Morocco

Fine-tuning Large Language Models (LLMs) on sensitive datasets carries a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII), which can violate privacy regulations and compromise individual safety. In this work, we systematically investigate a critical and underexplored vulnerability: the exposure of PII that appears only in model inputs, not in training targets. Using both synthetic and real-world datasets, we design controlled extraction probes to quantify unintended PII memorization and study how factors such as language, PII frequency, task type, and model size influence memorization behavior. We further benchmark four privacy-preserving approaches including differential privacy, machine unlearning, regularization, and preference alignment, evaluating their trade-offs between privacy and task performance. Our results show that post-training methods generally provide more consistent privacy–utility trade-offs, while differential privacy achieves strong reduction in leakage in specific settings, although it can introduce training instability. These findings highlight the persistent challenge of memorization in fine-tuned LLMs and emphasize the need for robust, scalable privacy-preserving techniques.

EACL 2026

Unintended Memorization of Sensitive Information in Fine-Tuned Language Models

llm memorization

safety and alignment

security and privacy

prompting

fine-tuning

robustness

### 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)

technical paper

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/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://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</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 EACL 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.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

Grouping, e.g., grouping channels, which is widely used in current integer-based quantization, has become essential for the emerging MXFP4 format. Ideally, each group should contain channels with similar quantization scales. To guide such groups, existing work clusters the channels using scalar proxy, ignoring the token dimension, which we find suboptimal. In this paper, we propose TAGQuant, a simple yet powerful enhancement for such ``group-wise'' quantization. By strategically shuffling channels to group those with similar token-wise activation distributions, TAGQuant ensures better clustering of large- and small-range values. This shuffle operation is hardware-efficient, and seamlessly integrated into the quantization process with only 0.01x latency overhead. TAGQuant reduces relative GSM8K error in both INT4 and MXFP4 formats, by up to 86\% in Llama-3.1-8B-Instruct compared to baselines, validating the effectiveness of our channel shuffling approach for group-wise quantization. Code is publicly available.

TAGQuant: Token-Aware Clustering for Group-Wise Quantization

Register-Transfer Level (RTL) synthesis and summarization are central to hardware design automation but remain challenging for Large Language Models (LLMs) due to rigid HDL syntax, limited supervision, and weak alignment with natural language. Existing prompting and retrieval-augmented generation (RAG) methods have not incorporated symbolic planning, limiting their structural precision. We introduce SYMDIREC, a neuro-symbolic framework that decomposes RTL tasks into symbolic subgoals, retrieves relevant code via a fine-tuned retriever, and assembles verified outputs through LLM reasoning. Supporting both Verilog and VHDL without LLM fine-tuning, SYMDIREC achieves ~20% higher Pass@1 rates for synthesis and 15–20% ROUGE-L improvements for summarization over prompting and RAG baselines, demonstrating the benefits of symbolic guidance in RTL tasks.

SYMDIREC: A Neuro-Symbolic Divide-Retrieve-Conquer Framework for Enhanced RTL Synthesis and Summarization

Instruction tuning is a widely used approach to improve the instruction-following ability of large language models (LLMs). Instruction-tuning datasets typically include a mixture of context-augmented and context-free examples, yet prior work has largely combined these data types without examining their distinct effects. In this paper, we investigate how training LLMs with or without context affects model behavior and downstream performance. First, in the text domain, we show that LLMs trained with context attend more strongly to the provided knowledge, achieving better grounding. We also observe that context-augmented training shifts how LLMs use knowledge: models store and leverage less on parametric knowledge and instead depend more on the provided context. Second, we observe that using LLM trained with context-augmented data as the backbone for vision-language models reduces hallucination and improves grounding in the visual domain. Finally, we explore practical strategies for real-world deployments where context availability varies. We show that maintaining separate context-augmented and context-free models and routing inputs between them yields more robust overall performance than training a single mixed model, as it better preserves their complementary strengths.

Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact

Large Language Models (LLMs) are increasingly embedded in the daily lives of individuals across diverse social classes. This widespread integration raises urgent concerns about the subtle, implicit biases these models may contain. In this work, we investigate such biases through the lens of ethical reasoning, analyzing model responses to scenarios in a new dataset we propose comprising 1,016 scenarios, systematically categorized into ethical, unethical, and neutral types. Our study focuses on dimensions that are socially influential but less explored, including (i) residency status, (ii) political ideology, (iii) Fitness Status, (iv) educational attainment, and (v) attitudes toward AI. To assess LLMs’ behavior, we propose a baseline and employ one statistical test and one metric: a permutation test that reveals the presence of bias by comparing the probability distributions of ethical/unethical scenarios with the probability distribution of neutral scenarios on each demographic group, and a tendency measurement that captures the magnitude of bias with respect to the relative difference between probability distribution of ethical and unethical scenarios. Our evaluations of 12 prominent LLMs reveal persistent and nuanced biases across all five attributes, and Llama models exhibited the most pronounced biases. These findings highlight the need for refined ethical benchmarks and bias-mitigation tools in LLMs.

Detecting Subtle Biases: An Ethical Lens on Underexplored Areas in AI Language Models Biases

Curriculum learningвЂ”organizing training data from easy to hardвЂ”has improved efficiency across machine learning domains, yet remains underexplored for language model pretraining. We present the first systematic investigation of curriculum learning in LLM pretraining, training over 200 models across three strategies: vanilla curriculum learning, pacing-based sampling, and interleaved curricula. Using six difficulty metrics spanning linguistic and information-theoretic properties, we evaluate performance on eight benchmarks under three realistic scenarios: limited data, unlimited data, and continual training. Our experiments show that curriculum learning consistently accelerates convergence in early and mid-training phases, reducing training steps by 18-45% to reach baseline performance. When applied as a warmup strategy before standard random sampling, curriculum learning yields sustained improvements up to 3.5%. We identify compression ratio, lexical diversity (MTLD), and readability (Flesch Reading Ease) as the most effective difficulty signals. Our findings demonstrate that data orderingвЂ”orthogonal to existing data selection methodsвЂ”provides a practical mechanism for more efficient LLM pretraining.

Beyond Random Sampling: Efficient Language Model Pretraining via Curriculum Learning

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, 5.7 x 102 or 580?'' This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the *log-magnitudes* of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a *pair* of numerals encodes their *ranking*, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe's log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models' internal magnitude representations can enhance their numerical reasoning capabilities.

LLMs Know More About Numbers than They Can Say

Recently, the use of pretrained language models (PLMs) as soft knowledge bases has gained growing interest, sparking the development of knowledge probes to evaluate their factual knowledge retrieval capabilities. However, existing knowledge probes for generative PLMs that support multi-token entities exhibit quadratic time complexity O(n2), where n corresponds to the number of candidate entities, limiting the size of knowledge graphs used for probing. To address this, we propose DEcoder Embedding-based Relational (DEER) probe, utilizing embedding vectors extracted from generative PLMs. DEER probe achieves effective time complexity of linear order O(n), supports rank-based evaluation metrics including Hit@ k, handles multi-token entity names and enables probing whilst disambiguating homographic tail-entity names. We empirically show that DEER-probe correlates with existing knowledge probes, validating its probing capability, and we demonstrate the practical benefits of its improved scalability.

Cosine Similarity as Logits?: A Scalable Knowledge Probe Using Embedding Vectors from Generative Language Models

Knowledge graphs (KGs) provide structured, verifiable grounding for large language models (LLMs), but current LLM-based systems commonly use KGs as auxiliary structures for text retrieval, leaving their intrinsic quality underexplored. In this work, we propose Wikontic, a multi-stage pipeline that constructs KGs from open-domain texts by extracting candidate triplets with qualifiers, enforcing Wikidata-based type and relation constraints, and normalizing entities to reduce duplication. The resulting KGs are compact, ontology-consistent, and well-connected; on MuSiQue, the correct answer entity appears in 96% of generated triplets. On HotpotQA, our triplets-only setup achieves 76.0 F1, and on MuSiQue 59.8 F1, matching or surpassing several retrieval-augmented generation baselines that still require textual context. In addition, Wikontic attains state-of-the-art information-retention performance on the MINE-1 benchmark (86%), outperforming prior KG construction methods. Wikontic is also efficient at build time: KG construction uses less than 1,000 output tokens, about 3 x fewer than AriGraph and < 1/20 of GraphRAG. The proposed pipeline improves the quality of the generated KG and offers a scalable solution for leveraging structured knowledge in LLMs. Wikontic is available at https://github.com/screemix/Wikontic.

Wikontic: Constructing Wikidata-Aligned, Ontology-Aware Knowledge Graphs with Large Language Models

Today, search is used by LLMs to support retrieval-augmented generation as well as by humans. In this new setting, how does what we want from a search engine change, and what are promising approaches to take? I will describe work on two themes related to these questions: (1) retrieving documents to support complex information needs in the context of long-form retrieval-augmented generation and (2) exploring learned sparse retrieval as a promising paradigm for first-stage retrieval. In the former line of work, I will discuss search systems aimed at maximizing coverage of relevant information and approaches for evaluating their ability to do so independent of a generation system. In the latter line of work, I will describe research positioning learned sparse retrieval as a compelling alternative to dense retrieval.

Search with Complex Topics and Learned Sparse Retrieval

Language is one of the richest and most complex human cognitive capacities. Yet, we lack a model organism to study its underlying neural mechanisms: unlike other important cognitive capacities, such as vision or memory, language does not have a clear counterpart in non-human animals, leaving a gap in our ability to develop and test mechanistic hypotheses. In recent years, large language models (LLMs) have emerged as the closest computational analogs we have but how can we use them effectively as model organisms for language in the human brain? In this talk, I will discuss the promise and challenges of this approach. I will present our recent work on brain-tuning using naturalistic brain recordings to refine LLMs so that their internal representations and processing better align with human neural data. But beyond representational similarity, a key question remains: do LLMs rely on mechanisms that are similar to those in the brain? And at what level of abstraction should we assess this similarity? This research direction aims to transform LLMs from mere engineering artifacts into powerful scientific tools for uncovering how the brain supports our most distinctive cognitive ability.

Downloads

Next from EACL 2026

TAGQuant: Token-Aware Clustering for Group-Wise Quantization

Similar lecture

Privacy Ripple Effects from Adding or Removing Personal Information in Language Model Training

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 EACL 2026

TAGQuant: Token-Aware Clustering for Group-Wise Quantization

Similar lecture

Privacy Ripple Effects from Adding or Removing Personal Information in Language Model Training

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads