China

Current benchmarks for evaluating Large Language Models (LLMs) often do not exhibit enough writing style diversity, with many adhering primarily to standardized conventions. Such benchmarks do not fully capture the rich variety of communication patterns exhibited by humans. Thus, it is possible that LLMs, which are optimized on these benchmarks, may demonstrate brittle performance when faced with &quot;non-standard&quot; input. In this work, we test this hypothesis by rewriting evaluation prompts using persona-based LLM prompting, a low-cost method to emulate diverse writing styles. Our results show that, even with identical semantic content, variations in writing style and prompt formatting significantly impact the estimated performance of the LLM under evaluation. Notably, we identify distinct writing styles that consistently trigger either low or high performance across a range of models and tasks, irrespective of model family, size, or recency. Our work offers a scalable approach to augment existing benchmarks, improving the external validity of the assessments they provide for LLM performance across linguistic variations.

EMNLP 2025

Persona-Augmented Benchmarking: Evaluating LLMs Across Diverse Writing Styles

evaluation methodologies

linguistic variation

prompting

benchmarking

data augmentation

robustness

## 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.

Current benchmarks for evaluating Large Language Models (LLMs) often do not exhibit enough writing style diversity, with many adhering primarily to standardized conventions. Such benchmarks do not fully capture the rich variety of communication patterns exhibited by humans. Thus, it is possible that LLMs, which are optimized on these benchmarks, may demonstrate brittle performance when faced with "non-standard" input. In this work, we test this hypothesis by rewriting evaluation prompts using persona-based LLM prompting, a low-cost method to emulate diverse writing styles. Our results show that, even with identical semantic content, variations in writing style and prompt formatting significantly impact the estimated performance of the LLM under evaluation. Notably, we identify distinct writing styles that consistently trigger either low or high performance across a range of models and tasks, irrespective of model family, size, or recency. Our work offers a scalable approach to augment existing benchmarks, improving the external validity of the assessments they provide for LLM performance across linguistic variations.

poster

The rapid advancements in large language models (LLMs) have significantly improved their ability to generate natural language, making texts generated by LLMs increasingly indistinguishable from human-written texts. While recent research has primarily focused on using LLMs to classify text as either human-written and machine-generated texts, our study focus on characterizing these texts using a set of linguistic features across different linguistic levels such as morphology, syntax, and semantics. We select a dataset of human-written and machine-generated texts spanning 8 domains and produced by 11 different LLMs. We calculate different linguistic features such as dependency length and emotionality and we use them for characterizing human-written and machine-generated texts along with different sampling strategies, repetition controls and model release date. Our statistical analysis reveals that human-written texts tend to exhibit simpler syntactic structures and more diverse semantic content. Furthermore, we calculate the variability of our set of features across models and domains. Both human and machine texts show stylistic diversity across domains, with humans displaying greater variation in our features. Finally, we apply style embeddings to further test variability among human-written and machine-generated texts. Notably, newer models output text that is similarly variable, pointing to an homogenization of machine-generated texts.

Linguistic and Embedding-Based Profiling of Texts Generated by Humans and Large Language Models

The surge of user-generated online content presents a wealth of insights into customer preferences and market trends. However, the highly diverse, complex, and context-rich nature of such contents poses significant challenges to traditional opinion mining approaches. To address this, we introduce Online Opinion Mining Benchmark (OOMB), a novel dataset and evaluation protocol designed to assess the ability of large language models (LLMs) to mine opinions effectively from diverse and intricate online environments. OOMB provides extensive (entity, feature, opinion) tuple annotations and a comprehensive opinion-centric summary that highlights key opinion topics within each content, thereby enabling the evaluation of both the extractive and abstractive capabilities of models. Through our proposed benchmark, we conduct a comprehensive analysis of which aspects remain challenging and where LLMs exhibit adaptability, to explore whether they can effectively serve as opinion miners in realistic online scenarios. This study lays the foundation for LLM-based opinion mining and discusses directions for future research in this field.

Can Large Language Models be Effective Online Opinion Miners?

Large language models (LLMs) have demonstrated impressive performance in machine translation, but still struggle with unseen low-resource languages, especially those written in underrepresented scripts. To investigate whether LLMs can translate such languages with the help of linguistic resources, we introduce Lotus, a benchmark designed to evaluate translation for Mongolian (in traditional script) and Yi. Our study shows that while linguistic resources can improve translation quality as measured by automatic metrics, LLMs remain limited in their ability to handle these languages effectively. We hope our work provides insights for the low-resource NLP community and fosters further progress in machine translation for underrepresented languages. Our code and data are available.

Can Large Language Models Translate Unseen Languages in Underrepresented Scripts?

Adversarial text defense is a significant strategy to protect modern NLP models from being attacked. Typical text defense methods usually enhance the model’s robustness by model retraining or equipping it with a data preprocessing step, aiming to eliminate the non-robust features and preserve the robust ones. Although some efforts have been made to recognize the robust features, e.g., by the information bottleneck (IB) technique, how to fully disentangle the robust and non-robust representation remains a big challenge. To alleviate this problem, we propose a novel text defense method, named Disentangled Information Bottleneck (DisIB), with two major merits. Firstly, we separate the robust features and non-robust features with a disentangled two-line framework rather than the one-line compression network in IB. This prevents the loss of robust features caused by information compression and produces complete robust features. Secondly, we design a discriminator network to approximate the minimum mutual information of the two lines, which sufficiently disentangles robust and non-robust features. To validate the effectiveness of our DisIB, we conduct a total of 96 defense experiments on four datasets by defending four popular attack methods. Experimental results elaborate that our method significantly outperforms six baselines, with accuracy improvements ranging from 3.8% to 20.7%.

