China

Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce USimBench, a benchmark of 909 annotated human–LLM conversations on two interactive tasks---math tutoring and document creation. USimBench evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman&#39;s ρ of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation.

EMNLP 2025

SimulatorArena: Are User Simulators Reliable Proxies for Multi-Turn Evaluation of AI Assistants?

document creation

math tutoring

multi-turn conversation evaluation

user simulator evaluation

dialogue

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

NAACL 2025

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce USimBench, a benchmark of 909 annotated human–LLM conversations on two interactive tasks---math tutoring and document creation. USimBench evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman's ρ of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation.

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.

Methods for lexical semantic change detection quantify changes in the meaning of words over time. Prior research produced many methods that excel when evaluated on established benchmarks that consist of annotated sets of pre-selected \emph{target} words; the targets are chosen ahead of time due to the prohibitively large cost of manually annotating all words. However, performance measured on small curated sets of words cannot reveal how well these methods would perform when applied to \emph{discovering} semantic changes among the \emph{full} corpus vocabulary, which is the actual end goal for many applications. In this paper, we implement a top-k set-up to evaluate semantic change discovery (SCDisc) despite lacking complete annotations (although a contribution of our efforts is a 50% extension of the annotations in the commonly used LiverpoolFC benchmark). We use our evaluation setup to investigate a battery of semantic change detection methods that utilize three different kinds of contextualized representations, as well as recent techniques specifically designed to reduce false-discovery rates, such as statistical significance tests and metric scaling. We find that despite great performance on pre-existing benchmarks, \emph{all} methods of semantic change detection struggle at ranking known changes higher compared to other words in the vocabulary when presented with a \emph{natural} distribution of instances. Finally, we verify that the majority of words with high detected change scores do not in fact experience meaning changes. In fact, for most of the methods less than a third of highest ranked changes were annotated to have changed in meaning. Given the large performance discrepancies between semantic-change scoring and discovery ``in the wild'', we recommend that researchers direct more attention to semantic change discovery and include it in their suite of evaluations when proposing new methods.

Current Semantic-change Quantification Methods Struggle with Semantic Change Discovery in the Wild

In this paper, we introduce Spotlight, a novel paradigm for information extraction that produces concise, engaging narratives by highlighting the most compelling aspects of a document. Unlike traditional summaries, which prioritize comprehensive coverage, spotlights selectively emphasize intriguing content to foster deeper reader engagement with the source material. We formally differentiate spotlights from related constructs and support our analysis with a detailed benchmarking study using new datasets curated for this work. To generate high-quality spotlights, we propose a two-stage approach: fine-tuning a large language model on our benchmark data, followed by alignment via Direct Preference Optimization (DPO). Our comprehensive evaluation demonstrates that the resulting model not only identifies key elements with precision but also enhances readability and boosts the engagement value of the original document.

Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents

Large Language Models (LLMs) have revolutionized various Natural Language Generation (NLG) tasks, including Argument Summarization (ArgSum), a key subfield of Argument Mining (AM). This paper investigates the integration of state-of-the-art LLMs into ArgSum, including for its evaluation. In particular, we propose a novel prompt-based evaluation scheme, and validate it through a novel human benchmark dataset. Our work makes three main contributions: (i) the integration of LLMs into existing ArgSum frameworks, (ii) the development of a new LLM-based ArgSum system, benchmarked against prior methods, and (iii) the introduction of an advanced LLM-based evaluation scheme. We demonstrate that the use of LLMs substantially improves both the generation and evaluation of argument summaries, achieving state-of-the-art results and advancing the field of ArgSum. We also show that among the four LLMs integrated in (i) and (ii), Qwen-3-32B, despite having the fewest parameters, performs best, even surpassing GPT-4o, while LLaMA-3.3-70B consistently underperforms.

Argument Summarization and its Evaluation in the Era of Large Language Models

This work investigates capturing and modeling disagreement in Semantic Textual Similarity (STS), where sentence pairs are assigned ordinal similarity labels (0–5). Conventional STS systems average multiple annotator scores and focus on a single numeric estimate, overlooking label dispersion. By leveraging the disaggregated SemEval-2015 dataset (Soft-STS-15), the authors propose a disagreement-aware calibration strategy that treats STS as an ordinal distribution prediction problem. Using a cross-encoder trained with a distance-aware objective, they produce softer, probabilistic output and apply post-hoc temperature scaling to refine calibration. Results show improved performance in distance-based metrics and robust handling of ambiguous pairs, demonstrating that modeling disagreement benefits both calibration and ranking accuracy. The paper thus highlights the value of retaining and modeling full annotation distributions, rather than collapsing them to a single mean label.

Beyond Averages: Learning with Annotator Disagreement in STS

Topic models often fail to capture low-prevalence, domain-critical themes—so-called minority topics—such as mental health themes in online comments. While some existing methods can incorporate domain knowledge such as expected topical content, methods allowing guidance may require overly detailed expected topics, hindering the discovery of topic divisions and variation. We propose a topic modeling solution via a specially constrained NMF. We incorporate a seed word list characterizing minority content of interest, but we do not require experts to pre-specify their division across minority topics. Through prevalence constraints on minority topics and seed word content across topics, we learn distinct data-driven minority topics as well as majority topics. The constrained NMF is fitted via Karush-Kuhn-Tucker (KKT) conditions with multiplicative updates. We outperform several baselines on synthetic data in terms of topic purity, normalized mutual information, and also evaluate topic quality using Jensen-Shannon divergence (JSD). We conduct a case study on YouTube vlog comments, analyzing viewer discussion of mental health content; our model successfully identifies and reveals this domain relevant minority content.

