Austria

Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, red-teaming techniques are often limited, and malicious actors may discover new vulnerabilities that bypass safety fine-tuning, underscoring the need for ongoing research and innovative approaches. Notably, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing AI systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content mitigate bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. and use adversarial attacking strategies to induce biased responses from several open- and closed-source language models. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, addressing adversarial challenges to ensure AI safety in high-stakes industry deployments.

ACL 2025

Red-Teaming for Uncovering Societal Bias in Large Language Models

## Welcome to 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics!
This year, the conference is in Mexico City. NAACL was actually already planned for Mexico City in 2021, but due to the pandemic the entire conference was moved online. This year, finally, we get to go! So it is my sincere pleasure to welcome you to Mexico City, whether in person or virtually. Having the conference in Mexico City is a good opportunity to emphasize that NAACL is our flagship conference for ACL members not only in North America but also in Central and South America, even though NAACL has been bearing “North” in its name. At this year’s conference, we have a theme to match, with a theme track on the Languages of Latin America to showcase the linguistic diversity of the region.
 
The opportunity to present at NAACL should not depend on a researcher’s travel budget, or their family status. This is why it is so important to make virtual participation at NAACL as good an experience as possible – but we want to also provide a good experience for in-person participants. As a community, we are still working out the best way to do that. This year at NAACL, we are trying out a big virtual poster session ahead of the conference, with the hope that this will make make for a lively and interactive experience. At the same time, we are reducing virtual oral presentations, which seem to be particularly tricky to make to work well. A big thanks to the NAACL program chairs and to Luciana Benotti for all their ideas and work to improve the virtual experience. And participants, virtual as well as in-person: Please let us know what worked for you and what didn’t, so we can continue to improve hybrid conferences. 
 
I have been lucky to work with many amazing people. Without their insight, dedication and patience, and without the many hours of work they put in, NAACL would not have been possible. A huge thank you to the program chairs Helena Gomez, Kevin Duh, and Steve Bethard – you are the best! 
 
Finally, I would like to thank all authors, invited speakers and panelists, area chairs and reviewers, the volunteers organizing and chairing sessions, and all attendees, in-person and virtual. Thank you for helping us make NAACL 2024 come to life. 
 
Welcome and hope you all enjoy the conference! 
Katrin Erk 
The University of Texas at Austin 
NAACL-2024 General Chair 
*You can read the full Welcome message in the [Conference Handbook (downloadable)](https://drive.google.com/file/d/1H1NvW0VASQjkSCYgw3mr-3l4yYDSxz6B/view?usp=sharing)*

To access the event page you need to register [**here.**](http://acl.swoogo.com/naacl2024) 
Your access to the event page is limited based on your registration type. If you registered for workshops only, you will gain access to full workshops content on the day of the workshops program.

Please register!

NAACL 2024

2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics

workshop paper

### 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)

Long-context understanding is crucial for large language models (LLMs) and has become a fundamental capability for most LLMs. However, beyond the focus on "input-long", the ability to "output-long" is equally significant, yet it remains underexplored. To address this limitation, we propose a simple, efficient, and plug-in approach, Position ID Compression (PIC), to unlock the long-form text generation potential of LLMs. The idea is straightforward: by compressing the position ids of the context, we provoke and guide LLMs to generate coherent and longer output. Specifically, we find that directly reducing the position ids by a fixed ratio significantly impacts the generation quality. To mitigate this, we propose two variants of PIC: NTK-aware PIC and Dynamic PIC. Without additional training, both methods enable LLMs to extend their generation length by approximately 1.5 times without compromising generation quality. Furthermore, by integrating supervised fine-tuning (SFT) with PIC, we propose PIC-SFT, which further improves LLMs' long-form text generation capabilities, achieving top performance on HelloBench and LongBench-Write. Extensive experiments demonstrate the effectiveness of our approach.