Disentangled Information Bottleneck for Adversarial Text Defense

GUI agents powered by LLMs show promise in interacting with diverse digital environments. Among these, video games offer a valuable testbed due to their varied interfaces, with adventure games posing additional challenges through complex, narrative-driven interactions. Existing game benchmarks, however, lack diversity and rarely evaluate agents on completing entire storylines. To address this, we introduce FlashAdventure, a benchmark of 34 Flash-based adventure games designed to test full story arc completion and tackle the observation-behavior gap—the challenge of remembering and acting on earlier gameplay information. We also propose CUA-as-a-judge, an automated gameplay evaluator, and COAST, an agentic framework leveraging long-term clue memory to better plan and solve sequential tasks. Experiments show current GUI agents struggle with full story arcs, while COAST improves milestone completion by bridging the observation-behavior gap. Nonetheless, a marked discrepancy between humans and best-performing agents warrants continued research efforts to narrow this divide.

FlashAdventure: A Benchmark for GUI Agents Solving Full Story Arcs in Diverse Adventure Games

Scientific fact-checking has largely focused on textual and tabular sources, neglecting scientific charts—a primary medium for conveying quantitative evidence and supporting statistical reasoning in research communication. We introduce ClimateViz, the first large-scale benchmark for scientific fact-checking grounded in real-world, expert-curated scientific charts. ClimateViz comprises 49,862 claims paired with 2,896 visualizations, each labeled as support, refute, or not enough information. To enable interpretable verification, each instance includes structured knowledge graph explanations that capture statistical patterns, temporal trends, spatial comparisons, and causal relations. We conduct a comprehensive evaluation of state-of-the-art multimodal large language models, including proprietary and open-source ones, under zero-shot and few-shot settings. Our results show that current models struggle to perform fact-checking when statistical reasoning over charts is required: even the best-performing systems, such as Gemini 2.5 and InternVL 2.5, achieve only 76.2–77.8% accuracy in label-only output settings, which is far below human performance (89.3% and 92.7%). While few-shot prompting yields limited improvements, explanation-augmented outputs significantly enhance performance in some closed-source models, notably o3 and Gemini 2.5.

ClimateViz: A Benchmark for Statistical Reasoning and Fact Verification on Scientific Charts

Audio descriptions (ADs) are indispensable for blind or visually-impaired individuals (BVIs), enabling them to understand the narrative and appreciate the visual diversity of movies. There is an explosion of interest in automatic AD generation for trimmed clips and many new metrics have also been proposed. However, they typically compare single ground-truth ADs against their prediction. We posit that ADs should not be treated as independent captions and pivot to a video-level evaluation. We propose ADQA, a question-answering benchmark to evaluate whether the generated ADs would help BVIs appreciate and understand the story. We motivate the QA framework by quantifying the subjective nature of ADs through an alignment between two AD sources of the same movie. ADQA features visual appreciation (VA) questions about specific visual facts and narrative understanding (NU) questions created using plot sentences associated with videos. Evaluation of current AD generation methods show a large gap to human performance, estimated by using the second AD source. Based on our findings, we provide several recommendations for future work on AD generation.

What You See is What You Ask: Evaluating Audio Descriptions

Large Language Models (LLMs) are trained on massive web-crawled corpora, often containing personal information, copyrighted text, and benchmark datasets. This inadvertent inclusion in the training dataset, known as data leakage, poses significant risks and could compromise the safety of LLM outputs. Despite its criticality, existing studies do not examine how leaked instances in the pre-training data influence LLMs' output and detection capabilities. In this paper, we conduct an experimental survey to elucidate the relationship between data leakage in training datasets and its effects on the generation and detection by LLMs. Our experiments reveal that LLMs often generate outputs containing leaked information, even when there is little such data in the training dataset. Moreover, the fewer the leaked instances, the more difficult it becomes to detect such leakage. Finally, we demonstrate that enhancing leakage detection through few-shot learning can help mitigate the impact of the leakage rate in the training data on detection performance.

Investigating How Pre-training Data Leakage Affects Models' Reproduction and Detection Capabilities

Standardized benchmarks are central to evaluating and comparing model performance in Natural Language Processing (NLP). However, Large Language Models (LLMs) have exposed shortcomings in existing benchmarks, and so far there is no clear solution. In this paper, we survey a wide scope of benchmarking issues, and provide an overview of solutions as they are suggested in the literature. We observe that these solutions often tackle a limited number of issues, neglecting other facets. Therefore, we propose concrete checklists to cover all aspects of benchmarking issues, both for benchmark creation and usage. Additionally, we discuss the advantages of adding minimal-sized test-suites to benchmarking, ensuring downstream applicability on real-world use cases.

In Benchmarks We Trust ... Or Not?

We investigate the identification of idiomatic expressions—a semantically non-compositional subclass of multiword expressions (MWEs)—in running text using large language models (LLMs) without any fine-tuning. Instead, we adopt a prompt-based approach and evaluate a range of prompting strategies, including zero-shot, few-shot, and chain-of-thought variants, across multiple languages, datasets, and model types. Our experiments show that, with well-crafted prompts, LLMs can perform competitively with supervised models trained on annotated data. These findings highlight the potential of prompt-based LLMs as a flexible and effective alternative for idiomatic expression identification.

Downloads

Next from EMNLP 2025

Linguistic and Embedding-Based Profiling of Texts Generated by Humans and Large Language Models

Similar lecture

Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors