Nikolaos Aletras

Tong Liu, Cheng Qian, Matej Cief et al.

Tool-calling is a central component of modern large language model (LLM) agents, equipping them with skills beyond their parametric knowledge. This paper studies tool-calling along two complementary axes: effectiveness, i.e., how this capability is measured, and efficiency, i.e., how it is learned. On effectiveness, we systematically analyze tool-calling evaluation pipelines and show that results can be highly sensitive to seemingly minor, often undocumented implementation choices including the random seed, system prompt, multi-turn template construction, and how prior interaction/reasoning history is carried forward. These choices can lead to substantial differences in reported performance, especially in multi-turn settings where without rigorous standardization, leaderboard rankings are unreliable. On efficiency, we examine standard reinforcement learning (RL) for tool-calling and identify two sources of computational waste: (i) during rollouts, many prompts produce no learning signal, and (ii) during policy updates, optimization incurs high computational cost. Guided by these findings, we introduce two techniques that accelerate RL-based tool-calling training, achieving substantial wall-clock speedup without degrading performance.

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.

Miles Williams, Nikolaos Aletras

Quantization and pruning form the foundation of compression for neural networks, enabling efficient inference for large language models (LLMs). Recently, various quantization and pruning techniques have demonstrated remarkable performance in a post-training setting. They rely upon calibration data, a small set of unlabeled examples that are used to generate layer activations. However, no prior work has systematically investigated how the calibration data impacts the effectiveness of model compression methods. In this paper, we present the first extensive empirical study on the effect of calibration data upon LLM performance. We trial a variety of quantization and pruning methods, datasets, tasks, and models. Surprisingly, we find substantial variations in downstream task performance, contrasting existing work that suggests a greater level of robustness to the calibration data. Finally, we make a series of recommendations for the effective use of calibration data in LLM quantization and pruning.

Yida Mu, Kalina Bontcheva, Nikolaos Aletras

New events emerge over time influencing the topics of rumors in social media. Current rumor detection benchmarks use random splits as training, development and test sets which typically results in topical overlaps. Consequently, models trained on random splits may not perform well on rumor classification on previously unseen topics due to the temporal concept drift. In this paper, we provide a re-evaluation of classification models on four popular rumor detection benchmarks considering chronological instead of random splits. Our experimental results show that the use of random splits can significantly overestimate predictive performance across all datasets and models. Therefore, we suggest that rumor detection models should always be evaluated using chronological splits for minimizing topical overlaps.

Ahmed Alajrami, Nikolaos Aletras

Several pre-training objectives, such as masked language modeling (MLM), have been proposed to pre-train language models (e.g. BERT) with the aim of learning better language representations. However, to the best of our knowledge, no previous work so far has investigated how different pre-training objectives affect what BERT learns about linguistics properties. We hypothesize that linguistically motivated objectives such as MLM should help BERT to acquire better linguistic knowledge compared to other non-linguistically motivated objectives that are not intuitive or hard for humans to guess the association between the input and the label to be predicted. To this end, we pre-train BERT with two linguistically motivated objectives and three non-linguistically motivated ones. We then probe for linguistic characteristics encoded in the representation of the resulting models. We find strong evidence that there are only small differences in probing performance between the representations learned by the two different types of objectives. These surprising results question the dominant narrative of linguistically informed pre-training.

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.

Mali Jin, Daniel Preoţiuc-Pietro, A. Seza Doğruöz et al.

Bragging is a speech act employed with the goal of constructing a favorable self-image through positive statements about oneself. It is widespread in daily communication and especially popular in social media, where users aim to build a positive image of their persona directly or indirectly. In this paper, we present the first large scale study of bragging in computational linguistics, building on previous research in linguistics and pragmatics. To facilitate this, we introduce a new publicly available data set of tweets annotated for bragging and their types. We empirically evaluate different transformer-based models injected with linguistic information in (a) binary bragging classification, i.e., if tweets contain bragging statements or not; and (b) multi-class bragging type prediction including not bragging. Our results show that our models can predict bragging with macro F1 up to 72.42 and 35.95 in the binary and multi-class classification tasks respectively. Finally, we present an extensive linguistic and error analysis of bragging prediction to guide future research on this topic.

Mingchen Li, Junfan Chen, Samuel Mensah et al.

Due to the incompleteness of knowledge graphs (KGs), zero-shot link prediction (ZSLP) which aims to predict unobserved relations in KGs has attracted recent interest from researchers. A common solution is to use textual features of relations (e.g., surface name or textual descriptions) as auxiliary information to bridge the gap between seen and unseen relations. Current approaches learn an embedding for each word token in the text. These methods lack robustness as they suffer from the out-of-vocabulary (OOV) problem. Meanwhile, models built on character n-grams have the capability of generating expressive representations for OOV words. Thus, in this paper, we propose a Hierarchical N-Gram framework for Zero-Shot Link Prediction (HNZSLP), which considers the dependencies among character n-grams of the relation surface name for ZSLP. Our approach works by first constructing a hierarchical n-gram graph on the surface name to model the organizational structure of n-grams that leads to the surface name. A GramTransformer, based on the Transformer is then presented to model the hierarchical n-gram graph to construct the relation embedding for ZSLP. Experimental results show the proposed HNZSLP achieved state-of-the-art performance on two ZSLP datasets.

Yida Mu, Pu Niu, Nikolaos Aletras

The phenomenon of misinformation spreading in social media has developed a new form of active citizens who focus on tackling the problem by refuting posts that might contain misinformation. Automatically identifying and characterizing the behavior of such active citizens in social media is an important task in computational social science for complementing studies in misinformation analysis. In this paper, we study this task across different social media platforms (i.e., Twitter and Weibo) and languages (i.e., English and Chinese) for the first time. To this end, (1) we develop and make publicly available a new dataset of Weibo users mapped into one of the two categories (i.e., misinformation posters or active citizens); (2) we evaluate a battery of supervised models on our new Weibo dataset and an existing Twitter dataset which we repurpose for the task; and (3) we present an extensive analysis of the differences in language use between the two user categories.

Xiao Ao, Danae Sánchez Villegas, Daniel Preoţiuc-Pietro et al.

Parody is a figurative device used for mimicking entities for comedic or critical purposes. Parody is intentionally humorous and often involves sarcasm. This paper explores jointly modelling these figurative tropes with the goal of improving performance of political parody detection in tweets. To this end, we present a multi-encoder model that combines three parallel encoders to enrich parody-specific representations with humor and sarcasm information. Experiments on a publicly available data set of political parody tweets demonstrate that our approach outperforms previous state-of-the-art methods.

Constantinos Karouzos, Xingwei Tan, Nikolaos Aletras

Post-trained language models produce less varied outputs than their base counterparts. This output diversity collapse undermines inference-time scaling methods that rely on varied samples, and risks homogenizing model outputs on creative and value-laden tasks. Prior work attributes collapse to specific post-training methods, without separating the role of training data composition from the method, or the generation format from the model weights. We trace output diversity through three parallel post-training lineages of Olmo 3, Think (chain-of-thought distillation), Instruct (broad multi-source data), and RL-Zero, across 15 tasks and four text diversity metrics. We find that the location of collapse co-varies with data composition: the Think lineage loses most semantic diversity at supervised fine-tuning, and the effect of DPO is larger in Instruct than in Think. Suppressing chain-of-thought reasoning at inference in Think models drops accuracy on hard tasks, yet leaves answer-level diversity unchanged, showing that the collapse is embedded in the model weights by training data, not imposed by the generation format. Decomposing diversity loss on six verifiable tasks into a quality-control component (removal of incorrect outputs) and a residual component (genuine narrowing among correct outputs) reveals that the split is task-dependent, and Think models retain more correct-answer diversity than Instruct despite collapsing more in aggregate. Our results indicate that diversity collapse is determined during training by data composition and cannot be addressed at inference time alone.

Danae Sánchez Villegas, Catalina Goanta, Nikolaos Aletras

Influencer marketing involves a wide range of strategies in which brands collaborate with popular content creators (i.e., influencers) to leverage their reach, trust, and impact on their audience to promote and endorse products or services. Because followers of influencers are more likely to buy a product after receiving an authentic product endorsement rather than an explicit direct product promotion, the line between personal opinions and commercial content promotion is frequently blurred. This makes automatic detection of regulatory compliance breaches related to influencer advertising (e.g., misleading advertising or hidden sponsorships) particularly difficult. In this work, we (1) introduce a new Twitter (now X) dataset consisting of 15,998 influencer posts mapped into commercial and non-commercial categories for assisting in the automatic detection of commercial influencer content; (2) experiment with an extensive set of predictive models that combine text and visual information showing that our proposed cross-attention approach outperforms state-of-the-art multimodal models; and (3) conduct a thorough analysis of strengths and limitations of our models. We show that multimodal modeling is useful for identifying commercial posts, reducing the amount of false positives, and capturing relevant context that aids in the discovery of undisclosed commercial posts.

Catalina Goanta, Nikolaos Aletras, Ilias Chalkidis et al.

The scientific innovation in Natural Language Processing (NLP) and more broadly in artificial intelligence (AI) is at its fastest pace to date. As large language models (LLMs) unleash a new era of automation, important debates emerge regarding the benefits and risks of their development, deployment and use. Currently, these debates have been dominated by often polarized narratives mainly led by the AI Safety and AI Ethics movements. This polarization, often amplified by social media, is swaying political agendas on AI regulation and governance and posing issues of regulatory capture. Capture occurs when the regulator advances the interests of the industry it is supposed to regulate, or of special interest groups rather than pursuing the general public interest. Meanwhile in NLP research, attention has been increasingly paid to the discussion of regulating risks and harms. This often happens without systematic methodologies or sufficient rooting in the disciplines that inspire an extended scope of NLP research, jeopardizing the scientific integrity of these endeavors. Regulation studies are a rich source of knowledge on how to systematically deal with risk and uncertainty, as well as with scientific evidence, to evaluate and compare regulatory options. This resource has largely remained untapped so far. In this paper, we argue how NLP research on these topics can benefit from proximity to regulatory studies and adjacent fields. We do so by discussing basic tenets of regulation, and risk and uncertainty, and by highlighting the shortcomings of current NLP discussions dealing with risk assessment. Finally, we advocate for the development of a new multidisciplinary research space on regulation and NLP (RegNLP), focused on connecting scientific knowledge to regulatory processes based on systematic methodologies.

