United States

Large Language Models (LLMs) often exhibit homogenized cultural perspectives. While the World Values Survey (WVS) provides a gold standard for mapping human values, traditional direct prompting of LLMs on WVS often fails to access the model&#39;s latent cultural depth, leading to safety-aligned refusals or neutral responses. Here, we propose a generalizable framework for cultural evaluation and intervention that transitions from abstract queries to scenario-based behavioral probing. By extracting implicit token probabilities across 300 situational dilemmas, we bypass surface-level alignment to map the latent coordinates of LLMs cultural value. We further introduce activation steering to shift these internal alignments during the forward pass without retraining. Our results reveal that cultural adaptability is non-uniform. While Llama exhibits structural rigidity, Qwen and Gemma show high but volatile sensitivity across social domains such as family and the legal system. Finally, we quantify latent entanglement, demonstrating that models encode cultural axes as coupled ``modernization&#39;&#39; dimensions. This work establishes a computationally efficient framework for cultural steering, highlighting the structural complexities when navigating global value with LLMs.

ACL 2026

Cultural Value Alignment Via Latent Activation Steering in Large Language Models

### 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**. 

ACL 2025

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

Large Language Models (LLMs) often exhibit homogenized cultural perspectives. While the World Values Survey (WVS) provides a gold standard for mapping human values, traditional direct prompting of LLMs on WVS often fails to access the model's latent cultural depth, leading to safety-aligned refusals or neutral responses. Here, we propose a generalizable framework for cultural evaluation and intervention that transitions from abstract queries to scenario-based behavioral probing. By extracting implicit token probabilities across 300 situational dilemmas, we bypass surface-level alignment to map the latent coordinates of LLMs cultural value. We further introduce activation steering to shift these internal alignments during the forward pass without retraining. Our results reveal that cultural adaptability is non-uniform. While Llama exhibits structural rigidity, Qwen and Gemma show high but volatile sensitivity across social domains such as family and the legal system. Finally, we quantify latent entanglement, demonstrating that models encode cultural axes as coupled ``modernization'' dimensions. This work establishes a computationally efficient framework for cultural steering, highlighting the structural complexities when navigating global value with LLMs.

poster

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields 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.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

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 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

The Advanced Encryption Standard (AES) is currently considered the preferred standard of encryption for secure messaging by the National Institute of Standards and Technology. While its predecessor, the Data Encryption Standard (DES), can be analyzed in mere seconds with modern cryptanalysis algorithms, those same algorithms are unfeasible for cryptanalysis of AES. This is primarily because the key, or the list of values used during encryption, was increased from a length of 56 bits to 128 bits. The list of all possible combinations, also called the probability space, is tens of magnitudes larger in AES than DES. Current DES cryptanalysis methods use mathematical methods to find the most statistically likely approximate key in the entire probability space. The increased probability space in AES is too large to find an approximation using these formulas. However, these methods operate under the assumption that every key is possible. While this is mathematically true, when looked at from the lens of a linguist, many of these keys create messages that are impossible in language. Linguistic attestation is the documentation that a word, grapheme, sound, or other linguistic feature exists in a language. This thesis proposal presents an algorithm that eliminates AES keys using linguistic attestation. I apply the idea that any grammatical message encrypted by AES will not contain any "unattested" data in its input. The algorithm proposes trimming the probability space by removing keys that are unattested. Once a smaller probability space has been created, any method that searches the probability space for a solution may be applied to find the key from the new, smaller list of filtered keys.

Thesis Proposal: Bring Linguistics Back to Cryptanalysis - Using Attestation to Break the Advanced Encryption Standard

Theory of Mind (ToM)—the ability to attribute mental states to others that may diverge from reality—is operationalized through false-belief tasks: a character holds an outdated belief about an object's location (the ToM answer), which conflicts with where the object actually is (the reality answer). Large language models can pass such tests at short context lengths, yet their performance degrades sharply when input is padded with irrelevant tokens—a phenomenon we term token inflation. We use mechanistic interpretability to investigate why this degradation occurs in Llama 3-8B-Instruct. Using a suite of activation patching experiments, we show that token inflation corrupts belief-state representations during early-layer processing, and that this corruption cannot be recovered by the model's own later computation. The failure is specific to belief information: restoring early-layer activations from a clean short-context run rescues up to 16 percentage points of lost accuracy, whereas the same intervention using reality-state or random activations provides no benefit. A head-level analysis further localizes the damage to a distributed circuit of attention heads whose focus on belief-relevant tokens collapses under inflation. These findings generalize across four languages (EN, ZH, JA, KO) and three padding types (dot, Wikipedia, shuffled Wikipedia). Our results demonstrate that token inflation induces a representational collapse of belief states during early-layer processing—pinpointing a concrete mechanistic bottleneck that must be addressed for LLMs to reason reliably about others' mental states.

