Austria

Large Language Models (LLMs) are often used as automated judges to evaluate text, but their effectiveness can be hindered by various unintentional biases. We propose using linear classifying probes, trained by leveraging differences between contrasting pairs of prompts, to directly access LLMs&#39; latent knowledge and extract more accurate preferences. Through extensive experiments using models of varying size from four different families and six diverse datasets assessing text quality evaluation and common sense reasoning, we demonstrate that both supervised and unsupervised probing approaches consistently outperform traditional generation-based judgement while maintaining similar computational costs. These probes generalise under domain shifts and can even outperform finetuned evaluators with the same training data size. Our results suggest linear probing offers an accurate, robust and computationally efficient approach for LLM-as-judge tasks while providing interpretable insights into how models encode judgement-relevant knowledge. Our data and code will be openly released in the future.

ACL 2025

Improving Preference Extraction In LLMs By Identifying Latent Knowledge Through Classifying Probes

llm-as-a-judge

large language models

interpretability

## 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*

EMNLP 2025

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

Large Language Models (LLMs) are often used as automated judges to evaluate text, but their effectiveness can be hindered by various unintentional biases. We propose using linear classifying probes, trained by leveraging differences between contrasting pairs of prompts, to directly access LLMs' latent knowledge and extract more accurate preferences. Through extensive experiments using models of varying size from four different families and six diverse datasets assessing text quality evaluation and common sense reasoning, we demonstrate that both supervised and unsupervised probing approaches consistently outperform traditional generation-based judgement while maintaining similar computational costs. These probes generalise under domain shifts and can even outperform finetuned evaluators with the same training data size. Our results suggest linear probing offers an accurate, robust and computationally efficient approach for LLM-as-judge tasks while providing interpretable insights into how models encode judgement-relevant knowledge. Our data and code will be openly released in the future.

poster

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Pretrained language models have significantly advanced performance across various natural language processing tasks. However, adversarial attacks continue to pose a critical challenge to system built using these models, as they can be exploited with carefully crafted adversarial texts. Inspired by the ability of diffusion models to predict and reduce noise in computer vision, we propose a novel and flexible adversarial defense method for language classification tasks, DiffuseDef, which incorporates a diffusion layer as a denoiser between the encoder and the classifier. The diffusion layer is trained on top of the existing classifier, ensuring seamless integration with any model in a plug-and-play manner. During inference, the adversarial hidden state is first combined with sampled noise, then denoised iteratively and finally ensembled to produce a robust text representation. By integrating adversarial training, denoising, and ensembling techniques, we show that DiffuseDef improves over existing adversarial defense methods and achieves state-of-the-art performance against common black-box and white-box adversarial attacks.

DiffuseDef: Improved Robustness to Adversarial Attacks via Iterative Denoising

Skewness in imbalanced datasets affects Automatic Text Classification (ATC), leading to classifier bias toward the majority classes. This work examines undersampling methods to mitigate such bias in Small and Large Language Model (SLMs and LLMs) classifiers. Based on the limitations found in existing solutions, we propose two novel undersampling methods inspired by state-of-the-art Instance Selection techniques, relying on calibrated confidences and semantic difficulty estimates. We compare them against 19 baselines across 13 datasets, evaluating: (i) effectiveness, (ii) class imbalance bias, (iii) efficiency, (iv) scalability, and (v) consistency. Results show our methods uniquely reduce classifier bias (up to 56%) across all datasets without effectiveness loss while improving efficiency (1.6x speedup), scalability and reducing carbon emissions (up to 50%).

Instance-Selection-Inspired Undersampling Strategies for Bias Reduction in Small and Large Language Models for Binary Text Classification

Parameter-efficient fine-tuning (PEFT) addresses the memory footprint issue of full fine-tuning by modifying only a subset of model parameters. However, on datasets exhibiting spurious correlations, we observed that PEFT slows down the model's convergence on unbiased examples, while the convergence on biased examples remains fast. This leads to the model's overfitting on biased examples, causing significant performance degradation in out-of-distribution (OOD) scenarios. Traditional debiasing methods mitigate this issue by emphasizing unbiased examples during training but often come at the cost of in-distribution (ID) performance drops. To address this trade-off issue, we propose a curriculum debiasing framework that presents examples in a \textit{biased-to-unbiased} order. Our framework initially limits the model’s exposure to unbiased examples, which are harder to learn, allowing it to first establish a foundation on easier-to-converge biased examples. As training progresses, we gradually increase the proportion of unbiased examples in the training set, guiding the model away from reliance on spurious correlations. Compared to the original PEFT methods, our method accelerates convergence on unbiased examples by approximately twofold and improves ID and OOD performance by 1.2% and 8.0%, respectively.

Curriculum Debiasing: Toward Robust Parameter-Efficient Fine-Tuning Against Dataset Biases

The large language model (LLM)-as-judge paradigm has been used to meet the demand for a cheap, reliable, and fast evaluation of model outputs during AI system development and post-deployment monitoring. While judge models—LLMs finetuned to specialize in assessing and critiquing model outputs—have been touted as general purpose evaluators, they are typically evaluated only on non-contextual scenarios, such as instruction following. The omission of contextual settings—those where external information is used as context to generate an output—is surprising given the increasing prevalence of retrieval-augmented generation (RAG) and summarization use cases. Contextual assessment is uniquely challenging, as evaluation often depends on practitioner priorities, leading to conditional evaluation criteria (e.g., comparing responses based on factuality and then considering completeness if they are equally factual). To address the gap, we propose ContextualJudgeBench, a judge benchmark with 2,000 challenging response pairs across eight splits inspired by real-world contextual evaluation scenarios. We build our benchmark with a multi-pronged data construction pipeline that leverages both existing human annotations and model-based perturbations. Our comprehensive study across 11 judge models and 7 general purpose models, reveals that the contextual information and assessment criteria present a significant challenge to even state-of-the-art models. For example, o1, the best-performing model, barely reaches 55% consistent accuracy.

Does Context Matter? ContextualJudgeBench for Evaluating LLM-based Judges in Contextual Settings

Large Language Models (LLMs) demonstrate exceptional performance across diverse tasks by leveraging pre-trained (i.e., parametric) and external (i.e., contextual) knowledge. While substantial efforts have been made to enhance the utilization of both forms of knowledge, situations in which models lack relevant information remain underexplored. To investigate this challenge, we first present a controlled testbed featuring four distinct knowledge access scenarios, including the aforementioned edge case, revealing that conventional LLM usage exhibits insufficient robustness in handling all instances. Addressing this limitation, we propose Contrastive Decoding with Abstention (CDA), a novel training-free decoding method that allows LLMs to generate responses when relevant knowledge is available and to abstain otherwise. CDA estimates the relevance of both knowledge sources for a given input, adaptively deciding which type of information to prioritize and which to exclude. Through extensive experiments, we demonstrate that CDA can effectively perform accurate generation and abstention simultaneously, enhancing reliability and preserving user trust.

When to Speak, When to Abstain: Contrastive Decoding with Abstention

Supervised fine-tuning (SFT) is widely used to align large language models (LLMs) with information extraction (IE) tasks, such as named entity recognition (NER). However, annotating such fine-grained labels and training domain-specific models is costly. Existing works typically train a unified model across multiple domains, but such approaches lack adaptation and scalability since not all training data benefits target domains and scaling trained models remains challenging. We propose the SaM framework, which dynamically Selects and Merges expert models at inference time. Specifically, for a target domain, we select domain-specific experts pre-trained on existing domains based on (i) domain similarity to the target domain and (ii) performance on sampled instances, respectively. The experts are then merged to create task-specific models optimized for the target domain. By dynamically merging experts beneficial to target domains, we improve generalization across various domains without extra training. Additionally, experts can be added or removed conveniently, leading to great scalability. Extensive experiments on multiple benchmarks demonstrate our framework’s effectiveness, which outperforms the unified model by an average of 10%. We further provide insights into potential improvements, practical experience, and extensions of our framework.

Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models

We introduce the $\underline{Ko}rean \underline{G}rammar \underline{E}valuation Bench\underline{M}ark (KoGEM)$, designed to assess the linguistic competence of LLMs and humans in Korean. KoGEM consists of 1.5k multiple-choice QA pairs covering five main categories and 16 subcategories. The zero-shot evaluation of 27 LLMs of various sizes and types reveals that while LLMs perform remarkably well on straightforward tasks requiring primarily definitional knowledge, they struggle with tasks that demand the integration of real-world experiential knowledge, such as phonological rules and pronunciation. Furthermore, our in-depth analysis suggests that incorporating such experiential knowledge could enhance the linguistic competence of LLMs. With KoGEM, we not only highlight the limitations of current LLMs in linguistic competence but also uncover hidden facets of LLMs in linguistic competence, paving the way for enhancing comprehensive language understanding. Our code and dataset are available at: https://github.com/SungHo3268/KoGEM.

Polishing Every Facet of the GEM: Testing Linguistic Competence of LLMs and Humans in Korean

Language models excel in various tasks by making complex decisions, yet understanding the rationale behind these decisions remains a challenge. This paper investigates \emph{data-centric interpretability} in language models, focusing on the next-word prediction task. Using representer theorem, we identify two types of \emph{support samples}—those that either promote or deter specific predictions. Our findings reveal that being a support sample is an intrinsic property, predictable even before training begins. Additionally, while non-support samples are less influential in direct predictions, they play a critical role in preventing overfitting and shaping generalization and representation learning. Notably, the importance of non-support samples increases in deeper layers, suggesting their significant role in intermediate representation formation.These insights shed light on the interplay between data and model decisions, offering a new dimension to understanding language model behavior and interpretability.

On Support Samples of Next Word Prediction

Recent advancements in generative large language models (LLMs) have enabled wider applicability, accessibility, and flexibility. However, their reliability and trustworthiness are still in doubt, especially for concerns regarding individuals' data privacy. Great efforts have been made on privacy by building various evaluation benchmarks to study LLMs' privacy awareness and robustness from their generated outputs to their hidden representations. Unfortunately, most of these works adopt a narrow formulation of privacy and only investigate personally identifiable information (PII). In this paper, we follow the merit of the Contextual Integrity (CI) theory, which posits that privacy evaluation should not only cover the transmitted attributes but also encompass the whole relevant social context through private information flows. We present PrivaCI-Bench, a comprehensive contextual privacy evaluation benchmark targeted at legal compliance to cover well-annotated privacy and safety regulations, real court cases, privacy policies, and synthetic data built from the official toolkit to study LLMs' privacy and safety compliance. We evaluate the latest LLMs, including the recent reasoner models QwQ-32B and Deepseek R1. Our experimental results suggest that though LLMs can effectively capture key CI parameters inside a given context, they still require further advancements for privacy compliance.

PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliance

Rich user-item interactions are essential for building reliable recommender systems, as they reflect user preference patterns. However, minor language news recommendation platforms suffer from limited interactions due to a small user base. A natural solution is to apply well-established English recommender systems to minor language news recommendation, but the linguistic gap can lead to inaccurate modeling of minor language news content. Therefore, enabling few-shot minor language news recommender systems to capture both content information and preference patterns remains a challenge. Based on the observation that preference patterns are similar across languages, we propose a minor language news recommendation model by cross-lingual preference pattern transfer, named PPT. Our model adopts the widely used two-tower architecture and employs the large language model as the backbone of the news encoder. Through cross-lingual alignment, the strong English capability of the news encoder is extended to minor languages, thus enhancing news content representations. Additionally, through cross-lingual news augmentation, PPT simulates interactions of minor language news in the English domain, which facilitates the transfer of preference patterns from the many-shot English domain to the few-shot minor language domain. Extensive experiments on two real-world datasets across 15 minor languages demonstrate the superiority and generalization of our proposed PPT in addressing minor language news recommendation.

Downloads

Next from ACL 2025

DiffuseDef: Improved Robustness to Adversarial Attacks via Iterative Denoising

Similar lecture

Direct Judgement Preference Optimization