China

As large language models (LLMs) continue to improve, their evaluation increasingly centers on complex, high-level tasks, often at the expense of systematically assessing fundamental capabilities. To address this gap, recent work proposed LMentry, a compact benchmark comprising tasks that are trivial for humans but remain surprisingly difficult for LLMs. However, LMentry is limited to English, leaving its insights linguistically narrow. In this paper, we present Multi-LMentry, a ground-up recreation of LMentry that enables systematic evaluation of LLMs on basic reasoning and understanding tasks across nine diverse languages. Multi-LMentry includes English and expands to Catalan, German, Spanish, Basque, Galician, Korean, Italian, and Brazilian Portuguese, emphasizing the importance of cross-lingual and low-resource settings. To validate that Multi-LMentry is still trivial for humans, we demonstrate that L2 speakers with only elementary proficiency achieve near-perfect scores in a low-resource language, namely, Basque. Through extensive experiments, we reveal that state-of-the-art open-weight multilingual LLMs still fall short of human performance on elementary tasks in many languages. Our results expose new failure modes that remain hidden in monolingual evaluation, underscoring the need for rigorous, language-diverse ``unit tests&#39;&#39; of core model abilities.

EMNLP 2025

Multi-LMentry: Can Multilingual LLMs Solve Elementary Tasks Across Languages?

elementary-level

multilinguality

evaluation

**Welcome to EMNLP 2023!**

On behalf of the EMNLP 2023 Organizing Committee, I extend a warm and heartfelt welcome to all of you to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). It is with immense pleasure and excitement that we gather here in Singapore, a vibrant hub of innovation and technological advancement.

The conference program is packed with insightful presentations, thought-provoking workshops, and engaging networking opportunities. In addition to the technical sessions, we have also planned several social events that will provide you with the opportunity to connect with your colleagues.

The conference is held in person at the Resorts World Convention Centre in Singapore, and available on-line with the help of Underline.

For those in Singapore, I hope that you will find time to explore the exciting Sentosa Island, Gardens by the Bay, Singapore Botanic Gardens and the many other attractions unique to Singapore.

Of course, we are grateful to our sponsors and partners for their generous support of EMNLP 2023, whose contributions make it possible for us to host this world-class event.

Yuji Matsumoto (RIKEN AIP) 
EMNLP 2023 General Chair

To access the **EMNLP 2023** event page on Underline, you need to register for the Conference. 
Please follow **[this link](https://2023.emnlp.org/registration/)** for more details.

EMNLP 2023

EMNLP 2023 took place in Singapore from Dec 6th to Dec 10th, 2023.

As large language models (LLMs) continue to improve, their evaluation increasingly centers on complex, high-level tasks, often at the expense of systematically assessing fundamental capabilities. To address this gap, recent work proposed LMentry, a compact benchmark comprising tasks that are trivial for humans but remain surprisingly difficult for LLMs. However, LMentry is limited to English, leaving its insights linguistically narrow. In this paper, we present Multi-LMentry, a ground-up recreation of LMentry that enables systematic evaluation of LLMs on basic reasoning and understanding tasks across nine diverse languages. Multi-LMentry includes English and expands to Catalan, German, Spanish, Basque, Galician, Korean, Italian, and Brazilian Portuguese, emphasizing the importance of cross-lingual and low-resource settings. To validate that Multi-LMentry is still trivial for humans, we demonstrate that L2 speakers with only elementary proficiency achieve near-perfect scores in a low-resource language, namely, Basque. Through extensive experiments, we reveal that state-of-the-art open-weight multilingual LLMs still fall short of human performance on elementary tasks in many languages. Our results expose new failure modes that remain hidden in monolingual evaluation, underscoring the need for rigorous, language-diverse ``unit tests'' of core model abilities.

poster

## 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." 

*Message from the General Chair, Dirk Hovy*

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

Retrieval-Augmented Generation (RAG) is an effective approach to enhance the factual accuracy of large language models (LLMs) by retrieving information from external databases, which are typically composed of diverse sources, to supplement the limited internal knowledge of LLMs. However, the standard RAG often risks retrieving incorrect information, as it relies solely on relevance between a query and a document, overlooking the heterogeneous reliability of these sources. To address this issue, we propose Reliability-Aware RAG (RA-RAG), a multi-source RAG framework that estimates the reliability of sources and leverages this information to prioritize highly reliable and relevant documents, ensuring more robust and accurate response generation. Specifically, RA-RAG first estimates source reliability by cross-checking information across multiple sources. It then retrieves documents from the top- reliable and relevant sources and aggregates their information using weighted majority voting (WMV), where the selective retrieval ensures scalability while not compromising the performance. Comprehensive experiments show that RA-RAG consistently outperforms baselines in scenarios with heterogeneous source reliability while scaling efficiently as the number of sources increases. Furthermore, we demonstrate the ability of RA-RAG to estimate real-world sources' reliability, highlighting its practical applicability.

Retrieval-Augmented Generation with Estimation of Source Reliability

Language models often fail on idiomatic, figurative, or context-sensitive inputs - not due to generation errors, but due to misinterpretation during input processing. We propose an input-only, model-agnostic method for anticipating such failures using token-level likelihood features inspired by surprisal and the Uniform Information Density hypothesis. These features capture localised uncertainty in input comprehension and outperform standard baselines across five linguistically challenging datasets. We show that span-localised features improve error detection for larger models, while smaller models benefit from global patterns. Our method requires no access to outputs or activations, offering a lightweight and generalisable approach to pre-generation error prediction.

From Input Perception to Predictive Insight: Modeling Model Blind Spots Before They Become Errors

Although speech models are expected to align well with brain language processing when subjects are involved in listening, recent studies find that they cannot capture brain-relevant semantics beyond low-level features. Surprisingly, text models are better aligned since they can capture brain-relevant semantics. No previous study has investigated the alignment effectiveness of text/speech representations from multimodal models. Can speech embeddings from such multimodal models capture brain-relevant semantics through cross-modal interactions? Which modality can leverage this synergistic multimodal understanding for improved alignment with brain language processing? Can text/speech representations from such multimodal models outperform unimodal models? To address these questions, we systematically analyze multiple multimodal models, extracting both text- and speech-based representations to assess their alignment with MEG brain recordings during naturalistic story listening. We find that text embeddings from both multimodal and unimodal models significantly outperform speech embeddings from these models. Specifically, multimodal text embeddings exhibit a peak around 200 ms, suggesting that they benefit from speech embeddings, with heightened activity during this time period. However, speech embeddings from these multimodal models still show similar alignment compared to their unimodal counterparts, suggesting that they do not gain meaningful semantic benefits over text-based representations. These results highlight an asymmetry in cross-modal knowledge transfer where text modality benefits more from speech information but not vice versa.

Aligning Text/Speech Representations from Multimodal Models with MEG Brain Activity During Listening

Large language models, despite extensive alignment with human values and ethical principles, remain vulnerable to sophisticated jailbreak attacks that exploit their reasoning abilities. Existing safety measures often detect overt malicious intent but fail to address subtle, reasoning-driven vulnerabilities. In this work, we introduce POATE (Polar Opposite query generation, Adversarial Template construction, and Elaboration), a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses. POATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety. We conduct extensive evaluation across six diverse language model families of varying parameter sizes to demonstrate the robustness of the attack, achieving significantly higher attack success rates (44%) compared to existing methods. To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses. These methods enhance reasoning robustness and strengthen the model’s defense against adversarial exploits.

Turning Logic Against Itself: Probing Model Defenses Through Contrastive Questions

Authorship representation (AR) learning, which models an author's unique writing style, has demonstrated strong performance in authorship attribution tasks. However, prior research has primarily focused on monolingual settings—mostly in English—leaving the potential benefits of multilingual AR models underexplored. We introduce a novel method for multilingual AR learning that incorporates two key innovations: probabilistic content masking, which encourages the model to focus on stylistically indicative words rather than content-specific words, and language-aware batching, which improves contrastive learning by reducing cross-lingual interference. Our model is trained on over 4.5 million authors across 36 languages and 13 domains. It consistently outperforms monolingual baselines in 21 out of 22 non-English languages, achieving an average Recall@8 improvement of 4.85%, with a maximum gain of 15.91% in a single language. Furthermore, it exhibits stronger cross-lingual and cross-domain generalization compared to a monolingual model trained solely on English. Our analysis confirms the effectiveness of both proposed techniques, highlighting their critical roles in the model's improved performance.

Leveraging Multilingual Training for Authorship Representation: Enhancing Generalization across Languages and Domains

Simultaneous machine translation (SiMT) necessitates a robust read/write (R/W) policy to determine the optimal moments for translation, thereby balancing translation quality and latency. Effective timing in translation can align source and target tokens accurately. The attention mechanism within translation models inherently provides valuable alignment information. Building on this, previous research has attempted to modify the attention mechanism's structure to leverage its alignment properties during training, employing multi-task learning to derive the read/write policy. However, this multi-task learning approach may compromise the efficacy of the attention mechanism itself. This raises a natural question: why not directly learn the read/write policy from the well-trained attention mechanism? In this study, we propose DrFrattn, a method that directly learns adaptive policies from the attention mechanism. Experimental results across various benchmarks demonstrate that our approach achieves an improved balance between translation accuracy and latency.

DrFrattn: Directly Learn Adaptive Policy from Attention for Simultaneous Machine Translation

Cross-lingual transfer allows models to perform tasks in languages unseen during training and is often assumed to benefit from increased multilinguality. In this work, we challenge this assumption in the context of two underexplored, sense-aware tasks: polysemy disambiguation and lexical semantic change. Through a large-scale analysis across 28 languages, we show that multilingual training is neither necessary nor inherently beneficial for effective transfer. Instead, we find that confounding factors—such as fine-tuning data composition and evaluation artifacts—better account for the perceived advantages of multilinguality. Our findings call for more rigorous evaluations in multilingual NLP. We release fine-tuned models and benchmarks to support further research, with implications extending to low-resource and typologically diverse languages.

Multilinguality Does not Make Sense: Investigating Factors Behind Zero-Shot Cross-Lingual Transfer in Sense-Aware Tasks

The modern Hawaiian orthography contains several characters and diacritics that did not exist before the 1950s. We develop several models ranging in complexity on this task of Hawaiian orthography modernization. Our results demonstrate that simple statistical n-gram models surprisingly outperform neural seq2seq models and LLMs, highlighting the potential for traditional machine learning approaches in a low-resource setting.

Statistical and Neural Methods for Hawaiian Orthography Modernization

The exposure of large language models (LLMs) to copyrighted material during pre-training raises concerns about unintentional copyright infringement post deployment. This has driven the development of "copyright takedown" methods—post-training approaches aimed at preventing models from generating content substantially similar to copyrighted ones. While current mitigation approaches are somewhat effective for average-case risks, we demonstrate that they overlook worst-case copyright risks exhibited by the existence of long, verbatim quotes from copyrighted sources. We propose BloomScrub, a remarkably simple yet highly effective inference-time approach that provides certified copyright takedown. Our method repeatedly interleaves quote detection with rewriting techniques to transform potentially infringing segments. By leveraging efficient data sketches (Bloom filters), our approach enables scalable copyright screening—even for large-scale real-world corpora. When quotes beyond a length threshold cannot be removed, the system can abstain from responding, offering certified risk reduction. Experimental results show that BloomScrub reduces infringement risk, preserves utility, and accommodates different levels of enforcement stringency with adaptive abstention. Our results suggest that lightweight, inference-time methods can be surprisingly effective for copyright prevention.

Certified Mitigation of Worst-Case LLM Copyright Infringement

Uncertainty quantification (UQ) has emerged as a promising approach for detecting hallucinations and low-quality output of Large Language Models (LLMs). However, obtaining proper uncertainty scores is complicated by the conditional dependency between the generation steps of an autoregressive LLM, because it is hard to model it explicitly. Here, we propose to learn this dependency from attention-based features. In particular, we train a regression model that leverages LLM attention maps, probabilities on the current generation step, and recurrently computed uncertainty scores from previously generated tokens. To incorporate the recurrent features, we also suggest a two-staged training procedure. Our experimental evaluation on ten datasets and three LLMs shows that the proposed method is highly effective for selective generation, achieving substantial improvements over rivaling unsupervised and supervised approaches.

Downloads

Next from EMNLP 2025

Retrieval-Augmented Generation with Estimation of Source Reliability

Similar lecture

Don’t Trust ChatGPT when your Question is not in English: A Study of Multilingual Abilities and Types of LLMs