Bingru He, Xinyang Li, Xinyu Zhao, Zijian Zhao, Yidong Lyu, Zhen Li, Sijia Liu, Lin Qiu, Jinhao Ji, Lecheng Ruan, Yuxi Ma, Wenjuan Han, Yixin Zhu
| Challenge: | PersLEARN is a tool designed to facilitate the cultivation of scientific perspectives . junior researchers struggle to identify the perspectives reflected in the literature and struggle to develop their own viewpoints. |
| Approach: | They propose a tool to facilitate the cultivation of scientific perspectives by interacting with a prompt-based model and allowing students to develop their own perspectives explicitly. |
| Outcome: | The proposed tool outperforms baseline approaches across multiple domains of literature from different perspectives. |
Evaluating Large Language Models with Enterprise Benchmarks (2025.naacl-industry)
Copied to clipboard
Bing Zhang, Mikio Takeuchi, Ryo Kawahara, Shubhi Asthana, Maruf Hossain, Guang-Jie Ren, Kate Soule, Yifan Mai, Yada Zhu
| Challenge: | Existing benchmarks lack domain-specific datasets for evaluating large language models . existing benchmarks often lack domain specific datasets, which can be difficult to convert to standardized metrics or regulatory issues. |
| Approach: | They propose to use 25 publicly available domain-specific English benchmarks from diverse domains . they propose to combine a wide range of natural language processing tasks for holistic evaluation . |
| Outcome: | The proposed framework includes 25 publicly available domain-specific English benchmarks from diverse enterprise domains like financial services, legal, climate, cyber security, and 2 public Japanese finance benchmarks. |
Parallel sentences mining with transfer learning in an unsupervised setting (2021.naacl-srw)
Copied to clipboard
| Challenge: | Existing methods to mine parallel sentences in low-resource environments are not suitable for many low-level language pairs. |
| Approach: | They propose an approach based on transfer learning to mine parallel sentences in an unsupervised setting using bilingual corpora of low-resource language pairs. |
| Outcome: | The proposed model improves the performance of mined parallel sentences at two real-world low-resource language pairs compared with previous methods. |
Common Ground Tracking in Multimodal Dialogue (2024.lrec-main)
Copied to clipboard
Ibrahim Khalil Khebour, Kenneth Lai, Mariah Bradford, Yifan Zhu, Richard A. Brutti, Christopher Tam, Jingxuan Tu, Benjamin A. Ibarra, Nathaniel Blanchard, Nikhil Krishnaswamy, James Pustejovsky
| Challenge: | In dialogue modeling, there is considerable attention on “dialogue state tracking” (DST) but “common ground tracking” identifies the shared belief space held by all participants in a task-oriented dialogue: the task-relevant propositions all participants accept as true. |
| Approach: | They propose a method for automatically identifying the current set of shared beliefs and ”questions under discussion” of a group with a shared goal. |
| Outcome: | The proposed method predicts moves toward building common ground relative to ground truth in a multimodal interaction with an AI. |
Beyond Pedagogical Principles: Multi-Horizon Preference Optimization for Efficient Socratic Tutoring (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for developing LLMs are constrained by static data or sparse reward signals in online settings. |
| Approach: | They propose a framework that iteratively refines tutor agents using a multi-horizon reward function within a dynamic teacher-student simulation environment. |
| Outcome: | The proposed framework improves model performance and balances principles and effectiveness compared to baselines. |
ConFiguRe: Exploring Discourse-level Chinese Figures of Speech (2022.coling-1)
Copied to clipboard
| Challenge: | Figures of speech often deviate from their literal meanings to express deeper semantic implications. |
| Approach: | They propose a concept of figurative unit, which is the carrier of a figure, and build a Chinese corpus for Contextualized Figure Recognition. |
| Outcome: | The proposed model is based on 12 types of figures commonly used in Chinese . it shows that the proposed tasks are challenging for existing models . |
ChatKBQA: A Generate-then-Retrieve Framework for Knowledge Base Question Answering with Fine-tuned Large Language Models (2024.findings-acl)
Copied to clipboard
Haoran Luo, Haihong E, Zichen Tang, Shiyao Peng, Yikai Guo, Wentai Zhang, Chenghao Ma, Guanting Dong, Meina Song, Wei Lin, Yifan Zhu, Anh Tuan Luu
| Challenge: | Existing KBQA methods address inefficient knowledge retrieval and semantic parsing errors. |
| Approach: | They propose a generatethen-retrieve KBQA framework that generates logical form and replaces entities and relations with an unsupervised retrieval method to improve both generation and retrieval more directly. |
| Outcome: | Experimental results show that ChatKBQA achieves new state-of-the-art performance on standard KBQA datasets, WebQSP, and CWQ. |
