China

Using special tokens (e.g., gist, memory, or compressed tokens) to compress context information is a common practice for large language models (LLMs). However, existing approaches often neglect that position encodings inherently induce local inductive biases in models, causing the compression process to ignore holistic contextual dependencies. We propose **Enhanced Position Layout (EPL)**, a simple yet effective method that improves the context compression capability of LLMs by only adjusting position IDs, the numerical identifiers that specify token positions. EPL minimizes the distance between context tokens and their corresponding special tokens and at the same time maintains the sequence order in position IDs between context tokens, special tokens, and the subsequent tokens. Integrating EPL into our best performing context compression model results in 1.9 ROUGE-1 F1 improvement on out-of-domain question answering datasets in average. When extended to multimodal scenarios, EPL brings an average accuracy gain of 2.6 to vision compression LLMs.

EMNLP 2025

Position IDs Matter: An Enhanced Position Layout for Efficient Context Compression in Large Language Models

llm memory

long context llms

context compression

position encoding

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*





<br>

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

poster

Multi-hop question answering is a challenging task that requires capturing information from different positions in multiple documents. Recently, several methods propose to enhance Large Language Models (LLMs) by incorporating structured knowledge, aiming to grasp key information for solving this task. Despite certain achievements, they still face the following challenges: 1) The neglect of text-based reasoning capabilities. 2) Information redundancy between text and triples. 3) Information loss during structured knowledge extraction. To solve the above challenges, in this paper, we propose Dynamic Combination of Structured Knowledge (DCSK), a novel framework for integrating text-based and triple-based paradigms. Following Occam's Razor, DCSK dynamically determine the necessity of structured knowledge by the designed multi-faceted evaluation, which systematically assess the correctness, clarity, and informativeness of text-based prediction. For questions that require structured knowledge, we develop an iterative fact refiner that screens for question-relevant triples, verifies their factual adequacy, and thereby effectively excludes irrelevant and redundant information. Furthermore, based on the verification results, we construct an adaptive knowledge reasoner that dynamically adjusts the need for text supplementation, thus mitigating the information deficiency in selected triples. Extensive experiments on three MHQA datasets demonstrate the efficiency and effectiveness of DCSK.

Following Occam’s Razor: Dynamic Combination of Structured Knowledge for Multi-Hop Question Answering using LLMs

The combination of pre-trained speech encoders with large language models has enabled the development of speech LLMs that can handle a wide range of spoken language processing tasks. While these models are powerful and flexible, this very flexibility may make them more vulnerable to adversarial attacks. To examine the extent of this problem, in this work we investigate universal acoustic adversarial attacks on speech LLMs. Here a fixed, universal, adversarial audio segment is prepended to the original input audio. We initially investigate attacks that cause the model to either produce no output or to perform a modified task overriding the original prompt. We then extend the nature of the attack to be selective so that it activates only when specific input attributes, such as a speaker gender or spoken language, are present. Inputs without the targeted attribute should be unaffected, allowing fine-grained control over the model outputs. Our findings reveal critical vulnerabilities in Qwen2-Audio and Granite-Speech and suggest that similar speech LLMs may be susceptible to universal adversarial attacks. This highlights the need for more robust training strategies and improved resistance to adversarial attacks.

Universal Acoustic Adversarial Attacks for Flexible Control of Speech-LLMs

The increasing size and complexity of large language models (LLMs) raise concerns about their ability to “cheat” on standard Question Answering (QA) benchmarks by memorizing task-specific data. This undermines the validity of benchmark evaluations, as they no longer reflect genuine model capabilities but instead the effects of data leakage. While existing methods detect such leakage, they fail to address the long-term challenge of mitigating it. In this paper, we introduce LastingBench, a novel approach to reinforce and safeguard existing benchmarks against knowledge leakage. Our method involves identifying leakage points through perturbation-based detection, followed by counterfactual rewriting to disrupt memorization while preserving the benchmark's original evaluative intent. We demonstrate that our approach significantly reduces memorization effects in long-context QA benchmarks, providing a more accurate assessment of model reasoning and generalization abilities. Our experiments show that LastingBench not only uncovers substantial leakage in benchmarks like HotpotQA but also yields a more reliable evaluation of state-of-the-art models, ensuring that benchmarks remain effective and resilient over time.

LastingBench: Defend Benchmarks Against Knowledge Leakage

Machine translation (MT) research addressing gender inclusivity has gained attention for promoting non-exclusionary language representing all genders. However, existing resources are limited to short sources, most often single sentences, or single gender-fair formulation types, leaving questions about MT models' ability to use context and diverse inclusive forms. We introduce Glitter, a new English-German benchmark featuring extended passages with professional translations implementing three gender-fair alternatives: neutral rephrasing, typographical solutions (gender star), and neologistic forms (-ens endings). Our experiments reveal significant limitations in state-of-the-art language models, which default to masculine generics, struggle to interpret explicit gender cues in context, and rarely produce gender-fair translations. Through systematic prompting analysis designed to elicit fair language, we demonstrate that current models lack a fundamental understanding of source gender phenomena, failing to implement inclusive forms even when explicitly instructed. Glitter establishes a challenging benchmark, advancing research in gender-fair English-German MT. It highlights substantial room for improvement even among leading models and can serve to guide development of future MT models capable of accurately representing gender diversity.