PIC: Unlocking Long-Form Text Generation Capabilities of Large Language Models via Position ID Compression

RAG systems rely on rerankers to identify relevant documents. However, fine-tuning these models remains challenging due to the scarcity of annotated query-document pairs. Existing distillation-based approaches suffer from training-inference misalignment and fail to capture interdependencies among candidate documents. To overcome these limitations, we reframe the reranking process as an attention-mask problem and propose Gumbel Reranking, an end-to-end training framework for rerankers aimed at minimizing the training-inference gap. In our approach, reranker optimization is reformulated as learning a stochastic, document-wise Top-$k$ attention mask using the Gumbel Trick and Relaxed Top-$k$ Sampling. This formulation enables end-to-end optimization by minimizing the overall language loss. Experiments across various settings consistently demonstrate performance gains, including a 10.4% improvement in recall on HotpotQA for distinguishing indirectly relevant documents.

Gumbel Reranking: Differentiable End-to-End Reranker Optimization

Knowledge neuron theory provides a key approach to understanding the mechanisms of factual knowledge in Large Language Models (LLMs), which suggests that facts are stored within multi-layer perceptron neurons. This paper further explores **Degenerate Knowledge Neurons** (DKNs), where distinct sets of neurons can store identical facts, but unlike simple redundancy, they also participate in storing other different facts. Despite the novelty and unique properties of this concept, it has not been rigorously defined and systematically studied. Our contributions are: (1) We pioneer the study of structures in knowledge neurons by analyzing weight connection patterns, providing a comprehensive definition of DKNs from both functional and structural aspects. (2) Based on this definition, we develop the **Neuronal Topology Clustering** method, leading to a more accurate DKN identification. (3) We demonstrate the practical applications of DKNs in two aspects: guiding LLMs to learn new knowledge and relating to LLMs' robustness against input errors.

Cracking Factual Knowledge: A Comprehensive Analysis of Degenerate Knowledge Neurons in Large Language Models

We demonstrate that features, rather than neurons, serve as superior analytical units for understanding the mechanisms of factual knowledge in Language Models (LMs). Previous studies primarily utilize MLP neurons as units of analysis; however, neurons suffer from polysemanticity, leading to limited knowledge expression and poor interpretability. We first conduct preliminary experiments to validate that SAE can effectively decompose neurons into features. With this established, our core findings reveal three key advantages of features over neurons: (1) Features exhibit stronger influence on knowledge expression and superior interpretability. (2) Features demonstrate enhanced monosemanticity, showing distinct activation patterns between related and unrelated facts. (3) Feature-based method demonstrates superior performance over neuron-based approaches in erasing privacy-sensitive information from LMs. Additionally, we propose FeatureEdit, the first feature-based editing method. Code and dataset will be available.

The Knowledge Microscope: Features as Better Analytical Lenses than Neurons

Knowledge Graph Embedding (KGE) is a common approach for Knowledge Graphs (KGs) in AI tasks. Embedding dimensions depend on application scenarios. Requiring a new dimension means training a new KGE model from scratch, increasing cost and limiting efficiency and flexibility. In this work, we propose a novel KGE training framework MED. It allows one training to obtain a croppable KGE model for multiple scenarios with different dimensional needs. Sub-models of required dimensions can be directly cropped and used without extra training. In MED, we propose a mutual learning mechanism to improve the low-dimensional sub-models and make high-dimensional sub-models retain the low-dimensional sub-models' capacity, an evolutionary improvement mechanism to promote the high-dimensional sub-models to master the triple that the low-dimensional sub-models can not, and a dynamic loss weight to adaptively balance the multiple losses. Experiments on 4 KGE models across 4 standard KG completion datasets, 3 real-world scenarios using a large-scale KG, and extending MED to the BERT language model demonstrate its effectiveness, high efficiency, and flexible extensibility.

Croppable Knowledge Graph Embedding

In this work, we introduce MultiReflect, a novel multimodal self-reflective Retrieval Augmented Generation (RAG)-based automated fact-checking pipeline. MultiReflect is designed to address the challenges of rapidly outdated information, limitations in human query capabilities, and expert knowledge barriers in fact-checking. Our proposed pipeline leverages the latest advancements in Large Language Models (LLMs) and Retrieval Augmented Generation (RAG) to enhance fact verification across text and images. Specifically, by integrating multimodal data processing with RAG-based evidence reflection, our system improves the accuracy of fact-checking by utilizing internet-sourced verification. We evaluate our results on the VERITE benchmarks and using several multimodal LLMs, outperforming baselines in binary classification.

MultiReflect: Multimodal Self-Reflective RAG-based Automated Fact-Checking

Retrieval-augmented generation (RAG) is a powerful paradigm for leveraging external data to enhance the capabilities of large language models (LLMs). However, most existing RAG solutions are tailored for single-modality or limited multimodal scenarios, restricting their applicability in real-world contexts where diverse data sources—including text, tables, images, and videos—must be integrated seamlessly. In this work proposes a unified \textit{Multimodal Retrieval-augmented generation (mRAG)} system designed to unify information processing across all four modalities. Our pipeline ingests and indexes data from PDFs and videos using tools like Amazon Textract, Transcribe, Langfuse, and multimodal LLMs (e.g., Claude 3.5 Sonnet) for structured extraction and semantic enrichment. The dataset includes text queries, table lookups, image-based questions, and videos. Evaluation with the Deepeval framework shows improved retrieval accuracy and response quality, especially for structured text and tables. While performance on image and video queries is lower, the multimodal integration framework remains robust, underscoring the value of unified pipelines for diverse data.

Multimodal Retrieval-Augmented Generation: Unified Information Processing Across Text, Image, Table, and Video Modalities

Annotator disagreement poses a significant challenge in subjective tasks like hate speech detection. In this paper, we introduce a novel variant of the HateWiC task that explicitly models annotator agreement by estimating the proportion of annotators who classify the meaning of a term as hateful. To tackle this challenge, we explore the use of Llama 3 models fine-tuned through Direct Preference Optimization (DPO). Our experiments show that while LLMs perform well for majority-based hate classification, they struggle with the more complex agreement-aware task. DPO fine-tuning offers improvements, particularly when applied to instruction-tuned models. Yet, our results emphasize the need for improved modeling of subjectivity in hate classification and this study can serve as foundation for future advancements.

Using LLMs and Preference Optimization for Agreement-Aware HateWiC Classification

Language changes over time, including in the hate speech domain, which evolves quickly following social dynamics and cultural shifts. While NLP research has investigated the impact of language evolution on model training and has proposed several solutions for it, its impact on model benchmarking remains under-explored. Yet, hate speech benchmarks play a crucial role to ensure model safety. In this paper, we empirically evaluate the robustness of 20 language models across two evolving hate speech experiments, and we show the temporal misalignment between static and time-sensitive evaluations. Our findings call for time-sensitive linguistic benchmarks in order to correctly and reliably evaluate language models in the hate speech domain.

Hatevolution: What Static Benchmarks Don't Tell Us

Large language model training follows a standard pipeline: tokenization, pretraining, possibly mid-training, and post training or alignment. Despite its wild success, we understand relatively little about this recipe and are almost certainly missing many opportunities to improve it. In this talk, I will focus on three such cases. I’ll describe our work on data efficient post training (e.g. LIMA, ALMA, and s1) where we argue that nearly all advanced model capabilities ultimately come from the pretraining data, even if effective alignment is still essential for controlling model behavior. I will also describe new methods for extracting more signal from the pretraining data, including new hierarchical architectures for byte-level language models (e.g. BLT) that are both tokenizer-free and scale better than traditional BPE-based methods, especially in the long tail. Finally, I will discuss decentralized, modular training algorithms (e.g. BTM) that better isolate and control the influence of specific data on specific model components and behaviors. Together, these methods promise to simplify training and improve scaling, by centering and amplifying the influence of data in architecture design.

Downloads

Next from ACL 2025

PIC: Unlocking Long-Form Text Generation Capabilities of Large Language Models via Position ID Compression

Similar lecture

BiasKG: Adversarial Knowledge Graphs to Induce Bias in Large Language Models

.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 2025

PIC: Unlocking Long-Form Text Generation Capabilities of Large Language Models via Position ID Compression

Similar lecture

BiasKG: Adversarial Knowledge Graphs to Induce Bias in Large Language Models

Downloads