Constrained Non-negative Matrix Factorization for Guided Topic Modeling of Minority Topics

Whether language models (LMs) have inductive biases that favor typologically frequent grammatical properties over rare, implausible ones has been investigated, typically using artificial languages (ALs) (White and Cotterell, 2021; Kuribayashi et al., 2024). In this paper, we extend these works from two perspectives. First, we extend their context-free AL formalization by adopting Generalized Categorial Grammar (GCG) (Wood, 2014), which allows ALs to cover attested but previously overlooked constructions, such as unbounded dependency and mildly context-sensitive structures. Second, our evaluation focuses more on the generalization ability of LMs to process longer sentences, while previous studies did not. These lead to better extensibility of ALs and clearer conclusions than in the in-domain evaluation --- typologically plausible word orders tend to be easier for LMs to productively generalize.

Which Word Orders Facilitate Length Generalization in LMs? An Investigation with GCG-Based Artificial Languages

Interpretability is a key challenge in fostering trust for Large Language Models (LLMs), which stems from the complexity of extracting reasoning from model's parameters. We present the Frame Representation Hypothesis, a theoretically robust framework grounded in the Linear Representation Hypothesis (LRH) to interpret and control LLMs by modeling multi-token words. Prior research explored LRH to connect LLM representations with linguistic concepts, but was limited to single token analysis. As most words are composed of several tokens, we extend LRH to multi-token words, thereby enabling usage on any textual data with thousands of concepts. To this end, we propose words can be interpreted as frames, ordered sequences of vectors that better capture token-word relationships. Then, concepts can be represented as the average of word frames sharing a common concept. We showcase these tools through Top-k Concept-Guided Decoding, which can intuitively steer text generation using concepts of choice. We verify said ideas on Llama 3, Gemma 2, Phi 3 and Qwen-2-VL families, demonstrating gender and language biases, exposing harmful content, but also potential to remediate them, leading to safer and more transparent LLMs. Code is available at https://github.com/phvv-me/frame-representation-hypothesis.git

The Frame Representation Hypothesis: Multi-Token LLM Interpretability and Concept-Guided Text Generation

Detecting content generated by large language models (LLMs) is crucial for preventing misuse and building trustworthy AI systems. Although existing detection methods perform well, their robustness in out-of-distribution (OOD) scenarios is still lacking. In this paper, we hypothesize that, compared to features used by existing detection methods, the internal representations of LLMs contain more comprehensive and raw features that can more effectively capture and distinguish the statistical pattern differences between LLM-generated texts (LGT) and human-written texts (HWT). We validated this hypothesis across different LLMs and observed significant differences in neural activation patterns when processing these two types of texts. Based on this, we propose RepreGuard, an efficient statistics-based detection method. Specifically, we first employ a surrogate model to collect representation of LGT and HWT, and extract the distinct activation feature that can better identify LGT. We can classify the text by calculating the projection score of the text representations along this feature direction and comparing with a precomputed threshold. Experimental results show that RepreGuard outperforms all baselines with average 94.92% AUROC on both in-distribution (ID) and OOD scenarios, while also demonstrating robust resilience to various text sizes and mainstream attacks.

RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns

A broad range of NLP tasks involve selecting relevant text spans from given source texts. Despite this shared objective, such content selection tasks have traditionally been studied in isolation, each with its own modeling approaches, datasets, and evaluation metrics. In this work, we propose instruction-guided content selection (IGCS) as a beneficial unified framework for such settings, where the task definition and any instance-specific request are encapsulated as instructions to a language model. To promote this framework, we introduce IGCSBench, the first unified benchmark covering diverse content selection tasks. Further, we create a large generic synthetic dataset that can be leveraged for diverse content selection tasks, and show that transfer learning with these datasets often boosts performance, whether dedicated training for the targeted task is available or not. Finally, we address generic inference time issues that arise in LLM-based modeling of content selection, assess a generic evaluation metric, and overall propose the utility of our resources and methods for future content selection models.

A Unifying Scheme for Extractive Content Selection Tasks

Religion and spirituality (R/S) are complex and highly domain-dependent concepts which have long confounded researchers and policymakers. Due to their context-specificity, R/S are difficult to operationalize in conventional archival search strategies, particularly when datasets are very large, poorly accessible, and marked by information noise. As a result, considerable time investments and specialist knowledge is often needed to extract actionable insights related to R/S from general archival sources, increasing reliance on published literature and manual desk reviews. To address this challenge, we present SpiritRAG, an interactive Question Answering (Q\&A) system based on Retrieval-Augmented Generation (RAG). Built using 7,500 United Nations (UN) resolution documents related to R/S in the domains of health and education, SpiritRAG allows researchers and policymakers to conduct complex, context-sensitive database searches of very large datasets using an easily accessible, chat-based web interface. SpiritRAG is lightweight to deploy and leverages both UN documents and user provided documents as source material. A pilot test and evaluation with domain experts on 100 manually composed questions demonstrates the practical value and usefulness of SpiritRAG.

Downloads

Next from EMNLP 2025

Current Semantic-change Quantification Methods Struggle with Semantic Change Discovery in the Wild

Similar lecture

Aligning Black-box Language Models with Human Judgments