Glitter: A Multi-Sentence, Multi-Reference Benchmark for Gender-Fair German Machine Translation

Temporal knowledge graph (TKG) reasoning, a central task in temporal knowledge representation, focuses on predicting future facts by leveraging historical temporal contexts. However, current approaches face two major challenges: limited generalization to unseen facts and insufficient interpretability of reasoning processes. To address these challenges, this paper proposes the **D**enoising **L**ogic-based **T**emporal **K**nowledge **G**raph (**DLTKG**) framework, which employs a denoising diffusion process to complete reasoning tasks by introducing a noise source and a historical conditionguiding mechanism. Specifically, DLTKG constructs fuzzy entity representations by treating historical facts as noise sources, thereby enhancing the semantic associations between entities and the generalization ability for unseen facts. Additionally, a condition-based guidance mechanism, rooted in the relationship evolutionary paths, is designed to improve the interpretability of the reasoning process. Furthermore, we introduce a fine-tuning strategy that optimizes the denoising process by leveraging shortest path information between head entity and candidate entities. Experimental results on three benchmark datasets demonstrate that DLTKG outperforms state-of-the-art methods across multiple evaluation metrics. Our code is available at: https://anonymous.4open.science/r/DLTKG-7CCB

DLTKG: Denoising Logic-based Temporal Knowledge Graph Reasoning

We introduce MANTA, an automated pipeline that generates high-quality large-scale instruction fine-tuning datasets from massive web corpora while preserving their diversity and scalability. By extracting structured syllabi from web documents and leveraging high-performance LLMs, our approach enables highly effective query-response generation with minimal human intervention. Extensive experiments on 8B-scale LLMs demonstrate that fine-tuning on the MANTA-1M dataset significantly outperforms other massive dataset generation methodologies, particularly in knowledge-intensive tasks such as MMLU and MMLU-Pro, while also delivering superior performance across a broad spectrum of tasks. Moreover, MANTA supports seamless scalability by allowing the continuous integration of web corpus data, enabling expansion into domains requiring intensive knowledge.

MANTA: A Scalable Pipeline for Transmuting Massive Web Corpora into Instruction Datasets

Knowledge-intensive queries require accurate answers that are explicitly grounded in retrieved evidence. However, existing retrieval-augmented generation (RAG) approaches often struggle with query complexity, suffer from propagated reasoning errors, or rely on incomplete or noisy retrieval, limiting their effectiveness. To address these limitations, we introduce UniRAG, a unified RAG framework that integrates entity-grounded query decomposition, break-down reasoning, and iterative query rewriting. Specifically, UniRAG decomposes queries into semantically coherent sub-queries, explicitly verifies retrieved sub-facts through a dedicated reasoning module, and adaptively refines queries based on identified knowledge gaps, significantly improving answer completeness and reliability. Extensive benchmark evaluations on complex question-answering datasets, including multi-hop HotPotQA and 2WikiMultihopQA, biomedical MedMCQA and MedQA, and fact-verification FEVER and SciFact, demonstrate that UniRAG consistently achieves performance improvements across various state-of-the-art LLMs, such as LLaMA-3.1-8B, GPT-3.5-Turbo, and Gemini-1.5-Flash.

UniRAG: A Unified RAG Framework for Knowledge-Intensive Queries with Decomposition, Break-Down Reasoning, and Iterative Rewriting

Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks—boosting sentiment analysis from 35.3 \% to 87.5 \% and paraphrase detection from 56.1 \% to 87.0 \%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://anonymous.4open.science/r/AraReasoner41299

AraReasoner: Evaluating Reasoning-Based LLMs for Arabic NLP

Leveraging the autonomous decision-making capabilities of large language models (LLMs) has demonstrated superior performance in reasoning tasks. However, despite the success of iterative or agentic retrieval-augmented generation (RAG) techniques, these methods are often constrained to a single solution space when confronted with complex problems. In this paper, we propose a novel thinking pattern in RAG that integrates autonomous strategic planning with efficient reasoning actions, significantly activating intrinsic reasoning capabilities and expanding the solution space of specific tasks via Monte Carlo Tree Search (MCTS), which we refer to as AirRAG. Specifically, our approach designs five fundamental reasoning actions, which are expanded to a broad tree-based reasoning space using MCTS. The approach also incorporates self-consistency verification to explore potential reasoning paths and inference scaling law. Additionally, computationally optimal strategies are employed to allocate more inference resources to key actions, thereby enhancing overall performance. Experimental results demonstrate the effectiveness of AirRAG, showing significant performance gains on complex question-answering datasets. Furthermore, AirRAG is flexible and lightweight, making it easy to integrate with other advanced technologies and models.

AirRAG: Autonomous Strategic Planning and Reasoning Steer Retrieval Augmented Generation

The prevailing "trivia-centered paradigm" for evaluating the cultural alignment of large language models (LLMs) is increasingly inadequate as these models become more advanced and widely deployed. Existing approaches typically reduce culture to static facts or values, testing models via multiple-choice or short-answer questions that treat culture as isolated trivia. Such methods neglect the pluralistic and interactive realities of culture, and overlook how cultural assumptions permeate even ostensibly "neutral" evaluation settings. In this position paper, we argue for \textbf{intentionally cultural evaluation}: an approach that systematically examines the cultural assumptions embedded in all aspects of evaluation, not just in explicitly cultural tasks. We systematically characterize the what, how, and circumstances by which culturally contingent considerations arise in evaluation, and emphasize the importance of researcher positionality for fostering inclusive, culturally aligned NLP research. Finally, we discuss implications and future directions for moving beyond current benchmarking practices, discovering important applications that we don't know exist, and involving communities in evaluation design through HCI-inspired participatory methodologies.

Downloads

Next from EMNLP 2025

Following Occam’s Razor: Dynamic Combination of Structured Knowledge for Multi-Hop Question Answering using LLMs

Similar lecture

QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory