EMNLP 2025 Papers — Page 19
Conference on Empirical Methods in Natural Language Processing · 1809 papers
XLQA: A Benchmark for Locale-Aware Multilingual Open-Domain Question Answering
Keonwoo Roh, Seong-Whan Lee (Korea University)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a multilingual open-domain QA benchmark named XLQA tailored for regional differences, containing 3,000 English seed questions and their expanded versions in eight languages (English, Korean, Arabic, Hebrew, Japanese, Russian, Vietnamese, Simplified Chinese), with human-validated annotations of answer regional sensitivity.
XQuant: Achieving Ultra-Low Bit KV Cache Quantization with Cross-Layer Compression
Haoqi Yang (Wuhan University), Hai Zhao (Xiaomi Inc.)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed XQuant, a training-agnostic and plug-and-play KV cache quantization framework, to achieve extremely low-bitwidth KV cache compression, significantly reducing memory consumption during LLM inference.
You Are What You Train: Effects of Data Composition on Training Context-aware Machine Translation Models
Paweł Mąka (Maastricht University), Gerasimos Spanakis (Maastricht University)
Data-Centric LearningTransformerText
🎯 What it does: Investigated the impact of the sparsity of context-rich examples in training data on the performance of context-aware machine translation models and systematically verified the sparse hypothesis.
Your Language Model Can Secretly Write Like Humans: Contrastive Paraphrase Attacks on LLM-Generated Text Detectors
Hao Fang (Tsinghua University), Min Zhang (Harbin Institute of Technology)
Adversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a training-free adversarial rewriting method called CoPA, which utilizes contrastive word distribution adjustment to generate text that better aligns with human writing characteristics, thereby evading LLM text detectors.
Your RAG is Unfair: Exposing Fairness Vulnerabilities in Retrieval-Augmented Generation via Backdoor Attacks
Gaurav Bagwe (Clemson University), Lan Emily Zhang
Adversarial AttackContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes BiasRAG, a fairness-targeted backdoor attack framework for retrieval-augmented generation (RAG) systems, which implants bias during the pre-training phase through a two-phase strategy and further reinforces it later via knowledge base poisoning.
ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization
Seungyoun Yi, Sungrae Park (Upstage AI Research)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the ZERA framework to achieve zero-initialization automatic prompt optimization, evaluating and improving system prompts, user prompts, and task descriptions through eight principles;
Zero-shot Multimodal Document Retrieval via Cross-modal Question Generation
Yejin Choi (Yonsei University), Youngjae Yu (Seoul National University)
RetrievalDomain AdaptationTransformerLarge Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a zero-shot multi-modal document retrieval framework PREMIR based on cross-modal question generation.
zFLoRA: Zero-Latency Fused Low-Rank Adapters
Dhananjaya Gowda (Samsung Research), Junhyun Lee (Samsung Research)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed zero-latency fused low-rank adapter zFLoRA for task-specific fine-tuning of large language models (LLMs) without introducing additional matrix multiplications.
ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration
Haozhan Shen (Zhejiang University), Jianwei Yin (Zhejiang University)
Prompt EngineeringVision Language ModelImage
🎯 What it does: Propose a tree search algorithm called Zoom Eye, enabling multi-modal large language models to simulate human zooming operations during reasoning to acquire image details;