Zhixue Zhao

Zhixue Zhao, George Chrysostomou, Kalina Bontcheva et al.

Explanation faithfulness of model predictions in natural language processing is typically evaluated on held-out data from the same temporal distribution as the training data (i.e. synchronous settings). While model performance often deteriorates due to temporal variation (i.e. temporal concept drift), it is currently unknown how explanation faithfulness is impacted when the time span of the target data is different from the data used to train the model (i.e. asynchronous settings). For this purpose, we examine the impact of temporal variation on model explanations extracted by eight feature attribution methods and three select-then-predict models across six text classification tasks. Our experiments show that (i)faithfulness is not consistent under temporal variations across feature attribution methods (e.g. it decreases or increases depending on the method), with an attention-based method demonstrating the most robust faithfulness scores across datasets; and (ii) select-then-predict models are mostly robust in asynchronous settings with only small degradation in predictive performance. Finally, feature attribution methods show conflicting behavior when used in FRESH (i.e. a select-and-predict model) and for measuring sufficiency/comprehensiveness (i.e. as post-hoc methods), suggesting that we need more robust metrics to evaluate post-hoc explanation faithfulness.

George Chrysostomou, Zhixue Zhao, Miles Williams et al.

Despite the remarkable performance of generative large language models (LLMs) on abstractive summarization, they face two significant challenges: their considerable size and tendency to hallucinate. Hallucinations are concerning because they erode reliability and raise safety issues. Pruning is a technique that reduces model size by removing redundant weights, enabling more efficient sparse inference. Pruned models yield downstream task performance comparable to the original, making them ideal alternatives when operating on a limited budget. However, the effect that pruning has upon hallucinations in abstractive summarization with LLMs has yet to be explored. In this paper, we provide an extensive empirical study across five summarization datasets, two state-of-the-art pruning methods, and five instruction-tuned LLMs. Surprisingly, we find that hallucinations are less prevalent from pruned LLMs than the original models. Our analysis suggests that pruned models tend to depend more on the source document for summary generation. This leads to a higher lexical overlap between the generated summary and the source document, which could be a reason for the reduction in hallucination risk.

Simon Kurz, Jian-Jia Chen, Lucie Flek et al.

Recent advances in large language model (LLM) pruning have shown state-of-the-art (SotA) compression results in post-training and retraining-free settings while maintaining high predictive performance. However, previous research mainly considered calibrating based on English text, despite the multilingual nature of modern LLMs and their frequent use in non-English languages. This analysis paper conducts an in-depth investigation of the performance and internal representation changes associated with pruning multilingual language models for monolingual applications. We present the first comprehensive empirical study, comparing different calibration languages for pruning multilingual models across diverse languages, tasks, models, and SotA pruning techniques. We further analyze the latent subspaces, pruning masks, and individual neurons within pruned models. Our results reveal that while calibration on the target language effectively retains perplexity and yields high signal-to-noise ratios, it does not consistently improve downstream task performance. Further analysis of internal representations at three different levels highlights broader limitations of current pruning approaches: While they effectively preserve dominant information like language-specific features, this is insufficient to counteract the loss of nuanced, language-agnostic features that are crucial for knowledge retention and reasoning.

Zhixue Zhao, Nikolaos Aletras

In many real natural language processing application scenarios, practitioners not only aim to maximize predictive performance but also seek faithful explanations for the model predictions. Rationales and importance distribution given by feature attribution methods (FAs) provide insights into how different parts of the input contribute to a prediction. Previous studies have explored how different factors affect faithfulness, mainly in the context of monolingual English models. On the other hand, the differences in FA faithfulness between multilingual and monolingual models have yet to be explored. Our extensive experiments, covering five languages and five popular FAs, show that FA faithfulness varies between multilingual and monolingual models. We find that the larger the multilingual model, the less faithful the FAs are compared to its counterpart monolingual models.Our further analysis shows that the faithfulness disparity is potentially driven by the differences between model tokenizers. Our code is available: https://github.com/casszhao/multilingual-faith.

Zhixue Zhao, Nikolaos Aletras

Feature attribution methods (FAs) are popular approaches for providing insights into the model reasoning process of making predictions. The more faithful a FA is, the more accurately it reflects which parts of the input are more important for the prediction. Widely used faithfulness metrics, such as sufficiency and comprehensiveness use a hard erasure criterion, i.e. entirely removing or retaining the top most important tokens ranked by a given FA and observing the changes in predictive likelihood. However, this hard criterion ignores the importance of each individual token, treating them all equally for computing sufficiency and comprehensiveness. In this paper, we propose a simple yet effective soft erasure criterion. Instead of entirely removing or retaining tokens from the input, we randomly mask parts of the token vector representations proportionately to their FA importance. Extensive experiments across various natural language processing tasks and different FAs show that our soft-sufficiency and soft-comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness. Our code: https://github.com/casszhao/SoftFaith

Jason Chan, Zhixue Zhao, Robert Gaizauskas

Existing work investigates the reasoning capabilities of large language models (LLMs) to uncover their limitations, human-like biases and underlying processes. Such studies include evaluations of base LLMs (pre-trained on unlabeled corpora only) for this purpose. Our position paper argues that evaluating base LLMs' reasoning capabilities raises inherent methodological concerns that are overlooked in such existing studies. We highlight the fundamental mismatch between base LLMs' pretraining objective and normative qualities, such as correctness, by which reasoning is assessed. In particular, we show how base LLMs generate logically valid or invalid conclusions as coincidental byproducts of conforming to purely linguistic patterns of statistical plausibility. This fundamental mismatch challenges the assumptions that (a) base LLMs' outputs can be assessed as their bona fide attempts at correct answers or conclusions; and (b) conclusions about base LLMs' reasoning can generalize to post-trained LLMs optimized for successful instruction-following. We call for a critical re-examination of existing work that relies implicitly on these assumptions, and for future work to account for these methodological pitfalls.

Paul Youssef, Zhixue Zhao, Daniel Braun et al.

Large Language Models (LLMs) contain large amounts of facts about the world. These facts can become outdated over time, which has led to the development of knowledge editing methods (KEs) that can change specific facts in LLMs with limited side effects. This position paper argues that editing LLMs poses serious safety risks that have been largely overlooked. First, we note the fact that KEs are widely available, computationally inexpensive, highly performant, and stealthy makes them an attractive tool for malicious actors. Second, we discuss malicious use cases of KEs, showing how KEs can be easily adapted for a variety of malicious purposes. Third, we highlight vulnerabilities in the AI ecosystem that allow unrestricted uploading and downloading of updated models without verification. Fourth, we argue that a lack of social and institutional awareness exacerbates this risk, and discuss the implications for different stakeholders. We call on the community to (i) research tamper-resistant models and countermeasures against malicious model editing, and (ii) actively engage in securing the AI ecosystem.

Zhixue Zhao, Boxuan Shan

Feature attribution methods (FAs), such as gradients and attention, are widely employed approaches to derive the importance of all input features to the model predictions. Existing work in natural language processing has mostly focused on developing and testing FAs for encoder-only language models (LMs) in classification tasks. However, it is unknown if it is faithful to use these FAs for decoder-only models on text generation, due to the inherent differences between model architectures and task settings respectively. Moreover, previous work has demonstrated that there is no `one-wins-all' FA across models and tasks. This makes the selection of a FA computationally expensive for large LMs since input importance derivation often requires multiple forward and backward passes including gradient computations that might be prohibitive even with access to large compute. To address these issues, we present a model-agnostic FA for generative LMs called Recursive Attribution Generator (ReAGent). Our method updates the token importance distribution in a recursive manner. For each update, we compute the difference in the probability distribution over the vocabulary for predicting the next token between using the original input and using a modified version where a part of the input is replaced with RoBERTa predictions. Our intuition is that replacing an important token in the context should have resulted in a larger change in the model's confidence in predicting the token than replacing an unimportant token. Our method can be universally applied to any generative LM without accessing internal model weights or additional training and fine-tuning, as most other FAs require. We extensively compare the faithfulness of ReAGent with seven popular FAs across six decoder-only LMs of various sizes. The results show that our method consistently provides more faithful token importance distributions.

Ipek Baris Schlicht, Zhixue Zhao, Burcu Sayin et al.

Equitable access to reliable health information is vital for public health, but the quality of online health resources varies by language, raising concerns about inconsistencies in Large Language Models (LLMs) for healthcare. In this study, we examine the consistency of responses provided by LLMs to health-related questions across English, German, Turkish, and Chinese. We largely expand the HealthFC dataset by categorizing health-related questions by disease type and broadening its multilingual scope with Turkish and Chinese translations. We reveal significant inconsistencies in responses that could spread healthcare misinformation. Our main contributions are 1) a multilingual health-related inquiry dataset with meta-information on disease categories, and 2) a novel prompt-based evaluation workflow that enables sub-dimensional comparisons between two languages through parsing. Our findings highlight key challenges in deploying LLM-based tools in multilingual contexts and emphasize the need for improved cross-lingual alignment to ensure accurate and equitable healthcare information.

Paul Youssef, Zhixue Zhao, Christin Seifert et al.

Knowledge editing methods (KEs) can update language models' obsolete or inaccurate knowledge learned from pre-training. However, KEs can be used for malicious applications, e.g., inserting misinformation and toxic content. Knowing whether a generated output is based on edited knowledge or first-hand knowledge from pre-training can increase users' trust in generative models and provide more transparency. Driven by this, we propose a novel task: detecting edited knowledge in language models. Given an edited model and a fact retrieved by a prompt from an edited model, the objective is to classify the knowledge as either unedited (based on the pre-training), or edited (based on subsequent editing). We instantiate the task with four KEs, two LLMs, and two datasets. Additionally, we propose using the hidden state representations and the probability distributions as features for the detection. Our results reveal that, using these features as inputs to a simple AdaBoost classifiers establishes a strong baseline. This classifier requires only a limited amount of data and maintains its performance even in cross-domain settings. Last, we find it more challenging to distinguish edited knowledge from unedited but related knowledge, highlighting the need for further research. Our work lays the groundwork for addressing malicious model editing, which is a critical challenge associated with the strong generative capabilities of LLMs.

Marco Valentino, Geonhee Kim, Dhairya Dalal et al.

Large language models (LLMs) frequently demonstrate reasoning limitations, often conflating content plausibility (i.e., material inference) with logical validity (i.e., formal inference). This can result in biased inferences, where plausible arguments are incorrectly deemed logically valid or vice versa. Mitigating this limitation is critical, as it undermines the trustworthiness and generalizability of LLMs in applications that demand rigorous logical consistency. This paper investigates the problem of mitigating content biases on formal reasoning through activation steering. Specifically, we curate a controlled syllogistic reasoning dataset to disentangle formal validity from content plausibility. After localising the layers responsible for formal and material inference, we investigate contrastive activation steering methods for test-time interventions. An extensive empirical analysis on different LLMs reveals that contrastive steering consistently supports linear control over content biases. However, we observe that a static approach is insufficient for improving all the tested models. We then leverage the possibility to control content effects by dynamically determining the value of the steering parameters via fine-grained conditional methods. We found that conditional steering is effective on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy with a newly introduced kNN-based method (K-CAST). Finally, additional experiments reveal that steering for content effects is robust to prompt variations, incurs minimal side effects on language modeling capabilities, and can partially generalize to out-of-distribution reasoning tasks. Practically, this paper demonstrates that activation-level interventions can offer a scalable strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased formal reasoning.

Leixin Zhang, Steffen Eger, Yinjie Cheng et al.

Multimodal large language models (LLMs) have demonstrated impressive capabilities in generating high-quality images from textual instructions. However, their performance in generating scientific images--a critical application for accelerating scientific progress--remains underexplored. In this work, we address this gap by introducing ScImage, a benchmark designed to evaluate the multimodal capabilities of LLMs in generating scientific images from textual descriptions. ScImage assesses three key dimensions of understanding: spatial, numeric, and attribute comprehension, as well as their combinations, focusing on the relationships between scientific objects (e.g., squares, circles). We evaluate five models, GPT-4o, Llama, AutomaTikZ, Dall-E, and StableDiffusion, using two modes of output generation: code-based outputs (Python, TikZ) and direct raster image generation. Additionally, we examine four different input languages: English, German, Farsi, and Chinese. Our evaluation, conducted with 11 scientists across three criteria (correctness, relevance, and scientific accuracy), reveals that while GPT-4o produces outputs of decent quality for simpler prompts involving individual dimensions such as spatial, numeric, or attribute understanding in isolation, all models face challenges in this task, especially for more complex prompts.

Paul Youssef, Zhixue Zhao, Jörg Schlötterer et al.

In-context knowledge editing (IKE) enables efficient modification of large language model (LLM) outputs without parameter changes and at zero-cost. However, it can be misused to manipulate responses opaquely, e.g., insert misinformation or offensive content. Such malicious interventions could be incorporated into high-level wrapped APIs where the final input prompt is not shown to end-users. To address this issue, we investigate the detection and reversal of IKE-edits. First, we demonstrate that IKE-edits can be detected with high accuracy (F1 > 80\%) using only the top-10 output probabilities of the next token, even in a black-box setting, e.g. proprietary LLMs with limited output information. Further, we introduce the novel task of reversing IKE-edits using specially tuned reversal tokens. We explore using both continuous and discrete reversal tokens, achieving over 80\% accuracy in recovering original, unedited outputs across multiple LLMs. Our continuous reversal tokens prove particularly effective, with minimal impact on unedited prompts. Through analysis of output distributions, attention patterns, and token rankings, we provide insights into IKE's effects on LLMs and how reversal tokens mitigate them. This work represents a significant step towards enhancing LLM resilience against potential misuse of in-context editing, improving their transparency and trustworthiness.

Jake Vasilakes, Carolina Scarton, Zhixue Zhao

Social media's global reach amplifies the spread of information, highlighting the need for robust Natural Language Processing tasks like stance detection across languages and modalities. Prior research predominantly focuses on text-only inputs, leaving multimodal scenarios, such as those involving both images and text, relatively underexplored. Meanwhile, the prevalence of multimodal posts has increased significantly in recent years. Although state-of-the-art Vision-Language Models (VLMs) show promise, their performance on multimodal and multilingual stance detection tasks remains largely unexamined. This paper evaluates state-of-the-art VLMs on a newly extended dataset covering seven languages and multimodal inputs, investigating their use of visual cues, language-specific performance, and cross-modality interactions. Our results show that VLMs generally rely more on text than images for stance detection and this trend persists across languages. Additionally, VLMs rely significantly more on text contained within the images than other visual content. Regarding multilinguality, the models studied tend to generate consistent predictions across languages whether they are explicitly multilingual or not, although there are outliers that are incongruous with macro F1, language support, and model size.

Jason Chan, Robert Gaizauskas, Zhixue Zhao

Formal logic enables computers to reason in natural language by representing sentences in symbolic forms and applying rules to derive conclusions. However, in what our study characterizes as "rulebreaker" scenarios, this method can lead to conclusions that are typically not inferred or accepted by humans given their common sense and factual knowledge. Inspired by works in cognitive science, we create RULEBREAKERS, the first dataset for rigorously evaluating the ability of large language models (LLMs) to recognize and respond to rulebreakers (versus non-rulebreakers) in a human-like manner. Evaluating seven LLMs, we find that most models, including GPT-4o, achieve mediocre accuracy on RULEBREAKERS and exhibit some tendency to over-rigidly apply logical rules unlike what is expected from typical human reasoners. Further analysis suggests that this apparent failure is potentially associated with the models' poor utilization of their world knowledge and their attention distribution patterns. Whilst revealing a limitation of current LLMs, our study also provides a timely counterbalance to a growing body of recent works that propose methods relying on formal logic to improve LLMs' general reasoning capabilities, highlighting their risk of further increasing divergence between LLMs and human-like reasoning.

Jianxin Li, Liang Qu, Taotao Cai et al.

Artificial Intelligence Generated Content (AIGC) has rapidly emerged with the capability to generate different forms of content, including text, images, videos, and other modalities, which can achieve a quality similar to content created by humans. As a result, AIGC is now widely applied across various domains such as digital marketing, education, and public health, and has shown promising results by enhancing content creation efficiency and improving information delivery. However, there are few studies that explore the latest progress and emerging challenges of AIGC across different domains. To bridge this gap, this paper brings together 16 scholars from multiple disciplines to provide a cross-domain perspective on the trends and challenges of AIGC. Specifically, the contributions of this paper are threefold: (1) It first provides a broader overview of AIGC, spanning the training techniques of Generative AI, detection methods, and both the spread and use of AI-generated content across digital platforms. (2) It then introduces the societal impacts of AIGC across diverse domains, along with a review of existing methods employed in these contexts. (3) Finally, it discusses the key technical challenges and presents research propositions to guide future work. Through these contributions, this vision paper seeks to offer readers a cross-domain perspective on AIGC, providing insights into its current research trends, ongoing challenges, and future directions.

Paul Youssef, Zhixue Zhao, Christin Seifert et al.

Knowledge editing methods (KEs) are a cost-effective way to update the factual content of large language models (LLMs), but they pose a dual-use risk. While KEs are beneficial for updating outdated or incorrect information, they can be exploited maliciously to implant misinformation or bias. In order to defend against these types of malicious manipulation, we need robust techniques that can reliably detect, interpret, and mitigate adversarial edits. This work investigates the traceability and reversibility of knowledge edits, focusing on the widely used Rank-One Model Editing (ROME) method. We first show that ROME introduces distinctive distributional patterns in the edited weight matrices, which can serve as effective signals for locating the edited weights. Second, we show that these altered weights can reliably be used to predict the edited factual relation, enabling partial reconstruction of the modified fact. Building on this, we propose a method to infer the edited object entity directly from the modified weights, without access to the editing prompt, achieving over 95% accuracy. Finally, we demonstrate that ROME edits can be reversed, recovering the model's original outputs with $\geq$ 80% accuracy. Our findings highlight the feasibility of detecting, tracing, and reversing edits based on the edited weights, offering a robust framework for safeguarding LLMs against adversarial manipulations.

Yue Li, Zhixue Zhao, Carolina Scarton

In-context learning (ICL) performance is highly sensitive to prompt design, yet the impact of class label options (e.g. lexicon or order) in zero-shot classification remains underexplored. This study proposes LOADS (Label set Optimization via Activation Distribution kurtosiS), a post-hoc method for selecting optimal label sets in zero-shot ICL with large language models (LLMs). LOADS is built upon the observations in our empirical analysis, the first to systematically examine how label option design (i.e., lexical choice, order, and elaboration) impacts classification performance. This analysis shows that the lexical choice of the labels in the prompt (such as agree vs. support in stance classification) plays an important role in both model performance and model's sensitivity to the label order. A further investigation demonstrates that optimal label words tend to activate fewer outlier neurons in LLMs' feed-forward networks. LOADS then leverages kurtosis to measure the neuron activation distribution for label selection, requiring only a single forward pass without gradient propagation or labelled data. The LOADS-selected label words consistently demonstrate effectiveness for zero-shot ICL across classification tasks, datasets, models and languages, achieving maximum performance gain from 0.54 to 0.76 compared to the conventional approach of using original dataset label words.

Samuel Lewis-Lim, Xingwei Tan, Zhixue Zhao et al.

Chain-of-thought (CoT) prompting has emerged as a common technique for enhancing the reasoning abilities of large language models (LLMs). While extended reasoning can boost accuracy on complex tasks, it is often unnecessary and substantially increases token usage, limiting the practicality of reasoning models in many scenarios. Recent models, such as GPT-OSS and Qwen3, expose controls that enable users to adjust the length of CoT or determine whether it is used at all. Yet, it remains unclear when CoT should be used: on some tasks it improves performance, while on others it provides little benefit or even harms performance. We address this challenge with confidence-gated CoT, where a model invokes reasoning only when confidence in its direct answer is low. To this end, we present the first systematic study of training-free confidence estimation methods for CoT gating. Specifically, we evaluate four training-free confidence estimation methods and compare them to a random baseline and an oracle that always knows when CoT is needed. Through extensive experiments, we show that existing training-free confidence measures can reduce redundant CoT and outperform randomly invoked CoT. However, the utility of individual confidence measures is inconsistent, varying with both the dataset and the model, underscoring the difficulty of deploying confidence-gated CoT in practice. By analysing both strengths and failure modes, our study highlights the potential and limitations of current methods and paves the way toward more reliable adaptive gating of CoT.

Ipek Baris Schlicht, Burcu Sayin, Zhixue Zhao et al.

Equitable access to reliable health information is vital when integrating AI into healthcare. Yet, information quality varies across languages, raising concerns about the reliability and consistency of multilingual Large Language Models (LLMs). We systematically examine cross-lingual disparities in pre-training source and factuality alignment in LLM answers for multilingual healthcare Q&A across English, German, Turkish, Chinese (Mandarin), and Italian. We (i) constructed Multilingual Wiki Health Care (MultiWikiHealthCare), a multilingual dataset from Wikipedia; (ii) analyzed cross-lingual healthcare coverage; (iii) assessed LLM response alignment with these references; and (iv) conducted a case study on factual alignment through the use of contextual information and Retrieval-Augmented Generation (RAG). Our findings reveal substantial cross-lingual disparities in both Wikipedia coverage and LLM factual alignment. Across LLMs, responses align more with English Wikipedia, even when the prompts are non-English. Providing contextual excerpts from non-English Wikipedia at inference time effectively shifts factual alignment toward culturally relevant knowledge. These results highlight practical pathways for building more equitable, multilingual AI systems for healthcare.

Yue Li, Jake Vasilakes, Zhixue Zhao et al.

We introduce SCRum-9, the largest multilingual Stance Classification dataset for Rumour analysis in 9 languages, containing 7,516 tweets from X. SCRum-9 goes beyond existing stance classification datasets by covering more languages, linking examples to more fact-checked claims (2.1k), and including confidence-related annotations from multiple annotators to account for intra- and inter-annotator variability. Annotations were made by at least two native speakers per language, totalling more than 405 hours of annotation and 8,150 dollars in compensation. Further, SCRum-9 is used to benchmark five large language models (LLMs) and two multilingual masked language models (MLMs) in In-Context Learning (ICL) and fine-tuning setups. This paper also innovates by exploring the use of multilingual synthetic data for rumour stance classification, showing that even LLMs with weak ICL performance can produce valuable synthetic data for fine-tuning small MLMs, enabling them to achieve higher performance than zero-shot ICL in LLMs. Finally, we examine the relationship between model predictions and human uncertainty on ambiguous cases finding that model predictions often match the second-choice labels assigned by annotators, rather than diverging entirely from human judgments. SCRum-9 is publicly released to the research community with potential to foster further research on multilingual analysis of misleading narratives on social media.

Feng He, Chao Zhang, Zhixue Zhao

Implicit assumptions and priors are often necessary in text-to-image generation tasks, especially when textual prompts lack sufficient context. However, these assumptions can sometimes reflect outdated concepts, inaccuracies, or societal bias embedded in the training data. We present Embedding-only Editing (Embedit), a method designed to efficiently adjust implict assumptions and priors in the model without affecting its interpretation of unrelated objects or overall performance. Given a "source" prompt (e.g., "rose") that elicits an implicit assumption (e.g., rose is red) and a "destination" prompt that specifies the desired attribute (e.g., "blue rose"), Embedit fine-tunes only the word token embedding (WTE) of the target object ("rose") to optimize the last hidden state of text encoder in Stable Diffusion, a SOTA text-to-image model. This targeted adjustment prevents unintended effects on other objects in the model's knowledge base, as the WTEs for unrelated objects and the model weights remain unchanged. Consequently, when a prompt does not contain the edited object, all representations, and the model outputs are identical to those of the original, unedited model. Our method is highly efficient, modifying only 768 parameters for Stable Diffusion 1.4 and 2048 for XL in a single edit, matching the WTE dimension of each respective model. This minimal scope, combined with rapid execution, makes Embedit highly practical for real-world applications. Additionally, changes are easily reversible by restoring the original WTE layers. Our experimental results demonstrate that Embedit consistently outperforms previous methods across various models, tasks, and editing scenarios (both single and sequential multiple edits), achieving at least a 6.01% improvement (from 87.17% to 93.18%).

Samarth N Ramesh, Zhixue Zhao

As text-to-image models grow increasingly powerful and complex, their burgeoning size presents a significant obstacle to widespread adoption, especially on resource-constrained devices. This paper presents a pioneering study on post-training pruning of Stable Diffusion 2, addressing the critical need for model compression in text-to-image domain. Our study tackles the pruning techniques for the previously unexplored multi-modal generation models, and particularly examines the pruning impact on the textual component and the image generation component separately. We conduct a comprehensive comparison on pruning the model or the single component of the model in various sparsities. Our results yield previously undocumented findings. For example, contrary to established trends in language model pruning, we discover that simple magnitude pruning outperforms more advanced techniques in text-to-image context. Furthermore, our results show that Stable Diffusion 2 can be pruned to 38.5% sparsity with minimal quality loss, achieving a significant reduction in model size. We propose an optimal pruning configuration that prunes the text encoder to 47.5% and the diffusion generator to 35%. This configuration maintains image generation quality while substantially reducing computational requirements. In addition, our work uncovers intriguing questions about information encoding in text-to-image models: we observe that pruning beyond certain thresholds leads to sudden performance drops (unreadable images), suggesting that specific weights encode critical semantics information. This finding opens new avenues for future research in model compression, interoperability, and bias identification in text-to-image models. By providing crucial insights into the pruning behavior of text-to-image models, our study lays the groundwork for developing more efficient and accessible AI-driven image generation systems

Zhixue Zhao, Ziqi Zhang, Frank Hopfgartner

Toxic comment classification models are often found biased toward identity terms which are terms characterizing a specific group of people such as "Muslim" and "black". Such bias is commonly reflected in false-positive predictions, i.e. non-toxic comments with identity terms. In this work, we propose a novel approach to tackle such bias in toxic comment classification, leveraging the notion of subjectivity level of a comment and the presence of identity terms. We hypothesize that when a comment is made about a group of people that is characterized by an identity term, the likelihood of that comment being toxic is associated with the subjectivity level of the comment, i.e. the extent to which the comment conveys personal feelings and opinions. Building upon the BERT model, we propose a new structure that is able to leverage these features, and thoroughly evaluate our model on 4 datasets of varying sizes and representing different social media platforms. The results show that our model can consistently outperform BERT and a SOTA model devised to address identity term bias in a different way, with a maximum improvement in F1 of 2.43% and 1.91% respectively.