Yida Mu, Xingyi Song, Kalina Bontcheva et al.

A crucial aspect of a rumor detection model is its ability to generalize, particularly its ability to detect emerging, previously unknown rumors. Past research has indicated that content-based (i.e., using solely source posts as input) rumor detection models tend to perform less effectively on unseen rumors. At the same time, the potential of context-based models remains largely untapped. The main contribution of this paper is in the in-depth evaluation of the performance gap between content and context-based models specifically on detecting new, unseen rumors. Our empirical findings demonstrate that context-based models are still overly dependent on the information derived from the rumors' source post and tend to overlook the significant role that contextual information can play. We also study the effect of data split strategies on classifier performance. Based on our experimental results, the paper also offers practical suggestions on how to minimize the effects of temporal concept drift in static datasets during the training of rumor detection methods.

Tulika Bose, Nikolaos Aletras, Irina Illina et al.

Hate speech classifiers exhibit substantial performance degradation when evaluated on datasets different from the source. This is due to learning spurious correlations between words that are not necessarily relevant to hateful language, and hate speech labels from the training corpus. Previous work has attempted to mitigate this problem by regularizing specific terms from pre-defined static dictionaries. While this has been demonstrated to improve the generalizability of classifiers, the coverage of such methods is limited and the dictionaries require regular manual updates from human experts. In this paper, we propose to automatically identify and reduce spurious correlations using attribution methods with dynamic refinement of the list of terms that need to be regularized during training. Our approach is flexible and improves the cross-corpora performance over previous work independently and in combination with pre-defined dictionaries.

Atsuki Yamaguchi, Maggie Mi, Nikolaos Aletras

Language models (LMs) are pre-trained on raw text datasets to generate text sequences token-by-token. While this approach facilitates the learning of world knowledge and reasoning, it does not explicitly optimize for linguistic competence. To bridge this gap, we propose L2T, a pre-training framework integrating Language Learning Tasks alongside standard next-token prediction. Inspired by human language acquisition, L2T transforms raw text into structured input-output pairs to provide explicit linguistic stimulation. Pre-training LMs on a mixture of raw text and L2T data not only improves overall performance on linguistic competence benchmarks but accelerates its acquisition, while maintaining competitive performance on general reasoning tasks.

Huiyin Xue, Nikolaos Aletras

Transformer-based pre-trained language models are vocabulary-dependent, mapping by default each token to its corresponding embedding. This one-to-one mapping results into embedding matrices that occupy a lot of memory (i.e. millions of parameters) and grow linearly with the size of the vocabulary. Previous work on on-device transformers dynamically generate token embeddings on-the-fly without embedding matrices using locality-sensitive hashing over morphological information. These embeddings are subsequently fed into transformer layers for text classification. However, these methods are not pre-trained. Inspired by this line of work, we propose HashFormers, a new family of vocabulary-independent pre-trained transformers that support an unlimited vocabulary (i.e. all possible tokens in a corpus) given a substantially smaller fixed-sized embedding matrix. We achieve this by first introducing computationally cheap hashing functions that bucket together individual tokens to embeddings. We also propose three variants that do not require an embedding matrix at all, further reducing the memory requirements. We empirically demonstrate that HashFormers are more memory efficient compared to standard pre-trained transformers while achieving comparable predictive performance when fine-tuned on multiple text classification tasks. For example, our most efficient HashFormer variant has a negligible performance degradation (0.4\% on GLUE) using only 99.1K parameters for representing the embeddings compared to 12.3-38M parameters of state-of-the-art models.

Danae Sánchez Villegas, Daniel Preoţiuc-Pietro, Nikolaos Aletras

Effectively leveraging multimodal information from social media posts is essential to various downstream tasks such as sentiment analysis, sarcasm detection or hate speech classification. Jointly modeling text and images is challenging because cross-modal semantics might be hidden or the relation between image and text is weak. However, prior work on multimodal classification of social media posts has not yet addressed these challenges. In this work, we present an extensive study on the effectiveness of using two auxiliary losses jointly with the main task during fine-tuning multimodal models. First, Image-Text Contrastive (ITC) is designed to minimize the distance between image-text representations within a post, thereby effectively bridging the gap between posts where the image plays an important role in conveying the post's meaning. Second, Image-Text Matching (ITM) enhances the model's ability to understand the semantic relationship between images and text, thus improving its capacity to handle ambiguous or loosely related modalities. We combine these objectives with five multimodal models across five diverse social media datasets, demonstrating consistent improvements of up to 2.6 points F1. Our comprehensive analysis shows the specific scenarios where each auxiliary task is most effective.

Ahmed Alajrami, Katerina Margatina, Nikolaos Aletras

Understanding how and what pre-trained language models (PLMs) learn about language is an open challenge in natural language processing. Previous work has focused on identifying whether they capture semantic and syntactic information, and how the data or the pre-training objective affects their performance. However, to the best of our knowledge, no previous work has specifically examined how information loss in input token characters affects the performance of PLMs. In this study, we address this gap by pre-training language models using small subsets of characters from individual tokens. Surprisingly, we find that pre-training even under extreme settings, i.e. using only one character of each token, the performance retention in standard NLU benchmarks and probing tasks compared to full-token models is high. For instance, a model pre-trained only on single first characters from tokens achieves performance retention of approximately $90$\% and $77$\% of the full-token model in SuperGLUE and GLUE tasks, respectively.

Miles Williams, Nikolaos Aletras

The extensive memory footprint of language model (LM) fine-tuning poses a challenge for both researchers and practitioners. LMs use an embedding matrix to represent extensive vocabularies, forming a substantial proportion of the model parameters. While previous work towards memory-efficient fine-tuning has focused on minimizing the number of trainable parameters, reducing the memory footprint of the embedding matrix has yet to be explored. We first demonstrate that a significant proportion of the vocabulary remains unused during fine-tuning. We then propose a simple yet effective approach that leverages this finding to minimize memory usage. We show that our approach provides substantial reductions in memory usage across a wide range of models and tasks. Notably, our approach does not impact downstream task performance, while allowing more efficient use of computational resources.

Wenzhe Li, Nikolaos Aletras

Graph-based text representation focuses on how text documents are represented as graphs for exploiting dependency information between tokens and documents within a corpus. Despite the increasing interest in graph representation learning, there is limited research in exploring new ways for graph-based text representation, which is important in downstream natural language processing tasks. In this paper, we first propose a new heterogeneous word-character text graph that combines word and character n-gram nodes together with document nodes, allowing us to better learn dependencies among these entities. Additionally, we propose two new graph-based neural models, WCTextGCN and WCTextGAT, for modeling our proposed text graph. Extensive experiments in text classification and automatic text summarization benchmarks demonstrate that our proposed models consistently outperform competitive baselines and state-of-the-art graph-based models.

Kai Sun, Richong Zhang, Samuel Mensah et al.

Opinion target extraction (OTE) or aspect extraction (AE) is a fundamental task in opinion mining that aims to extract the targets (or aspects) on which opinions have been expressed. Recent work focus on cross-domain OTE, which is typically encountered in real-world scenarios, where the testing and training distributions differ. Most methods use domain adversarial neural networks that aim to reduce the domain gap between the labelled source and unlabelled target domains to improve target domain performance. However, this approach only aligns feature distributions and does not account for class-wise feature alignment, leading to suboptimal results. Semi-supervised learning (SSL) has been explored as a solution, but is limited by the quality of pseudo-labels generated by the model. Inspired by the theoretical foundations in domain adaptation [2], we propose a new SSL approach that opts for selecting target samples whose model output from a domain-specific teacher and student network disagree on the unlabelled target data, in an effort to boost the target domain performance. Extensive experiments on benchmark cross-domain OTE datasets show that this approach is effective and performs consistently well in settings with large domain shifts.

Huiyin Xue, Nikolaos Aletras

Scaling pre-trained language models has resulted in large performance gains in various natural language processing tasks but comes with a large cost in memory requirements. Inspired by the position embeddings in transformers, we aim to simplify and reduce the memory footprint of the multi-head attention (MHA) mechanism. We propose an alternative module that uses only a single shared projection matrix and multiple head embeddings (MHE), i.e. one per head. We empirically demonstrate that our MHE attention is substantially more memory efficient compared to alternative attention mechanisms while achieving high predictive performance retention ratio to vanilla MHA on several downstream tasks. MHE attention only requires a negligible fraction of additional parameters ($3nd$, where $n$ is the number of attention heads and $d$ the size of the head embeddings) compared to a single-head attention, while MHA requires $(3n^2-3n)d^2-3nd$ additional parameters.

Tulika Bose, Nikolaos Aletras, Irina Illina et al.

State-of-the-art approaches for hate-speech detection usually exhibit poor performance in out-of-domain settings. This occurs, typically, due to classifiers overemphasizing source-specific information that negatively impacts its domain invariance. Prior work has attempted to penalize terms related to hate-speech from manually curated lists using feature attribution methods, which quantify the importance assigned to input terms by the classifier when making a prediction. We, instead, propose a domain adaptation approach that automatically extracts and penalizes source-specific terms using a domain classifier, which learns to differentiate between domains, and feature-attribution scores for hate-speech classes, yielding consistent improvements in cross-domain evaluation.

Atsuki Yamaguchi, Terufumi Morishita, Aline Villavicencio et al.

Expanding the linguistic diversity of instruct large language models (LLMs) is crucial for global accessibility but is often hindered by the reliance on costly specialized target language labeled data and catastrophic forgetting during adaptation. We tackle this challenge under a realistic, low-resource constraint: adapting instruct LLMs using only unlabeled target language data. We introduce Source-Shielded Updates (SSU), a selective parameter update strategy that proactively preserves source knowledge. Using a small set of source data and a parameter importance scoring method, SSU identifies parameters critical to maintaining source abilities. It then applies a column-wise freezing strategy to protect these parameters before adaptation. Experiments across five typologically diverse languages and 7B and 13B models demonstrate that SSU successfully mitigates catastrophic forgetting. It reduces performance degradation on monolingual source tasks to just 3.4% (7B) and 2.8% (13B) on average, a stark contrast to the 20.3% and 22.3% from full fine-tuning. SSU also achieves target-language performance highly competitive with full fine-tuning, outperforming it on all benchmarks for 7B models and the majority for 13B models.

Anthony Hughes, Alexander Goldberg, Prince Jha et al.

Safety classifiers are essential safeguards within generative AI systems, filtering harmful content or identifying at-risk users when interacting with large language models. Despite their necessity, these models are trained on sensitive datasets including discussions of self-harm and mental health, raising important, yet poorly understood, privacy concerns. Membership inference attacks (MIAs) allow adversaries to infer membership of examples used to train models. In this work, we hypothesize that identifying the examples on which the classifier is least confident are informative for an adversary to infer membership. This reflects a localized failure of generalization, where the model relies on memorization to resolve ambiguity in the training set. To investigate this, we introduce a new boundary-targeted selection strategy that identifies low confidence examples that amplify the signal of an examples membership within a training set. Our experimental results show that an adversary can recover 19\% of the conversations a safety classifier flagged as indicating user distress, at a 5\% false-positive rate, on a classifier fine-tuned for detecting a user who may require emotional support. This is $3.5$ times more than attacking using state-of-the-art MIA methods alone. Finally, we characterize the boundary laying examples and show that content-based filtering is ineffective for protection, and existing noise strategies can effectively mitigate susceptibility of these examples.

Danae Sánchez Villegas, Samuel Lewis-Lim, Nikolaos Aletras et al.

Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear. We analyze reasoning dynamics in 18 VLMs covering instruction-tuned and reasoning-trained models from two different model families. We track confidence over Chain-of-Thought (CoT), measure the corrective effect of reasoning, and evaluate the contribution of intermediate reasoning steps. We find that models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, we show that models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues, but their longer and fluent CoTs can still appear visually grounded while actually following textual cues, obscuring modality reliance. In contrast, instruction-tuned models refer to the cues less explicitly, but their shorter traces reveal inconsistencies with the visual input. Taken together, these findings indicate that CoT provides only a partial view of how different modalities drive VLM decisions, with important implications for the transparency and safety of multimodal systems.

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

Daniel Preotiuc-Pietro, Mihaela Gaman, Nikolaos Aletras

Complaining is a basic speech act regularly used in human and computer mediated communication to express a negative mismatch between reality and expectations in a particular situation. Automatically identifying complaints in social media is of utmost importance for organizations or brands to improve the customer experience or in developing dialogue systems for handling and responding to complaints. In this paper, we introduce the first systematic analysis of complaints in computational linguistics. We collect a new annotated data set of written complaints expressed in English on Twitter.\footnote{Data and code is available here: \url{https://github.com/danielpreotiuc/complaints-social-media}} We present an extensive linguistic analysis of complaining as a speech act in social media and train strong feature-based and neural models of complaints across nine domains achieving a predictive performance of up to 79 F1 using distant supervision.

Mingzi Cao, Xingwei Tan, Mahmud Elahi Akhter et al.

Deduction, induction, and abduction are fundamental reasoning paradigms, core for human logical thinking. Although improving Large Language Model (LLM) reasoning has attracted significant research efforts, the extent to which the fundamental paradigms induce generalization has yet to be systematically explored. In this study, we shed light on how the interplay between these core paradigms influences LLMs' reasoning behavior. To this end, we first collect a new dataset of reasoning trajectories from symbolic tasks, each targeting one of the three fundamental paradigms, to abstract from concrete world knowledge. Then, we investigate effective ways for inducing these skills into LLMs. We experiment with a battery of methods including simple fine-tuning, and more complex approaches to increase model depth, or transform a dense model to a mixture-of-experts. We comprehensively evaluate induced models on realistic out-of-domain tasks, that are entirely formulated in natural language and contain real-world knowledge. Our results reveal that our approach yields strong generalizability with substantial performance gains (up to $14.60$) across realistic tasks.

Constantinos Karouzos, Xingwei Tan, Nikolaos Aletras

Preference tuning aligns pretrained language models to human judgments of quality, helpfulness, or safety by optimizing over explicit preference signals rather than likelihood alone. Prior work has shown that preference-tuning degrades performance and reduces helpfulness when evaluated outside the training domain. However, the extent to which adaptation strategies mitigate this domain shift remains unexplored. We address this challenge by conducting a comprehensive and systematic study of alignment generalization under domain shift. We compare five popular alignment objectives and various adaptation strategies from source to target, including target-domain supervised fine-tuning and pseudo-labeling, across summarization and question-answering helpfulness tasks. Our findings reveal systematic differences in generalization across alignment objectives under domain shift. We show that adaptation strategies based on pseudo-labeling can substantially reduce domain-shift degradation

Atsuki Yamaguchi, Aline Villavicencio, Nikolaos Aletras

The development of state-of-the-art generative large language models (LLMs) disproportionately relies on English-centric tokenizers, vocabulary and pre-training data. Despite the fact that some LLMs have multilingual capabilities, recent studies have shown that their inference efficiency deteriorates when generating text in languages other than English. This results in increased inference time and costs. Cross-lingual vocabulary adaptation (CVA) methods have been proposed for adapting models to a target language aiming to improve downstream performance. However, the effectiveness of these methods on increasing inference efficiency of generative LLMs has yet to be explored. In this paper, we perform an empirical study of five CVA methods on four generative LLMs (including monolingual and multilingual models) across four typologically-diverse languages and four natural language understanding tasks. We find that CVA substantially contributes to LLM inference speedups of up to 271.5\%. We also show that adapting LLMs that have been pre-trained on more balanced multilingual data results in downstream performance comparable to the original models.

Xingwei Tan, Marco Valentino, Mahmud Elahi Akhter et al.

Large Language Models (LLMs) are known to acquire reasoning capabilities through shared inference patterns in pre-training data, which are further elicited via Chain-of-Thought (CoT) practices. However, whether fundamental reasoning patterns, such as induction, deduction, and abduction, can be decoupled from specific problem instances remains a critical challenge for model controllability, and for shedding light on reasoning controllability. In this paper, we present the first systematic investigation of this problem through the lens of reasoning conflicts: an explicit tension between parametric and contextual information induced by mandating logical schemata that deviate from those expected for a target task. Our evaluation reveals that LLMs consistently prioritize sensibility over compliance, favoring task-appropriate reasoning patterns despite conflicting instructions. Notably, task accuracy is not strictly determined by sensibility, with models often maintaining high performance even when using conflicting patterns, suggesting a reliance on internalized parametric memory that increases with model size. We further demonstrate that reasoning conflicts are internally detectable, as confidence scores significantly drop during conflicting episodes. Probing experiments confirm that reasoning types are linearly encoded from middle-to-late layers, indicating the potential for activation-level controllability. Leveraging these insights, we steer models towards compliance, increasing instruction following by up to 29%. Overall, our findings establish that while LLM reasoning is anchored to concrete instances, active mechanistic interventions can effectively decouple logical schemata from data, offering a path toward improved controllability, faithfulness, and generalizability.

Peter Vickers, Loïc Barrault, Emilio Monti et al. · amazon-science, meta-ai

In Natural Language Processing (NLP) classification tasks such as topic categorisation and sentiment analysis, model generalizability is generally measured with standard metrics such as Accuracy, F-Measure, or AUC-ROC. The diversity of metrics, and the arbitrariness of their application suggest that there is no agreement within NLP on a single best metric to use. This lack suggests there has not been sufficient examination of the underlying heuristics which each metric encodes. To address this we compare several standard classification metrics with more 'exotic' metrics and demonstrate that a random-guess normalised Informedness metric is a parsimonious baseline for task performance. To show how important the choice of metric is, we perform extensive experiments on a wide range of NLP tasks including a synthetic scenario, natural language understanding, question answering and machine translation. Across these tasks we use a superset of metrics to rank models and find that Informedness best captures the ideal model characteristics. Finally, we release a Python implementation of Informedness following the SciKitLearn classifier format.

Anthony Hughes, Ning Ma, Nikolaos Aletras

In sensitive domains such as medical and legal, protecting sensitive information is critical, with protective laws strictly prohibiting the disclosure of personal data. This poses challenges for sharing valuable data such as medical reports and legal cases summaries. While language models (LMs) have shown strong performance in text summarization, it is still an open question to what extent they can provide privacy-preserving summaries from non-private source documents. In this paper, we perform a comprehensive study of privacy risks in LM-based summarization across two closed- and four open-weight models of different sizes and families. We experiment with both prompting and fine-tuning strategies for privacy-preservation across a range of summarization datasets including medical and legal domains. Our quantitative and qualitative analysis, including human evaluation, shows that LMs frequently leak personally identifiable information in their summaries, in contrast to human-generated privacy-preserving summaries, which demonstrate significantly higher privacy protection levels. These findings highlight a substantial gap between current LM capabilities and expert human expert performance in privacy-sensitive summarization tasks.

Atsuki Yamaguchi, Terufumi Morishita, Aline Villavicencio et al.

Vocabulary expansion (VE) is the de-facto approach to language adaptation of large language models (LLMs) by adding new tokens and continuing pre-training on target data. While this is effective for base models trained on unlabeled data, it poses challenges for chat models trained to follow instructions through labeled conversation data. Directly adapting the latter with VE on target unlabeled data may result in forgetting chat abilities. While ideal, target chat data is often unavailable or costly to create for low-resource languages, and machine-translated alternatives are not always effective. To address this issue, previous work proposed using a base and chat model from the same family. This method first adapts the base LLM with VE on target unlabeled data and then converts it to a chat model by adding a chat vector (CV) derived from the weight difference between the source base and chat models. We propose ElChat, a new language adaptation method for chat LLMs that adapts a chat model directly on target unlabeled data, without a base model. It elicits chat abilities by injecting information from the source chat model. ElChat offers more robust and competitive target language and safety performance while achieving superior English, chat, and instruction-following abilities compared to CV.

Mali Jin, Daniel Preoţiuc-Pietro, A. Seza Doğruöz et al.

Bragging is the act of uttering statements that are likely to be positively viewed by others and it is extensively employed in human communication with the aim to build a positive self-image of oneself. Social media is a natural platform for users to employ bragging in order to gain admiration, respect, attention and followers from their audiences. Yet, little is known about the scale of bragging online and its characteristics. This paper employs computational sociolinguistics methods to conduct the first large scale study of bragging behavior on Twitter (U.S.) by focusing on its overall prevalence, temporal dynamics and impact of demographic factors. Our study shows that the prevalence of bragging decreases over time within the same population of users. In addition, younger, more educated and popular users in the U.S. are more likely to brag. Finally, we conduct an extensive linguistics analysis to unveil specific bragging themes associated with different user traits.

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

Recent work has demonstrated that Chain-of-Thought (CoT) often yields limited gains for soft-reasoning problems such as analytical and commonsense reasoning. CoT can also be unfaithful to a model's actual reasoning. We investigate the dynamics and faithfulness of CoT in soft-reasoning tasks across instruction-tuned, reasoning and reasoning-distilled models. Our findings reveal differences in how these models rely on CoT, and show that CoT influence and faithfulness are not always aligned.

Odysseas S. Chlapanis, Dimitrios Galanis, Nikolaos Aletras et al.

We introduce GreekBarBench, a benchmark that evaluates LLMs on legal questions across five different legal areas from the Greek Bar exams, requiring citations to statutory articles and case facts. To tackle the challenges of free-text evaluation, we propose a three-dimensional scoring system combined with an LLM-as-a-judge approach. We also develop a meta-evaluation benchmark to assess the correlation between LLM-judges and human expert evaluations, revealing that simple, span-based rubrics improve their alignment. Our systematic evaluation of 13 proprietary and open-weight LLMs shows that even though the best models outperform average expert scores, they fall short of the 95th percentile of experts.

Miles Williams, George Chrysostomou, Vitor Jeronymo et al.

Compression techniques such as pruning and quantization offer a solution for more efficient deployment of language models (LMs), albeit with small performance drops in benchmark performance. However, general-purpose LM compression methods can negatively affect performance in specialized domains (e.g. biomedical or legal). Recent work has sought to address this, yet requires computationally expensive full-parameter fine-tuning. To this end, we propose cross-calibration, a novel training-free approach for improving the domain performance of compressed LMs. Our approach effectively leverages Hessian-based sensitivity to identify weights that are influential for both in-domain and general performance. Through extensive experimentation, we demonstrate that cross-calibration substantially outperforms existing approaches on domain-specific tasks, without compromising general performance. Notably, these gains come without additional computational overhead, displaying remarkable potential towards extracting domain-specialized compressed models from general-purpose LMs.

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.

Huiyin Xue, Nafise Sadat Moosavi, Nikolaos Aletras

The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.

Anthony Hughes, Vasisht Duddu, N. Asokan et al.

Language models (LMs) may memorize personally identifiable information (PII) from training data, enabling adversaries to extract it during inference. Existing defense mechanisms such as differential privacy (DP) reduce this leakage, but incur large drops in utility. Based on a comprehensive study using circuit discovery to identify the computational circuits responsible PII leakage in LMs, we hypothesize that specific PII leakage circuits in LMs should be responsible for this behavior. Therefore, we propose PATCH (Privacy-Aware Targeted Circuit PatcHing), a novel approach that first identifies and subsequently directly edits PII circuits to reduce leakage. PATCH achieves better privacy-utility trade-off than existing defenses, e.g., reducing recall of PII leakage from LMs by up to 65%. Finally, PATCH can be combined with DP to reduce recall of residual leakage of an LM to as low as 0.01%. Our analysis shows that PII leakage circuits persist even after the application of existing defense mechanisms. In contrast, PATCH can effectively mitigate their impact.

Ahmed Alajrami, Xingwei Tan, Nikolaos Aletras

Instruction-tuning plays a vital role in enhancing the task-solving abilities of large language models (LLMs), improving their usability in generating helpful responses on various tasks. However, previous work has demonstrated that they are sensitive to minor variations in instruction phrasing. In this paper, we explore whether introducing perturbations in instruction-tuning data can enhance LLMs' resistance against noisy instructions. We focus on how instruction-tuning with perturbations, such as removing stop words or shuffling words, affects LLMs' performance on the original and perturbed versions of widely-used benchmarks (MMLU, BBH, GSM8K). We further assess learning dynamics and potential shifts in model behavior. Surprisingly, our results suggest that instruction-tuning on perturbed instructions can, in some cases, improve downstream performance. These findings highlight the importance of including perturbed instructions in instruction-tuning, which can make LLMs more resilient to noisy user inputs.

Mingzi Cao, Xi Wang, Nikolaos Aletras

Scaling Large Language Models (LLMs) yields performance gains but incurs substantial training costs. Depth up-scaling offers training efficiency by adding new layers to pre-trained models. However, most existing methods copy or average weights from base layers, neglecting neuron permutation differences. This limitation can potentially cause misalignment that harms performance. Inspired by applying Optimal Transport (OT) for neuron alignment, we propose Optimal Transport Depth Up-Scaling (OpT-DeUS). OpT-DeUS aligns and fuses Transformer blocks in adjacent base layers via OT for new layer creation, to mitigate neuron permutation mismatch between layers. OpT-DeUS achieves better overall performance and offers improved training efficiency than existing methods for continual pre-training and supervised fine-tuning across different model sizes. To further evaluate the impact of interpolation positions, our extensive analysis shows that inserting new layers closer to the top results in higher training efficiency due to shorter back-propagation time while obtaining additional performance gains.

Chuan Meng, Francesco Tonolini, Fengran Mo et al.

Proactive search in conversations (PSC) aims to reduce user effort in formulating explicit queries by proactively retrieving useful relevant information given conversational context. Previous work in PSC either directly uses this context as input to off-the-shelf ad-hoc retrievers or further fine-tunes them on PSC data. However, ad-hoc retrievers are pre-trained on short and concise queries, while the PSC input is longer and noisier. This input mismatch between ad-hoc search and PSC limits retrieval quality. While fine-tuning on PSC data helps, its benefits remain constrained by this input gap. In this work, we propose Conv2Query, a novel conversation-to-query framework that adapts ad-hoc retrievers to PSC by bridging the input gap between ad-hoc search and PSC. Conv2Query maps conversational context into ad-hoc queries, which can either be used as input for off-the-shelf ad-hoc retrievers or for further fine-tuning on PSC data. Extensive experiments on two PSC datasets show that Conv2Query significantly improves ad-hoc retrievers' performance, both when used directly and after fine-tuning on PSC.

Miles Williams, George Chrysostomou, Nikolaos Aletras

Quantization and pruning are fundamental approaches for model compression, enabling efficient inference for language models. In a post-training setting, state-of-the-art quantization and pruning methods require calibration data, a small set of unlabeled examples. Conventionally, this is randomly sampled web text, aiming to reflect the model training data. However, this poses two key problems: (1) unrepresentative calibration examples can harm model performance, and (2) organizations increasingly avoid releasing model training data. In this paper, we propose self-calibration as a solution. Our approach requires no external data, instead leveraging the model itself to generate synthetic calibration data, with a view to better approximating the pre-training data distribution. We extensively compare the performance of self-calibration with several baselines, across a variety of models, compression methods, and tasks. Our approach proves consistently competitive in maximizing downstream task performance, frequently outperforming even using real data.

Atsuki Yamaguchi, Aline Villavicencio, Nikolaos Aletras

Large language models (LLMs) have shown remarkable capabilities in many languages beyond English. Yet, LLMs require more inference steps when generating non-English text due to their reliance on English-centric tokenizers and vocabulary, resulting in higher usage costs to non-English speakers. Vocabulary expansion with target language tokens is a widely used cross-lingual vocabulary adaptation approach to remedy this issue. Despite its effectiveness in inference speedup, previous work on vocabulary expansion has focused on high-resource settings assuming access to a substantial amount of target language data to effectively initialize the embeddings of the new tokens and adapt the LLM to the target language. However, vocabulary expansion in low-resource settings has yet to be explored. In this article, we investigate vocabulary expansion in low-resource settings by considering embedding initialization methods and continual pre-training strategies. Through extensive experiments across typologically diverse languages, tasks and models, we establish a set of strategies to perform vocabulary expansion for faster inference, while striving to maintain competitive downstream performance to baselines. This is achieved with only 30K sentences ($\sim$0.01GB text data) from the target language.

Yue Feng, Yunlong Jiao, Animesh Prasad et al.

User Satisfaction Modeling (USM) is one of the popular choices for task-oriented dialogue systems evaluation, where user satisfaction typically depends on whether the user's task goals were fulfilled by the system. Task-oriented dialogue systems use task schema, which is a set of task attributes, to encode the user's task goals. Existing studies on USM neglect explicitly modeling the user's task goals fulfillment using the task schema. In this paper, we propose SG-USM, a novel schema-guided user satisfaction modeling framework. It explicitly models the degree to which the user's preferences regarding the task attributes are fulfilled by the system for predicting the user's satisfaction level. SG-USM employs a pre-trained language model for encoding dialogue context and task attributes. Further, it employs a fulfillment representation layer for learning how many task attributes have been fulfilled in the dialogue, an importance predictor component for calculating the importance of task attributes. Finally, it predicts the user satisfaction based on task attribute fulfillment and task attribute importance. Experimental results on benchmark datasets (i.e. MWOZ, SGD, ReDial, and JDDC) show that SG-USM consistently outperforms competitive existing methods. Our extensive analysis demonstrates that SG-USM can improve the interpretability of user satisfaction modeling, has good scalability as it can effectively deal with unseen tasks and can also effectively work in low-resource settings by leveraging unlabeled data.