Lucie Flek

Nicholas Kluge Corrêa, Aniket Sen, Shiza Fatimah et al.

We present Tucano 2, a fully open suite of large language models (LLMs) with 0.5-3.7 billion parameters, designed to address certain gaps in open-source development for Portuguese LLMs. Following our previous works, we now extend our dataset, GigaVerbo-v2, to a new degree of quality and scale, while also introducing a new synthetic dataset, GigaVerbo-v2 Synth, aimed at filling missing gaps in GigaVerbo-v2, and two post-training datasets, GigaVerbo-v2 SFT and GigaVerbo-v2 Preferences, that allow Portuguese LLMs to be trained in domains like retrieval augmented generation, coding, tool use, chain-of-thought reasoning, and many other domains of interest. Through extensive ablation studies, we design both pretraining and continual pretraining recipes for the Tucano 2 suite (Base, Instruct, and Think), which achieve state-of-the-art performance on several Portuguese-language modeling benchmarks. We also extend and refine the evaluation harness introduced in our earlier work, yielding a comprehensive evaluation suite that provides strong signals across different pretraining, continual pretraining, and post-training regimes. All artifacts associated with Tucano 2 are openly released, including training recipes, logs, and source code, ensuring that our work is reproducible, accessible, and extendable by the broader Portuguese NLP community.

Ipek Baris Schlicht, Lucie Flek, Paolo Rosso

Check-worthiness detection is the task of identifying claims, worthy to be investigated by fact-checkers. Resource scarcity for non-world languages and model learning costs remain major challenges for the creation of models supporting multilingual check-worthiness detection. This paper proposes cross-training adapters on a subset of world languages, combined by adapter fusion, to detect claims emerging globally in multiple languages. (1) With a vast number of annotators available for world languages and the storage-efficient adapter models, this approach is more cost efficient. Models can be updated more frequently and thus stay up-to-date. (2) Adapter fusion provides insights and allows for interpretation regarding the influence of each adapter model on a particular language. The proposed solution often outperformed the top multilingual approaches in our benchmark tasks.

Vahid Sadiri Javadi, Martin Potthast, Lucie Flek

When searching for products, the opinions of others play an important role in making informed decisions. Subjective experiences about a product can be a valuable source of information. This is also true in sales conversations, where a customer and a sales assistant exchange facts and opinions about products. However, training an AI for such conversations is complicated by the fact that language models do not possess authentic opinions for their lack of real-world experience. We address this problem by leveraging product reviews as a rich source of product opinions to ground conversational AI in true subjective narratives. With OpinionConv, we develop the first conversational AI for simulating sales conversations. To validate the generated conversations, we conduct several user studies showing that the generated opinions are perceived as realistic. Our assessors also confirm the importance of opinions as an informative basis for decision-making.

Victor Petrén Bach Hansen, Atula Tejaswi Neerkaje, Ramit Sawhney et al.

The performance cost of differential privacy has, for some applications, been shown to be higher for minority groups; fairness, conversely, has been shown to disproportionally compromise the privacy of members of such groups. Most work in this area has been restricted to computer vision and risk assessment. In this paper, we evaluate the impact of differential privacy on fairness across four tasks, focusing on how attempts to mitigate privacy violations and between-group performance differences interact: Does privacy inhibit attempts to ensure fairness? To this end, we train $(\varepsilon,δ)$-differentially private models with empirical risk minimization and group distributionally robust training objectives. Consistent with previous findings, we find that differential privacy increases between-group performance differences in the baseline setting; but more interestingly, differential privacy reduces between-group performance differences in the robust setting. We explain this by reinterpreting differential privacy as regularization.

Magnus Jørgenvåg, David Kaczér, Lasse Ruttert et al.

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with interleaving on-policy safety data performing best.

Allison Lahnala, Charles Welch, David Jurgens et al.

We review the state of research on empathy in natural language processing and identify the following issues: (1) empathy definitions are absent or abstract, which (2) leads to low construct validity and reproducibility. Moreover, (3) emotional empathy is overemphasized, skewing our focus to a narrow subset of simplified tasks. We believe these issues hinder research progress and argue that current directions will benefit from a clear conceptualization that includes operationalizing cognitive empathy components. Our main objectives are to provide insight and guidance on empathy conceptualization for NLP research objectives and to encourage researchers to pursue the overlooked opportunities in this area, highly relevant, e.g., for clinical and educational sectors.

Niclas Heilig, Jan Kirchhoff, Florian Stumpe et al.

Medical diagnosis is the process of making a prediction of the disease a patient is likely to have, given a set of symptoms and observations. This requires extensive expert knowledge, in particular when covering a large variety of diseases. Such knowledge can be coded in a knowledge graph -- encompassing diseases, symptoms, and diagnosis paths. Since both the knowledge itself and its encoding can be incomplete, refining the knowledge graph with additional information helps physicians making better predictions. At the same time, for deployment in a hospital, the diagnosis must be explainable and transparent. In this paper, we present an approach using diagnosis paths in a medical knowledge graph. We show that those graphs can be refined using latent representations with RDF2vec, while the final diagnosis is still made in an explainable way. Using both an intrinsic as well as an expert-based evaluation, we show that the embedding-based prediction approach is beneficial for refining the graph with additional valid conditions.

Lucie Flek, Oliver Janik, Philipp Alexander Jung et al.

In this paper, we present a new algorithm, MiniFool, that implements physics-inspired adversarial attacks for testing neural network-based classification tasks in particle and astroparticle physics. While we initially developed the algorithm for the search for astrophysical tau neutrinos with the IceCube Neutrino Observatory, we apply it to further data from other science domains, thus demonstrating its general applicability. Here, we apply the algorithm to the well-known MNIST data set and furthermore, to Open Data data from the CMS experiment at the Large Hadron Collider. The algorithm is based on minimizing a cost function that combines a $χ^2$ based test-statistic with the deviation from the desired target score. The test statistic quantifies the probability of the perturbations applied to the data based on the experimental uncertainties. For our studied use cases, we find that the likelihood of a flipped classification differs for both the initially correctly and incorrectly classified events. When testing changes of the classifications as a function of an attack parameter that scales the experimental uncertainties, the robustness of the network decision can be quantified. Furthermore, this allows testing the robustness of the classification of unlabeled experimental data.

Shangrui Nie, Kian Omoomi, Lucie Flek et al.

Pluralism, the capacity to engage with diverse perspectives without collapsing them into a single viewpoint, is critical for developing large language models that faithfully reflect human heterogeneity. Yet this characteristic has not been carefully examined in the LLM research community and remains absent from most alignment studies. Debate-oriented sources provide a natural entry point for pluralism research. Previous work builds on online debate sources but remains constrained by costly human validation. Other debate-rich platforms such as Reddit and Kialo also offer promising material: Reddit provides linguistic diversity and scale but lacks clear argumentative structure, while Kialo supplies explicit pro/con graphs but remains overly concise and detached from natural discourse. We introduce PERSPECTRA, a pluralist benchmark that integrates the structural clarity of Kialo debate graphs with the linguistic diversity of real Reddit discussions. Using a controlled retrieval-and-expansion pipeline, we construct 3,810 enriched arguments spanning 762 pro/con stances on 100 controversial topics. Each opinion is expanded to multiple naturalistic variants, enabling robust evaluation of pluralism. We initialise three tasks with PERSPECTRA: opinion counting (identifying distinct viewpoints), opinion matching (aligning supporting stances and discourse to source opinions), and polarity check (inferring aggregate stance in mixed discourse). Experiments with state-of-the-art open-source and proprietary LLMs, highlight systematic failures, such as overestimating the number of viewpoints and misclassifying concessive structures, underscoring the difficulty of pluralism-aware understanding and reasoning. By combining diversity with structure, PERSPECTRA establishes the first scalable, configurable benchmark for evaluating how well models represent, distinguish, and reason over multiple perspectives.

Khashayar Alavi, Zhastay Yeltay, Lucie Flek et al.

When LLM agents work together, they seem to be more powerful than a single LLM in mathematical question answering. However, are they also more robust to adversarial inputs? We investigate this question using adversarially perturbed math questions. These perturbations include punctuation noise with three intensities (10, 30, and 50 percent), plus real-world and human-like typos (WikiTypo, R2ATA). Using a unified sampling-and-voting framework (Agent Forest), we evaluate six open-source models (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B) across four benchmarks (GSM8K, MATH, MMLU-Math, MultiArith), with various numbers of agents n from one to 25 (1, 2, 5, 10, 15, 20, 25). Our findings show that (1) Noise type matters: punctuation noise harm scales with its severity, and the human typos remain the dominant bottleneck, yielding the largest gaps to Clean accuracy and the highest ASR even with a large number of agents. And (2) Collaboration reliably improves accuracy as the number of agents, n, increases, with the largest gains from one to five agents and diminishing returns beyond 10 agents. However, the adversarial robustness gap persists regardless of the agent count.

Philip Bechtle, Lucie Flek, Philipp Alexander Jung et al.

In High Energy Physics, as in many other fields of science, the application of machine learning techniques has been crucial in advancing our understanding of fundamental phenomena. Increasingly, deep learning models are applied to analyze both simulated and experimental data. In most experiments, a rigorous regime of testing for physically motivated systematic uncertainties is in place. The numerical evaluation of these tests for differences between the data on the one side and simulations on the other side quantifies the effect of potential sources of mismodelling on the machine learning output. In addition, thorough comparisons of marginal distributions and (linear) feature correlations between data and simulation in "control regions" are applied. However, the guidance by physical motivation, and the need to constrain comparisons to specific regions, does not guarantee that all possible sources of deviations have been accounted for. We therefore propose a new adversarial attack - the CONSERVAttack - designed to exploit the remaining space of hypothetical deviations between simulation and data after the above mentioned tests. The resulting adversarial perturbations are consistent within the uncertainty bounds - evading standard validation checks - while successfully fooling the underlying model. We further propose strategies to mitigate such vulnerabilities and argue that robustness to adversarial effects must be considered when interpreting results from deep learning in particle physics.

Joan Plepi, Béla Neuendorf, Lucie Flek et al.

Instead of using a single ground truth for language processing tasks, several recent studies have examined how to represent and predict the labels of the set of annotators. However, often little or no information about annotators is known, or the set of annotators is small. In this work, we examine a corpus of social media posts about conflict from a set of 13k annotators and 210k judgements of social norms. We provide a novel experimental setup that applies personalization methods to the modeling of annotators and compare their effectiveness for predicting the perception of social norms. We further provide an analysis of performance across subsets of social situations that vary by the closeness of the relationship between parties in conflict, and assess where personalization helps the most.

Ramit Sawhney, Atula Tejaswi Neerkaje, Ivan Habernal et al.

Clinical NLP tasks such as mental health assessment from text, must take social constraints into account - the performance maximization must be constrained by the utmost importance of guaranteeing privacy of user data. Consumer protection regulations, such as GDPR, generally handle privacy by restricting data availability, such as requiring to limit user data to 'what is necessary' for a given purpose. In this work, we reason that providing stricter formal privacy guarantees, while increasing the volume of user data in the model, in most cases increases benefit for all parties involved, especially for the user. We demonstrate our arguments on two existing suicide risk assessment datasets of Twitter and Reddit posts. We present the first analysis juxtaposing user history length and differential privacy budgets and elaborate how modeling additional user context enables utility preservation while maintaining acceptable user privacy guarantees.

Allison Lahnala, Charles Welch, Béla Neuendorf et al.

Large pre-trained neural language models have supported the effectiveness of many NLP tasks, yet are still prone to generating toxic language hindering the safety of their use. Using empathetic data, we improve over recent work on controllable text generation that aims to reduce the toxicity of generated text. We find we are able to dramatically reduce the size of fine-tuning data to 7.5-30k samples while at the same time making significant improvements over state-of-the-art toxicity mitigation of up to 3.4% absolute reduction (26% relative) from the original work on 2.3m samples, by strategically sampling data based on empathy scores. We observe that the degree of improvement is subject to specific communication components of empathy. In particular, the cognitive components of empathy significantly beat the original dataset in almost all experiments, while emotional empathy was tied to less improvement and even underperforming random samples of the original data. This is a particularly implicative insight for NLP work concerning empathy as until recently the research and resources built for it have exclusively considered empathy as an emotional concept.

Shangrui Nie, Michael Fromm, Charles Welch et al.

While preliminary findings indicate that multilingual LLMs exhibit reduced bias compared to monolingual ones, a comprehensive understanding of the effect of multilingual training on bias mitigation, is lacking. This study addresses this gap by systematically training six LLMs of identical size (2.6B parameters) and architecture: five monolingual models (English, German, French, Italian, and Spanish) and one multilingual model trained on an equal distribution of data across these languages, all using publicly available data. To ensure robust evaluation, standard bias benchmarks were automatically translated into the five target languages and verified for both translation quality and bias preservation by human annotators. Our results consistently demonstrate that multilingual training effectively mitigates bias. Moreover, we observe that multilingual models achieve not only lower bias but also superior prediction accuracy when compared to monolingual models with the same amount of training data, model architecture, and size.

Flora Sakketou, Allison Lahnala, Liane Vogel et al.

There is an increasing need for the ability to model fine-grained opinion shifts of social media users, as concerns about the potential polarizing social effects increase. However, the lack of publicly available datasets that are suitable for the task presents a major challenge. In this paper, we introduce an innovative annotated dataset for modeling subtle opinion fluctuations and detecting fine-grained stances. The dataset includes a sufficient amount of stance polarity and intensity labels per user over time and within entire conversational threads, thus making subtle opinion fluctuations detectable both in long term and in short term. All posts are annotated by non-experts and a significant portion of the data is also annotated by experts. We provide a strategy for recruiting suitable non-experts. Our analysis of the inter-annotator agreements shows that the resulting annotations obtained from the majority vote of the non-experts are of comparable quality to the annotations of the experts. We provide analyses of the stance evolution in short term and long term levels, a comparison of language usage between users with vacillating and resolute attitudes, and fine-grained stance detection baselines.

Charles Welch, Joan Plepi, Béla Neuendorf et al.

Studies on interpersonal conflict have a long history and contain many suggestions for conflict typology. We use this as the basis of a novel annotation scheme and release a new dataset of situations and conflict aspect annotations. We then build a classifier to predict whether someone will perceive the actions of one individual as right or wrong in a given situation. Our analyses include conflict aspects, but also generated clusters, which are human validated, and show differences in conflict content based on the relationship of participants to the author. Our findings have important implications for understanding conflict and social norms.

Fabian Lechner, Allison Lahnala, Charles Welch et al.

The potential to provide patients with faster information access while allowing medical specialists to concentrate on critical tasks makes medical domain dialog agents appealing. However, the integration of large-language models (LLMs) into these agents presents certain limitations that may result in serious consequences. This paper investigates the challenges and risks of using GPT-3-based models for medical question-answering (MedQA). We perform several evaluations contextualized in terms of standard medical principles. We provide a procedure for manually designing patient queries to stress-test high-risk limitations of LLMs in MedQA systems. Our analysis reveals that LLMs fail to respond adequately to these queries, generating erroneous medical information, unsafe recommendations, and content that may be considered offensive.

Severino Trotta, Lucie Flek, Charles Welch

Recent language modeling performance has been greatly improved by the use of external memory. This memory encodes the context so that similar contexts can be recalled during decoding. This similarity depends on how the model learns to encode context, which can be altered to include other attributes, such as style. We construct and evaluate an architecture for this purpose, using corpora annotated for politeness, formality, and toxicity. Through extensive experiments and human evaluation we demonstrate the potential of our method to generate text while controlling style. We find that style-specific datastores improve generation performance, though results vary greatly across styles, and the effect of pretraining data and specific styles should be explored in future work.

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.

Akbar Karimi, Lucie Flek

The class imbalance problem can cause machine learning models to produce an undesirable performance on the minority class as well as the whole dataset. Using data augmentation techniques to increase the number of samples is one way to tackle this problem. We introduce a novel counterfactual data augmentation by verb replacement for the identification of medical claims. In addition, we investigate the impact of this method and compare it with 3 other data augmentation techniques, showing that the proposed method can result in a significant (relative) improvement in the minority class.

Heidi Campana Piva, Shaina Ashraf, Maziar Kianimoghadam Jouneghani et al.

Conspiracy theories are anti-authoritarian narratives that lead to social conflict, impacting how people perceive political information. To help in understanding this issue, we introduce the Conspiracy Frame: a fine-grained semantic representation of conspiratorial narratives derived from frame-semantics and semiotics, which spawned the Conspiracy Frames (Con.Fra.) dataset: a corpus of Telegram messages annotated at span-level. The Conspiracy Frame and Con.Fra. dataset contribute to the implementation of a more generalizable understanding and recognition of conspiracy theories. We observe the ability of LLMs to recognize this phenomenon in-domain and out-of-domain, investigating the role that frames may have in supporting this task. Results show that, while the injection of frames in an in-context approach does not lead to clear increase of performance, it has potential; the mapping of annotated spans with FrameNet shows abstract semantic patterns (e.g., `Kinship', `Ingest\_substance') that potentially pave the way for a more semantically- and semiotically-aware detection of conspiratorial narratives.

Mohammad H. A. Monfared, Lucie Flek, Akbar Karimi

We propose an agentic data augmentation method for Aspect-Based Sentiment Analysis (ABSA) that uses iterative generation and verification to produce high quality synthetic training examples. To isolate the effect of agentic structure, we also develop a closely matched prompting-based baseline using the same model and instructions. Both methods are evaluated across three ABSA subtasks (Aspect Term Extraction (ATE), Aspect Sentiment Classification (ATSC), and Aspect Sentiment Pair Extraction (ASPE)), four SemEval datasets, and two encoder-decoder models: T5-Base and Tk-Instruct. Our results show that the agentic augmentation outperforms raw prompting in label preservation of the augmented data, especially when the tasks require aspect term generation. In addition, when combined with real data, agentic augmentation provides higher gains, consistently outperforming prompting-based generation. These benefits are most pronounced for T5-Base, while the more heavily pretrained Tk-Instruct exhibits smaller improvements. As a result, augmented data helps T5-Base achieve comparable performance with its counterpart.

Shivam Rawat, Lucie Flek, Akbar Karimi

Scientific literature in astronomy is rapidly expanding, making it increasingly important to automate the extraction of key entities and contextual information from research papers. In this paper, we present an encoder-based system for extracting knowledge from astronomy articles. Our objective is to develop models capable of classifying telescope references, detecting auxiliary semantic attributes, and recognizing instrument mentions from textual content. To this end, we implement a multi-task transformer-based system built upon the SciBERT model and fine-tuned for astronomy corpora classification. To carry out the fine-tuning, we stochastically sample segments from the training data and use majority voting over the test segments at inference time. Our system, despite its simplicity and low-cost implementation, significantly outperforms the open-weight GPT baseline.

Gilles Nawezi, Lucie Flek, Charles Welch

Recent language models have been improved by the addition of external memory. Nearest neighbor language models retrieve similar contexts to assist in word prediction. The addition of locality levels allows a model to learn how to weight neighbors based on their relative location to the current text in source documents, and have been shown to further improve model performance. Nearest neighbor models have been explored for controllable generation but have not examined the use of locality levels. We present a novel approach for this purpose and evaluate it using automatic and human evaluation on politeness, formality, supportiveness, and toxicity textual data. We find that our model is successfully able to control style and provides a better fluency-style trade-off than previous work.

Shivam Rawat, Lucie Flek, Florian Mai et al.

Reasoning in large language models is often treated as a monolithic capability, but its observed gains may arise from more basic operations. We study reasoning through two such primitives, recall and state-tracking, and ask whether hybrid architectures that combine attention-based retrieval with recurrent state updates are better suited than attention-only models for tasks that jointly require both. Using matched Olmo3 transformer and hybrid models in instruction-tuned and reasoning-augmented variants, we evaluate these models on a set of controlled tasks involving a mixture of state-tracking and recall primitives, state-based recall. Across tasks, we notice that reasoning augmentation provides the largest overall improvement, substantially extending the range of difficulty over which models remain effective. We also notice that in certain tasks, the hybrid reasoning model remains substantially more robust as sequential dependence increases. In contrast, the transformer reasoning model degrades sharply in performance as task difficulty increases beyond a given threshold. These results suggest that reasoning tokens and architectural inductive biases contribute at different levels of the computational process: explicit reasoning can expand a model's effective operating range, but its benefit depends on how well the underlying architecture supports persistent state propagation. Given the small size of our case study, which involves a limited set of models and tasks, we present these findings as suggestive rather than conclusive and leave broader validation across model families, scales, and task variations to future work.

Charles Welch, Allison Lahnala, Vasudha Varadarajan et al.

We present a low-compute non-generative system for implementing interview-style conversational agents which can be used to facilitate qualitative data collection through controlled interactions and quantitative analysis. Use cases include applications to tracking attitude formation or behavior change, where control or standardization over the conversational flow is desired. We show how our system can be easily adjusted through an online administrative panel to create new interviews, making the tool accessible without coding. Two case studies are presented as example applications, one regarding the Expressive Interviewing system for COVID-19 and the other a semi-structured interview to survey public opinion on emerging neurotechnology. Our code is open-source, allowing others to build off of our work and develop extensions for additional functionality.

Mehdi Ali, Manuel Brack, Max Lübbering et al.

High-quality multilingual training data is essential for effectively pretraining large language models (LLMs). Yet, the availability of suitable open-source multilingual datasets remains limited. Existing state-of-the-art datasets mostly rely on heuristic filtering methods, restricting both their cross-lingual transferability and scalability. Here, we introduce JQL, a systematic approach that efficiently curates diverse and high-quality multilingual data at scale while significantly reducing computational demands. JQL distills LLMs' annotation capabilities into lightweight annotators based on pretrained multilingual embeddings. These models exhibit robust multilingual and cross-lingual performance, even for languages and scripts unseen during training. Evaluated empirically across 35 languages, the resulting annotation pipeline substantially outperforms current heuristic filtering methods like Fineweb2. JQL notably enhances downstream model training quality and increases data retention rates. Our research provides practical insights and valuable resources for multilingual data curation, raising the standards of multilingual dataset development.

Shangrui Nie, Florian Mai, David Kaczér et al.

Large language models implicitly encode preferences over human values, yet steering them often requires large training data. In this work, we investigate a simple approach: Can we reliably modify a model's value system in downstream behavior by training it to answer value survey questions accordingly? We first construct value profiles of several open-source LLMs by asking them to rate a series of value-related descriptions spanning 20 distinct human values, which we use as a baseline for subsequent experiments. We then investigate whether the value system of a model can be governed by fine-tuning on the value surveys. We evaluate the effect of finetuning on the model's behavior in two ways; first, we assess how answers change on in-domain, held-out survey questions. Second, we evaluate whether the model's behavior changes in out-of-domain settings (situational scenarios). To this end, we construct a contextualized moral judgment dataset based on Reddit posts and evaluate changes in the model's behavior in text-based adventure games. We demonstrate that our simple approach can not only change the model's answers to in-domain survey questions, but also produces substantial shifts (value alignment) in implicit downstream task behavior.

Ondrej Sotolar, Vojtech Formanek, Alok Debnath et al.

Empathetic response generation is a desirable aspect of conversational agents, crucial for facilitating engaging and emotionally intelligent multi-turn conversations between humans and machines. Leveraging large language models for this task has shown promising results, yet challenges persist in ensuring both the empathetic quality of the responses and retention of the generalization performance of the models. We propose a novel approach where we construct theory-driven preference datasets based on emotion grounding and use them to align LLMs with preference optimization algorithms to address these challenges. To evaluate empathetic response generation, we employ the EmpatheticDialogues dataset, assessing empathy with the diff-Epitome and BERTscore metrics and with multi-dimensional human evaluation. Additionally, we measure diversity and emotional valence using feature-based methods. We also evaluate the impact of training on the generalization performance using the MMLU benchmark and tasks from the Open LLM Leaderboard. The results show that LLMs can be aligned for empathetic response generation by preference optimization while retaining their general performance and that emotion grounding can guide preference dataset creation. We make all datasets, source code, and models publicly available. https://github.com/justtherightsize/empo

Mounika Marreddy, Subba Reddy Oota, Venkata Charan Chinni et al.

Analyzing user opinion changes in long conversation threads is extremely critical for applications like enhanced personalization, market research, political campaigns, customer service, targeted advertising, and content moderation. Unfortunately, previous studies on stance and dogmatism in user conversations have focused on training models using datasets annotated at the post level, treating each post as independent and randomly sampling posts from conversation threads. Hence, first, we build a dataset for studying user opinion fluctuations in 764 long multi-user Reddit conversation threads, called USDC. USDC contains annotations for 2 tasks: i) User Stance classification, which involves labeling a user's stance in a post within a conversation on a five-point scale; ii) User Dogmatism classification, which involves labeling a user's overall opinion in the conversation on a four-point scale. Besides being time-consuming and costly, manual annotations for USDC are challenging because: 1) Conversation threads could be very long, increasing the chances of noisy annotations; and 2) Interpreting instances where a user changes their opinion within a conversation is difficult because often such transitions are subtle and not expressed explicitly. Hence, we leverage majority voting on zero-shot, one-shot, and few-shot annotations from Mistral Large and GPT-4 to automate the annotation process. Human annotations on 200 test conversations achieved inter-annotator agreement scores of 0.49 for stance and 0.50 for dogmatism with these LLM annotations, indicating a reasonable level of consistency between human and LLM annotations. USDC is then used to finetune and instruction-tune multiple deployable small language models like LLaMA, Falcon and Vicuna for the stance and dogmatism classification tasks. We make the code and dataset publicly available [https://github.com/mounikamarreddy/USDC].

Shivam Rawat, Lucie Flek

Agentic AI systems are increasingly being integrated into scientific workflows, yet their behavior under realistic conditions remains insufficiently understood. We evaluate CMBAgent across two workflow paradigms and eighteen astrophysical tasks. In the One-Shot setting, access to domain-specific context yields an approximately ~6x performance improvement (0.85 vs. ~0 without context), with the primary failure mode being silent incorrect computation - syntactically valid code that produces plausible but inaccurate results. In the Deep Research setting, the system frequently exhibits silent failures across stress tests, producing physically inconsistent posteriors without self-diagnosis. Overall, performance is strong on well-specified tasks but degrades on problems designed to probe reasoning limits, often without visible error signals. These findings highlight that the most concerning failure mode in agentic scientific workflows is not overt failure, but confident generation of incorrect results. We release our evaluation framework to facilitate systematic reliability analysis of scientific AI agents.

Matthias Schott, Lucie Flek

Machine-learning models in high-energy physics are often trained on simulated data, where fully simulated samples are computationally expensive while fast simulation provides large statistics at reduced realism. In this work, we systematically study transfer learning between fast-simulated and fully simulated datasets in a realistic LHC environment. We consider three representative tasks, signal-background classification, quark-gluon jet tagging, and missing transverse energy reconstruction, using dense neural networks, graph neural networks, and transformer-based architectures. Models are pretrained on ATLAS-like fast simulation and adapted to CMS-like fast simulation and to fully simulated ATLAS Open Data. Across all tasks, pretrained models consistently outperform independently trained baselines and require significantly less target-domain training data, typically reducing the needed statistics by about a factor of two. These results demonstrate that fast simulation can be used to learn robust, reusable representations and motivate publishing trained models as reusable scientific assets beyond large foundation models.

Matthias Schott, Lucie Flek

Accurate Monte Carlo (MC) modelling in high-energy physics is challenging, particularly in complex scenarios where simulations fail to reproduce observed data. In practice, experimental information is often limited to one-dimensional (1D) distributions, while mismodelling arises in a multidimensional feature space. This restricts traditional correction methods, as one-dimensional reweighting ignores correlations and fully multidimensional approaches require large target datasets. We propose a neural network-based method that operates under these constraints by learning a transformation of simulated events that reproduces the available 1D target distributions while remaining close to the original simulation. This minimal-deviation principle preserves the global correlation structure of the baseline model while enabling targeted corrections of mismodelled features. Using controlled studies with simulated pseudo-data, we show that the method improves agreement with target distributions and maintains a consistent multidimensional structure. The approach is designed for complex, high-dimensional analyses where traditional techniques are insufficient, providing a scalable way to enhance MC modelling under limited information.

Lucie Flek, Philipp Alexander Jungs, Akbar Karimi et al.

Neural networks (NNs) are inherently multidimensional classifiers that learn complex, non-linear relationships among input observables. While their flexibility enables unprecedented performance in high-energy physics (HEP) analyses, it also makes them sensitive to small variations in their inputs. Consequently, the propagation and estimation of systematic uncertainties in NN-based models remain an open challenge. There are indications that uncertainties derived in control regions or from nominal variations of input features can underestimate the true model uncertainty, potentially leaving biases unaccounted for. Inspired by insights from adversarial-attack studies in machine learning, we explore how subtle perturbations, fully consistent with the experimental uncertainties on the input observables, can lead to substantial changes in NN outputs, while keeping the one-dimensional and correlated input distributions nearly unchanged. Using a set of representative HEP tasks, including event classification and object identification, and testing across a variety of network architectures, we demonstrate that networks can be systematically "fooled" at significant rates within the allowed uncertainty envelopes. Building on this observation, we introduce a quantitative framework to probe and measure the hidden sensitivity of neural networks to realistic experimental variations, providing a practical path to evaluate and control their systematic uncertainty in physics analyses.

Shiza Fatimah, Aniket Sen, Sophia Falk et al.

The dominance of large multilingual foundation models has widened linguistic inequalities in Natural Language Processing (NLP), often leaving low-resource languages underrepresented. This paper introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address this gap. Unlike prior Hindi models that rely on continual pretraining from opaque multilingual foundations, LilMoo is developed through a fully transparent and reproducible pipeline optimized for limited compute environments. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned (LLM-as-a-judge) methods, complemented by bilingual augmentation with curated English data. Using this dataset, we explore various training recipes for small-scale language models. Across comprehensive evaluation suites, LilMoo consistently outperforms comparably sized multilingual baselines such as Qwen2.5-0.5B and Qwen3-0.6B, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.

Christian Nickel, Laura Schrewe, Florian Mai et al.

Theory of Mind (ToM) refers to an agent's ability to model the internal states of others. Contributing to the debate whether large language models (LLMs) exhibit genuine ToM capabilities, our study investigates their ToM robustness using perturbations on false-belief tasks and examines the potential of Chain-of-Thought prompting (CoT) to enhance performance and explain the LLM's decision. We introduce a handcrafted, richly annotated ToM dataset, including classic and perturbed false belief tasks, the corresponding spaces of valid reasoning chains for correct task completion, subsequent reasoning faithfulness, task solutions, and propose metrics to evaluate reasoning chain correctness and to what extent final answers are faithful to reasoning traces of the generated CoT. We show a steep drop in ToM capabilities under task perturbation for all evaluated LLMs, questioning the notion of any robust form of ToM being present. While CoT prompting improves the ToM performance overall in a faithful manner, it surprisingly degrades accuracy for some perturbation classes, indicating that selective application is necessary.

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.

Zain Ul Abedin, Shahzeb Qamar, Lucie Flek et al.

While Large Language Models (LLMs) have shown impressive capabilities in math problem-solving tasks, their robustness to noisy inputs is not well-studied. We propose ArithmAttack to examine how robust the LLMs are when they encounter noisy prompts that contain extra noise in the form of punctuation marks. While being easy to implement, ArithmAttack does not cause any information loss since words are not added or deleted from the context. We evaluate the robustness of eight LLMs, including LLama3, Mistral, Mathstral, and DeepSeek on noisy GSM8K and MultiArith datasets. Our experiments suggest that all the studied models show vulnerability to such noise, with more noise leading to poorer performances.

Ipek Baris Schlicht, Defne Altiok, Maryanne Taouk et al.

This paper addresses debiasing in news editing and evaluates the effectiveness of conversational Large Language Models in this task. We designed an evaluation checklist tailored to news editors' perspectives, obtained generated texts from three popular conversational models using a subset of a publicly available dataset in media bias, and evaluated the texts according to the designed checklist. Furthermore, we examined the models as evaluator for checking the quality of debiased model outputs. Our findings indicate that none of the LLMs are perfect in debiasing. Notably, some models, including ChatGPT, introduced unnecessary changes that may impact the author's style and create misinformation. Lastly, we show that the models do not perform as proficiently as domain experts in evaluating the quality of debiased outputs.

Olufunke O. Sarumi, Béla Neuendorf, Joan Plepi et al.

Recent trends in natural language processing research and annotation tasks affirm a paradigm shift from the traditional reliance on a single ground truth to a focus on individual perspectives, particularly in subjective tasks. In scenarios where annotation tasks are meant to encompass diversity, models that solely rely on the majority class labels may inadvertently disregard valuable minority perspectives. This oversight could result in the omission of crucial information and, in a broader context, risk disrupting the balance within larger ecosystems. As the landscape of annotator modeling unfolds with diverse representation techniques, it becomes imperative to investigate their effectiveness with the fine-grained features of the datasets in view. This study systematically explores various annotator modeling techniques and compares their performance across seven corpora. From our findings, we show that the commonly used user token model consistently outperforms more complex models. We introduce a composite embedding approach and show distinct differences in which model performs best as a function of the agreement with a given dataset. Our findings shed light on the relationship between corpus statistics and annotator modeling performance, which informs future work on corpus construction and perspectivist NLP.

Timo Saala, Lucie Flek, Alexander Jung et al.

Correlations between input parameters play a crucial role in many scientific classification tasks, since these are often related to fundamental laws of nature. For example, in high energy physics, one of the common deep learning use-cases is the classification of signal and background processes in particle collisions. In many such cases, the fundamental principles of the correlations between observables are often better understood than the actual distributions of the observables themselves. In this work, we present a new adversarial attack algorithm called Random Distribution Shuffle Attack (RDSA), emphasizing the correlations between observables in the network rather than individual feature characteristics. Correct application of the proposed novel attack can result in a significant improvement in classification performance - particularly in the context of data augmentation - when using the generated adversaries within adversarial training. Given that correlations between input features are also crucial in many other disciplines. We demonstrate the RDSA effectiveness on six classification tasks, including two particle collision challenges (using CERN Open Data), hand-written digit recognition (MNIST784), human activity recognition (HAR), weather forecasting (Rain in Australia), and ICU patient mortality (MIMIC-IV), demonstrating a general use case beyond fundamental physics for this new type of adversarial attack algorithms.

Allison Lahnala, Charles Welch, David Jurgens et al.

Conceptual operationalizations of empathy in NLP are varied, with some having specific behaviors and properties, while others are more abstract. How these variations relate to one another and capture properties of empathy observable in text remains unclear. To provide insight into this, we analyze the transfer performance of empathy models adapted to empathy tasks with different theoretical groundings. We study (1) the dimensionality of empathy definitions, (2) the correspondence between the defined dimensions and measured/observed properties, and (3) the conduciveness of the data to represent them, finding they have a significant impact to performance compared to other transfer setting features. Characterizing the theoretical grounding of empathy tasks as direct, abstract, or adjacent further indicates that tasks that directly predict specified empathy components have higher transferability. Our work provides empirical evidence for the need for precise and multidimensional empathy operationalizations.

Amirhossein Aliakbarzadeh, Lucie Flek, Akbar Karimi

Large Language Models (LLMs) are trained on Web data that might contain spelling errors made by humans. But do they become robust to similar real-world noise? In this paper, we investigate the effect of real-world spelling mistakes on the performance of 9 language models, with parameters ranging from 0.2B to 13B, in 3 different NLP tasks, namely Natural Language Inference (NLI), Name Entity Recognition (NER), and Intent Classification (IC). We perform our experiments on 6 different languages and build a dictionary of real-world noise for them using the Wikipedia edit history. We show that the performance gap of the studied models on the clean and noisy test data averaged across all the datasets and languages ranges from 2.3 to 4.3 absolute percentage points. In addition, mT5 models, in general, show more robustness compared to BLOOM, Falcon, and BERT-like models. In particular, mT5 (13B), was the most robust on average overall, across the 3 tasks, and in 4 of the 6 languages.