DyBBT: Dynamic Balance via Bandit-inspired Targeting for Dialog Policy with Cognitive Dual Systems (2026.acl-long)
Copied to clipboard
| Challenge: | Task oriented dialog systems often rely on static exploration strategies that do not adapt to dynamic dialog contexts. |
| Approach: | They propose a dialog policy learning framework that formalizes the exploration challenge through a structured cognitive state space C. |
| Outcome: | The proposed framework achieves SOTA performance in success rate, efficiency, and generalization. |
FANNO: Augmenting High-Quality Instruction Data with Open-Sourced LLMs Only (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent studies explore approaches to synthesize instruction data with open-sourced LLMs but require high-quality human-crafted seed data. |
| Approach: | They propose an end-to-end framework to synthesize high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data. |
| Outcome: | The proposed framework synthesizes high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data. |
DecoupledESC: Enhancing Emotional Support Generation via Strategy-Response Decoupled Preference Optimization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing ESC data entangles psychological strategies and response content, making it difficult to construct high-quality preference pairs. |
| Approach: | They propose a Decoupled ESC framework that decomposes the ESC task into two sequential subtasks: strategy planning and empathic response generation. |
| Outcome: | The proposed framework outperforms baselines, reducing preference bias and improving response quality. |
Bidirectional LMs are Better Knowledge Memorizers? A Benchmark for Real-world Knowledge Injection (2026.acl-long)
Copied to clipboard
Yuwei Zhang, Wenhao Yu, Shangbin Feng, Yifan Zhu, Letian Peng, Jayanth Srinivasa, Gaowen Liu, Jingbo Shang
| Challenge: | Existing knowledge injection benchmarks for large language models lack standardized testing grounds. |
| Approach: | They propose a knowledge injection benchmark that leverages recently-added and expert-curated facts from Wikipedia’s “Did You Know...” entries. |
| Outcome: | The proposed framework improves reliability accuracy by 29.1%. |
Can MLLMs Reason Beyond Language? VisReason: A Comprehensive Benchmark for Vision-Centric Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in multimodal large language models demonstrate strong performance on visual reasoning benchmarks. |
| Approach: | They propose a benchmark for vision-centric reasoning that integrates visual and textual information for non-trivial reasoning. |
| Outcome: | The proposed benchmark exposes gaps between humans and current MLLMs and reveals limited benefits from test-time reasoning strategies. |
LongEmbed: Extending Embedding Models for Long Context Retrieval (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing embedding models support only 512 input tokens, hindering their application in scenarios requiring long inputs. |
| Approach: | They evaluate the performance of existing embedding models by using a new benchmark and a training-free context window extension strategy. |
| Outcome: | The proposed model extends the input window of existing models by several folds. |
Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have revolutionized open-domain dialogue agents but face challenges in multi-character role-playing (MCRP) scenarios. |
| Approach: | They propose a framework for efficient multi-character role-playing that employs a dynamic low-rank adapter strategy and distinct LoRA blocks for each character. |
| Outcome: | Neeko employs a dynamic low-rank adapter (LoRA) strategy, enabling it to adapt seamlessly to diverse characters. |
CoUDA: Coherence Evaluation via Unified Data Augmentation (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing data augmentations for coherence evaluation rely on heuristic rules and lack designing criteria. |
| Approach: | They propose a data augmentation framework that breaks down coherence into global and local aspects and designs augmentation strategies for both aspects. |
| Outcome: | The proposed framework surpasses recent models in scoring and ranking tasks with 233M parameters. |
Concise Math Reasoning via Difficulty-Aware Distillation (2026.findings-acl)
Copied to clipboard
Yifan Wu, Jingze Shi, Bingheng Wu, Jiayi Zhang, Xiaotian Lin, Yizhang Zhu, Zhaoyang Yu, Bang Liu, Chenglin Wu, Nan Tang, Yuyu Luo
| Challenge: | Human experts tackle difficult math problems by identifying and executing a few pivotal steps rather than listing every intermediate thought. |
| Approach: | They propose a method for producing training data that mirrors concise human reasoning by rewriting a problem's solution to retain only the essential steps. |
| Outcome: | The proposed method outperforms models trained on 800k long CoT and cuts training and inference costs. |