Believing is Seeing: How Token Inflation Mechanistically Erodes Theory of Mind in Large Language Models

Sentiment analysis involves analyzing text to determine whether the sentiment expressed is positive, negative, or neutral. In the context of online reviews, such as those on Yelp, sentiment analysis helps businesses assess customer satisfaction and identify areas for improvement. Given the large volume of user-generated content, restaurants often struggle to extract actionable insights from feedback, making sentiment analysis an efficient tool for categorizing reviews and highlighting customer concerns. This study focuses on sentiment analysis of Yelp reviews. The main research question is: How can Natural Language Processing (NLP) combined with statistical machine learning methods be applied to classify sentiment in Yelp reviews and provide actionable insights for improving customer satisfaction, service quality, and business performance? The study used 21,000 Yelp reviews, utilizing NLP approaches - tokenization, stop-word removal, and vectorization. Comparative classification predictive modeling and analysis were done across traditional machine learning (Logistic Regression, Support Vector Machine (SVM), Naïve Bayes, Random Forest), deep learning methods (CNN, LSTM, BiLSTM, GRU, RNN), and an advanced transformer-based (RoBERTa) model. Results showed that RoBERTa outperformed the other candidate methods. These findings highlight the potential of advanced NLP techniques to offer businesses practical ways to address customer complaints, enhance service quality, and drive overall business performance.

Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods

Structured span extraction research is siloed by context length, annotation task, and domain, making it difficult to assess how well large language models (LLMs) generalize across realistic extraction settings. We introduce SSA (Structured Span Annotation), a unified evaluation framework bringing together five datasets across four domains: finance, biomedicine, affective analysis, and privacy, under a common JSONL format with character-level offsets. We conduct an exploratory study evaluating seven models (three closed, four open-weight) under three prompting configurations: zero-shot, definition-augmented, and few-shot, formulating extraction as inline XML generation where models reproduce the document with tagged spans. Our results reveal two distinct performance regimes: on tasks requiring complex ontology reasoning, zero-shot performance is near zero (e.g., 0.00\% F1 on FiNER-139) but improves substantially with label definitions (e.g., Claude Opus 4.6 rises from 8.8\% to 57.5\% F1); on pattern-based tasks like PII detection, definitions consistently hurt performance across all models. These findings suggest that prompting strategy must be matched to task structure, and that unified evaluation frameworks spanning varied domains and input lengths are essential for understanding LLM extraction capabilities.

Semantic Span Annotation: An Exploratory Study of LLM Annnotation

Current automated content moderation systems fail to protect children from harmful YouTube content, particularly in under-resourced, code-switched settings. These systems are often text-only, English-centric, and operate as 'black boxes,' lacking the multimodal understanding and transparency needed for effective moderation. This thesis proposes a novel hybrid framework for the explainable multimodal detection of harmful content in videos with code-switching. The proposed framework integrates a fine-tuned classifier for accurate, scalable detection with an LLM-powered module that synthesizes the classifier’s internal evidential signals (e.g., text attention and visual heat maps) to generate faithful, human-readable rationales for each decision. As a primary case study, the framework will be developed and validated on an English–Filipino code-switched dataset. Expected contributions include a new dataset publicly available under controlled access (de-identified transcripts, blurred frames, extracted feature representations, and metadata via data-sharing agreement) and a blueprint for building more equitable, transparent, and trustworthy AI safety systems.

Thesis Proposal: An Explainable Multimodal Framework for Detecting Harmful Content in Code-Switched Children’s Media

Agentic retrieval-augmented generation (RAG) frameworks such as Search-R1 (Jin et al., 2025) improve performance on complex, multi-hop questions by interleaving reasoning and retrieval, but they often retrieve previously seen information and struggle to integrate evidence into the current reasoning state. We study lightweight test-time modifications to Search-R1 that target these inefficiencies without changing training: contextualization of retrieved passages, de-duplication of previously seen documents, and a hybrid of both. Evaluated on 500-example subsets of HotpotQA (Yang et al., 2018) and Natural Questions (Kwiatkowski et al., 2019) using exact match (EM), LLM-based semantic matching, and average retrieval turns, our best method - contextualization with GPT-4.1-mini - improves exact match by 5.6% and reduces retrieval turns by 10.5% relative to the Search-R1 baseline. These results suggest that improved use of retrieved evidence, rather than enforcing retrieval diversity alone, is the main driver of better answer accuracy and efficiency.

Test-Time Strategies for More Efficient and Accurate Agentic RAG

Eye-tracking metrics offer valuable insights into human visual attention during language comprehension, yet existing corpora lack diverse machine-generated text samples. To bridge this gap, we introduce Gaze Responses for Evaluating AI Texts (GREAT), a comprehensive dataset and software framework capturing human eye-movement patterns during screen reading of passages generated by large language models (LLMs). The dataset includes raw eye-movement recordings, reading-time measures, and post-reading evaluations for LLM-generated passage pairs from MT-Bench, alongside rigorous validation metrics. The collected eye-tracking metrics demonstrate strong explanatory power in predicting text quality. When integrated with negative log-likelihood (NLL), a commonly used metric for evaluating text quality, substantially enhance model performance across all standard statistical criteria. These findings demonstrate that eye-tracking data effectively complement probabilistic metrics, improving predictive accuracy for text quality assessment. The full dataset and some processing code are publicly available at https://anonymous.4open.science/r/eye-track.

Eye Movement Features Can Predict Human Preferences on Machine-Generated Texts

Multilingual language models now cover more languages than ever, yet script-sharing low-resource languages remain vulnerable to failures driven by script and dominant-language priors. This dissertation studies one such failure mode, $semantic$ $interference$, in Square Bai Script, where many forms resemble Chinese characters but often differ in meaning. We argue that current adaptation pipelines underperform not only because Bai is low-resource, but because they treat visible overlap as safe transfer by default. Building on an expert-validated corpus of 28,382 Bai-Chinese sentence pairs, an out-of-domain epigraphic set and a reproducible encoding pipeline, the dissertation will (1) diagnose semantic interference, (2) compare adaptation strategies under realistic compute constraints, and (3) estimate when shared-script transfer helps or harms adaptation. The long-term goal is Bai-capable understanding and generation. The dissertation addresses the prerequisite problem of safe and effective adaptation in a script-sharing low-resource setting.

Thesis Proposal: Diagnosing and Mitigating Semantic Interference in Script-Sharing Low-Resource Language Models: A Case Study on Square Bai Script

Fringe platforms like Gab harbor high volumes of hate speech due to minimal moderation and insular communities. Our study examines the factors that determine how hate speech amplifies on these platforms. We scraped a novel dataset of 5K+ threads and 50K+ responses from four fringe platforms (Gab, 4chan, Stormfront, and Vanguard), where hate speech amplification was measured using platform-specific engagement metrics. We trained both Generalized Linear Models and Gradient Boosted Tree models to estimate how several features influence the amplification of hate speech on fringe platforms, and used Shapley value estimates to identify the relative importance of the features. Our analysis shows that research insights from social network analysis (SNA) of mainstream sites like X do not directly generalize to fringe platforms. For instance, our experiments show that using features like thread structure and disagreements in early response windows can give up to 74% RMSE lift in predicting reply counts for hateful posts on fringe platforms, on top of a baseline model with features like hate intensity and thread age that would be considered predictive by regular SNA methods.

Analyzing Hate Speech Amplification on Fringe Platforms

As large language models (LLM) trained on massive corpora scraped from the web exhibit the capability to reproduce sensitive and copyright-protected data, the field of machine unlearning has emerged to address the arising ethical and legal concerns. While previous research has provided a unified evaluation of LLM unlearning methods, this unification remains constrained to English-only models and datasets. We aim to address the prevailing fragmentation in recent cross-lingual unlearning research by extending existing unified benchmarks with multilingual data. To that end, we plan to compile a dataset of parallel translations of question-answer pairs consisting of real-world facts and synthetic personally identifiable information. Moreover, we will focus on mitigating model degradation during unlearning by selectively editing only those layers that contain the given knowledge.

Downloads

Next from ACL 2026

Thesis Proposal: Bring Linguistics Back to Cryptanalysis - Using Attestation to Break the Advanced Encryption Standard

Similar lecture

Cultural Learning-Based Culture Adaptation of Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

Thesis Proposal: Bring Linguistics Back to Cryptanalysis - Using Attestation to Break the Advanced Encryption Standard

Similar lecture

Cultural Learning-Based Culture Adaptation of Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads