United States

Warning: This paper contains offensive language. Ensuring that large language models (LLMs) do not generate harmful text is critical for their safe deployment. A common failure mode involves producing toxic responses to otherwise innocuous prompts. While various detoxification methods have been proposed, the underlying mechanisms that drive toxic generation in LLMs are not yet fully understood. Our work aims to provide a mechanistic understanding of toxic generation against innocuous-seeming adversarial prompts through the lens of memory localization. We find evidence of localization of toxic memories in the early Multilayer Perceptron (MLP) layers of GPT-2-XL. We further investigate the effects of editing and deleting these toxic memories in MLP layers to reduce toxic generation. Editing significantly reduces toxic generation, from 62.86% to 28.61%. However, this reduction comes with a trade-off in generation quality as perplexity increases from 78.18 on GPT2-XL against the adversarial prompts to 106.06 after editing. Localization-informed deletion achieves a better toxicity-perplexity tradeoff compared to random early layer editing, which reduces toxicity but leads to greater perplexity increases.

NAACL 2025

On Localizing and Deleting Toxic Memories in Large Language Models

mechanistic interpretability

unlearning

model editing

toxicity

alignment

safety

### 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)

poster

### 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!

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

Interventions targeting the representation space of language models (LMs) have emerged as an effective means to influence model behavior. Such methods are employed, for example, to eliminate or alter the encoding of demographic information such as gender within the model's representations and, in so doing, create a counterfactual representation. However, because the intervention operates within the representation space, understanding precisely what aspects of the text it modifies poses a challenge. In this paper, we give a method to convert representation counterfactuals into string counterfactuals. We demonstrate that this approach enables us to analyze the linguistic alterations corresponding to a given representation space intervention and to interpret the features utilized to encode a specific concept. Moreover, the resulting counterfactuals can be used to mitigate bias in classification through data augmentation.

A Practical Method for Generating String Counterfactuals

The performance of Large Language Models~(LLMs) on many tasks is greatly limited by the knowledge learned during pre-training and stored in the model's parameters. Low-rank adaptation (LoRA) is a popular and efficient training technique for updating or domain-specific adaptation of LLMs. In this study, we investigate how new facts can be incorporated into the LLM using LoRA without compromising the previously learned knowledge. We fine-tuned Llama-3.1-8B-instruct using LoRA with varying amounts of new knowledge. Our experiments have shown that the best results are obtained when the training data contains a mixture of known and new facts. However, this approach is still potentially harmful because the model's performance on external question-answering benchmarks declines after such fine-tuning. When the training data is biased towards certain entities, the model tends to regress to few overrepresented answers. In addition, we found that the model becomes more confident and refuses to provide an answer in only few cases. These findings highlight the potential pitfalls of LoRA-based LLM updates and underscore the importance of training data composition and tuning parameters to balance new knowledge integration and general model capabilities.

How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?

Trustworthiness in model predictions is crucial for safety-critical applications in the real world. However, deep neural networks often suffer from the issues of uncertainty estimation, such as miscalibration. In this study, we propose k-Nearest Neighbor Uncertainty Estimation (kNN-UE), which is a new uncertainty estimation method that uses not only the distances from the neighbors, but also the ratio of labels in the neighbors. Experiments on sentiment analysis, natural language inference, and named entity recognition show that our proposed method outperforms the baselines and recent density-based methods in several calibration and uncertainty metrics. Moreover, our analyses indicate that approximate nearest neighbor search techniques reduce the inference overhead without significantly degrading the uncertainty estimation performance when they are appropriately combined.

Efficient Nearest Neighbor based Uncertainty Estimation for Natural Language Processing Tasks

The k-nearest-neighbor language model (kNNLM), one of the retrieval-augmented language models, improves the perplexity for given text by directly accessing a large datastore built from any text data during inference. A widely held hypothesis for the success of kNN-LM is that its explicit memory, i.e., the datastore, enhances predictions for long-tail phenomena.
However, prior works have primarily shown its ability to retrieve long-tail contexts, leaving the model’s performance remain underexplored in estimating the probabilities of long-tail target tokens during inference. In this paper, we
investigate the behavior of kNN-LM on lowfrequency tokens, examining prediction probability, retrieval accuracy, token distribution in the datastore, and approximation error of the product quantization. Our experimental results reveal that kNN-LM does not improve prediction performance for low-frequency tokens but mainly benefits high-frequency tokens regardless of long-tail contexts in the datastore.1

Long-Tail Crisis in Nearest Neighbor Language Models

Sparse Autoencoders (SAEs) are a promising approach for extracting neural network representations by learning a sparse and overcomplete decomposition of the network's internal activations. However, SAEs are traditionally trained considering only activation values and not the effect those activations have on downstream computations. This limits the information available to learn features, and biases the autoencoder towards neglecting features which are represented with small activation values but strongly influence model outputs. To address this, we introduce Gradient SAEs (g-SAEs), which modify the k-sparse autoencoder architecture by augmenting the TopK activation function to rely on the gradients of the input activation when selecting the k elements. For a given sparsity level, g-SAEs produce reconstructions that are more faithful to original network performance when propagated through the network. Additionally, we find evidence that g-SAEs learn latents that are on average more effective at steering models in arbitrary contexts. By considering the downstream effects of activations, our approach leverages the dual nature of neural network features as both representations, retrospectively, and actions, prospectively. While previous methods have approached the problem of feature discovery primarily focused on the former aspect, g-SAEs represent a step towards accounting for the latter as well.

Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning

In tasks such as summarization and open-book question answering (QA), Large Language Models (LLMs) frequently experience ''contextual hallucination'', where they generate irrelevant or incorrect responses despite having access to accurate information in the input. This issue often stems from the models' propensity to prioritize self-generated content over input context, leading to a disregard for pertinent details. To address this challenge, we introduce, Guided Attention Map Editing (GAME), an innovative approach that dynamically adjusts attention maps to enhance contextual relevance. During inference, GAME employs a trained classifier to identify attention maps likely to induce hallucinations and implements targeted interventions. These interventions, guided by gradient-informed ``edit directions'', strategically redistribute attention weights across various heads to efficiently mitigate hallucination. Extensive evaluations on challenging summarization and open-book QA tasks demonstrate that GAME consistently and significantly reduces hallucinations across diverse open-source models, thereby improving the reliability and applicability of LLMs.

Gradient-guided Attention Map Editing: Towards Efficient Contextual Hallucination Mitigation

Pre-trained language models (PLMs) are known to be susceptible to perturbations to the input text, but existing works do not explicitly focus on linguistically grounded attacks, which are subtle and more prevalent in nature. In this paper, we study whether PLMs are agnostic to linguistically grounded attacks or not. To this end, we offer the first study addressing this, investigating different Indic languages and various downstream tasks. Our findings reveal that although PLMs are susceptible to linguistic perturbations, when compared to non-linguistic attacks, PLMs exhibit a slightly lower susceptibility to linguistic attacks. This highlights that even constrained attacks are effective. Moreover, we investigate the implications of these outcomes across a range of languages, encompassing diverse language families and different scripts.

Are Language Models Agnostic to Linguistically Grounded Perturbations? A Case Study of Indic Languages

Utilizing large language models (LLMs) to rank a set of items has become a common approach in recommendation and retrieval systems. Typically, these systems focus on ordering a substantial number of documents in a monotonic order based on a given query. However, real-world scenarios often present a different challenge: ranking a comparatively smaller set of items, but according to a variety of diverse and occasionally conflicting conditions. In this paper, we define and explore the task of multi-conditional ranking by introducing MCRank, a benchmark tailored for assessing multi-conditional ranking across various item types and conditions. Our analysis of LLMs using MCRank indicates a significant decrease in performance as the number and complexity of items and conditions grow. To overcome this limitation, we propose a novel decomposed reasoning method, consisting of EXtracting and Sorting the conditions, and then Iteratively Ranking the items (EXSIR). Our extensive experiments show that this decomposed reasoning method enhances LLMs' performance significantly, achieving up to a 14.4% improvement over existing LLMs. We also provide a detailed analysis of LLMs performance across various condition categories, and examine the effectiveness of decomposition step. Furthermore, we compare our method with existing approaches such as Chain-of-Thought and existing ranking models, demonstrating the superiority of our approach and complexity of MCR task. We will make our dataset and code publicly available.

Multi-Conditional Ranking with Large Language Models

Recent research efforts have explored the potential of leveraging natural language inference (NLI) techniques to enhance relation extraction (RE). In this vein, we introduce MetaEntail-RE, a novel adaptation method that harnesses NLI principles to enhance RE performance. Our approach follows past works by verbalizing relation classes into class-indicative hypotheses, aligning a traditionally multi-class classification task to one of textual entailment. We introduce three key enhancements: (1) Meta-class analysis which, instead of labeling non-entailed premise-hypothesis pairs with the less informative ''neutral'' entailment label, provides additional context by analyzing overarching meta-relationships between classes; (2) Feasible hypothesis filtering, which removes unlikely hypotheses from consideration based on domain knowledge derived from data; and (3) Group-based prediction selection, which further improves performance by selecting highly confident predictions. MetaEntail-RE is conceptually simple and empirically powerful, yielding significant improvements over conventional relation extraction techniques and other NLI formulations. We observe surprisingly large F1 gains of 17.6 points on BioRED and 13.4 points on ReTACRED compared to conventional methods, underscoring the versatility of MetaEntail-RE across both biomedical and general domains.

Entangled Relations: Leveraging NLI and Meta-analysis to Enhance Biomedical Relation Extraction

Large language models (LLMs) have been used to generate query expansions augmenting original queries for improving information search. Recent studies also explore providing LLMs with initial retrieval results to generate query expansions more grounded to document corpus. However, these methods mostly focus on enhancing textual similarities between search queries and target documents, overlooking document relations. For queries like ''Find me a highly rated camera for wildlife photography compatible with my Nikon F-Mount lenses'', existing methods may generate expansions that are semantically similar but structurally unrelated to user intents. To handle such semi-structured queries with both textual and relational requirements, in this paper we propose a knowledge-aware query expansion framework, augmenting LLMs with structured document relations from knowledge graph (KG). To further address the limitation of entity-based scoring in existing KG-based methods, we leverage document texts as rich KG node representations and use document-based relation filtering for our Knowledge-Aware Retrieval (KAR). Extensive experiments on three datasets of diverse domains show the advantages of our method compared against state-of-the-art baselines on textual and relational semi-structured retrieval.

Downloads

Next from NAACL 2025

A Practical Method for Generating String Counterfactuals

Similar lecture

Contrastive Perplexity for Controlled Generation: An Application in Detoxifying Large Language Models