Allison Lahnala, Vasudha Varadarajan, Lucie Flek et al.

The proliferation of ideological movements into extremist factions via social media has become a global concern. While radicalization has been studied extensively within the context of specific ideologies, our ability to accurately characterize extremism in more generalizable terms remains underdeveloped. In this paper, we propose a novel method for extracting and analyzing extremist discourse across a range of online community forums. By focusing on verbal behavioral signatures of extremist traits, we develop a framework for quantifying extremism at both user and community levels. Our research identifies 11 distinct factors, which we term ``The Extremist Eleven,'' as a generalized psychosocial model of extremism. Applying our method to various online communities, we demonstrate an ability to characterize ideologically diverse communities across the 11 extremist traits. We demonstrate the power of this method by analyzing user histories from members of the incel community. We find that our framework accurately predicts which users join the incel community up to 10 months before their actual entry with an AUC of $>0.6$, steadily increasing to AUC ~0.9 three to four months before the event. Further, we find that upon entry into an extremist forum, the users tend to maintain their level of extremism within the community, while still remaining distinguishable from the general online discourse. Our findings contribute to the study of extremism by introducing a more holistic, cross-ideological approach that transcends traditional, trait-specific models.

Vahid Sadiri Javadi, Johanne R. Trippas, Yash Kumar Lal et al.

Narratives are widely recognized as a powerful tool for structuring information and facilitating comprehension of complex ideas in various domains such as science communication. This paper investigates whether incorporating narrative elements can assist Large Language Models (LLMs) in solving complex problems more effectively. We propose a novel approach, Story of Thought (SoT), integrating narrative structures into prompting techniques for problem-solving. This approach involves constructing narratives around problem statements and creating a framework to identify and organize relevant information. Our experiments show that using various LLMs with SoT consistently surpasses using them with other techniques on physics, chemistry, math, and biology questions in both the GPQA and JEEBench datasets. The narrative-based information curation process in SoT enhances problem comprehension by contextualizing critical in-domain information and highlighting causal relationships within the problem space.

Florian Mai, David Kaczér, Nicholas Kluge Corrêa et al.

Two core challenges of alignment are 1) scalable oversight and 2) accounting for the dynamic nature of human values. While solutions like recursive reward modeling address 1), they do not simultaneously account for 2). We sketch a roadmap for a novel algorithmic framework that trains a superhuman reasoning model to decompose complex tasks into subtasks that are still amenable to human-level guidance. Our approach relies on what we call the part-to-complete generalization hypothesis, which states that the alignment of subtask solutions generalizes to the alignment of complete solutions. We advocate for the need to measure this generalization and propose ways to improve it in the future.

Olufunke O. Sarumi, Charles Welch, Lucie Flek et al.

In this work, we evaluate annotator disagreement in Word-in-Context (WiC) tasks exploring the relationship between contextual meaning and disagreement as part of the CoMeDi shared task competition. While prior studies have modeled disagreement by analyzing annotator attributes with single-sentence inputs, this shared task incorporates WiC to bridge the gap between sentence-level semantic representation and annotator judgment variability. We describe three different methods that we developed for the shared task, including a feature enrichment approach that combines concatenation, element-wise differences, products, and cosine similarity, Euclidean and Manhattan distances to extend contextual embedding representations, a transformation by Adapter blocks to obtain task-specific representations of contextual embeddings, and classifiers of varying complexities, including ensembles. The comparison of our methods demonstrates improved performance for methods that include enriched and task-specfic features. While the performance of our method falls short in comparison to the best system in subtask 1 (OGWiC), it is competitive to the official evaluation results in subtask 2 (DisWiC).

Tobias Bystrich, Lukas Hamm, Maria Hassan et al.

Due to the scarcity of labeled dialectal speech, audio dialect classification is a challenging task for most languages, including Swiss German. In this work, we explore the ability of large language models (LLMs) as agents in understanding the dialects and whether they can show comparable performance to models such as HuBERT in dialect classification. In addition, we provide an LLM baseline and a human linguist one. Our approach uses phonetic transcriptions produced by ASR systems and combines them with linguistic resources such as dialect feature maps, vowel history, and rules. Our findings indicate that, when linguistic information is provided, the LLM predictions improve. The human baseline shows that automatically generated transcriptions can be beneficial for such classifications, but also presents opportunities for improvement.

Tianyi Zhang, Florian Mai, Lucie Flek

Continual pretraining promises to adapt large language models (LLMs) to new domains using only unlabeled test-time data, but naively applying standard self-supervised objectives to instruction-tuned models is known to degrade their instruction-following capability and semantic representations. Existing fixes assume access to the original base model or rely on knowledge from an external domain-specific database - both of which pose a realistic barrier in settings where the base model weights are withheld for safety reasons or reliable external corpora are unavailable. In this work, we propose Instruction-Knowledge-Aware Continual Adaptation (IKnow), a simple and general framework that formulates novel self-supervised objectives in the instruction-response dialogue format. Rather than depend- ing on external resources, IKnow leverages domain knowledge embedded within the text itself and learns to encode it at a deeper semantic level.