Katharina von der Wense

Mario Sanz-Guerrero, Manuel Mager, Katharina von der Wense

Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template's effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an "ownership bias" -- models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.

Carolin Holtermann, Minh Duc Bui, Kaitlyn Zhou et al.

Hundreds of millions of people rely on large language models (LLMs) for education, work, and even healthcare. Yet these models are known to reproduce and amplify social biases present in their training data. Moreover, text-based interfaces remain a barrier for many, for example, users with limited literacy, motor impairments, or mobile-only devices. Voice interaction promises to expand accessibility, but unlike text, speech carries identity cues that users cannot easily mask, raising concerns about whether accessibility gains may come at the cost of equitable treatment. Here we show that audio-enabled LLMs exhibit systematic gender discrimination, shifting responses toward gender-stereotyped adjectives and occupations solely on the basis of speaker voice, and amplifying bias beyond that observed in text-based interaction. Thus, voice interfaces do not merely extend text models to a new modality but introduce distinct bias mechanisms tied to paralinguistic cues. Complementary survey evidence ($n=1,000$) shows that infrequent chatbot users are most hesitant to undisclosed attribute inference and most likely to disengage when such practices are revealed. To demonstrate a potential mitigation strategy, we show that pitch manipulation can systematically regulate gender-discriminatory outputs. Overall, our findings reveal a critical tension in AI development: efforts to expand accessibility through voice interfaces simultaneously create new pathways for discrimination, demanding that fairness and accessibility be addressed in tandem.

Sagi Shaier, Lawrence E. Hunter, Katharina von der Wense

Both standalone language models (LMs) as well as LMs within downstream-task systems have been shown to generate statements which are factually untrue. This problem is especially severe for low-resource languages, where training data is scarce and of worse quality than for high-resource languages. In this opinion piece, we argue that LMs in their current state will never be fully trustworthy in critical settings and suggest a possible novel strategy to handle this issue: by building LMs such that can cite their sources - i.e., point a user to the parts of their training data that back up their outputs. We first discuss which current NLP tasks would or would not benefit from such models. We then highlight the expected benefits such models would bring, e.g., quick verifiability of statements. We end by outlining the individual tasks that would need to be solved on the way to developing LMs with the ability to cite. We hope to start a discussion about the field's current approach to building LMs, especially for low-resource languages, and the role of the training data in explaining model generations.

Minh Duc Bui, Katharina von der Wense, Anne Lauscher

Warning: this paper contains content that may be offensive or upsetting Hate speech moderation on global platforms poses unique challenges due to the multimodal and multilingual nature of content, along with the varying cultural perceptions. How well do current vision-language models (VLMs) navigate these nuances? To investigate this, we create the first multimodal and multilingual parallel hate speech dataset, annotated by a multicultural set of annotators, called Multi3Hate. It contains 300 parallel meme samples across 5 languages: English, German, Spanish, Hindi, and Mandarin. We demonstrate that cultural background significantly affects multimodal hate speech annotation in our dataset. The average pairwise agreement among countries is just 74%, significantly lower than that of randomly selected annotator groups. Our qualitative analysis indicates that the lowest pairwise label agreement-only 67% between the USA and India-can be attributed to cultural factors. We then conduct experiments with 5 large VLMs in a zero-shot setting, finding that these models align more closely with annotations from the US than with those from other cultures, even when the memes and prompts are presented in the dominant language of the other culture. Code and dataset are available at https://github.com/MinhDucBui/Multi3Hate.

Sagi Shaier, Mario Sanz-Guerrero, Katharina von der Wense

This study investigates whether repeating questions within prompts influences the performance of large language models (LLMs). We hypothesize that reiterating a question within a single prompt might enhance the model's focus on key elements of the query. We evaluate five recent LLMs -- including GPT-4o-mini, DeepSeek-V3, and smaller open-source models -- on three reading comprehension datasets under different prompt settings, varying question repetition levels (1, 3, or 5 times per prompt). Our results demonstrate that question repetition can increase models' accuracy by up to $6\%$. However, across all models, settings, and datasets, we do not find the result statistically significant. These findings provide insights into prompt design and LLM behavior, suggesting that repetition alone does not significantly impact output quality.

Sagi Shaier, Kevin Bennett, Lawrence Hunter et al.

State-of-the-art question answering (QA) models exhibit a variety of social biases (e.g., with respect to sex or race), generally explained by similar issues in their training data. However, what has been overlooked so far is that in the critical domain of biomedicine, any unjustified change in model output due to patient demographics is problematic: it results in the unfair treatment of patients. Selecting only questions on biomedical topics whose answers do not depend on ethnicity, sex, or sexual orientation, we ask the following research questions: (RQ1) Do the answers of QA models change when being provided with irrelevant demographic information? (RQ2) Does the answer of RQ1 differ between knowledge graph (KG)-grounded and text-based QA systems? We find that irrelevant demographic information change up to 15% of the answers of a KG-grounded system and up to 23% of the answers of a text-based system, including changes that affect accuracy. We conclude that unjustified answer changes caused by patient demographics are a frequent phenomenon, which raises fairness concerns and should be paid more attention to.

Minh Duc Bui, Manuel Mager, Peter Herbert Kann et al.

Meenzerisch, the dialect spoken in the German city of Mainz, is also the traditional language of the Mainz carnival, a yearly celebration well known throughout Germany. However, Meenzerisch is on the verge of dying out-a fate it shares with many other German dialects. Natural language processing (NLP) has the potential to help with the preservation and revival efforts of languages and dialects. However, so far no NLP research has looked at Meenzerisch. This work presents the first research in the field of NLP that is explicitly focused on the dialect of Mainz. We introduce a digital dictionary-an NLP-ready dataset derived from an existing resource (Schramm, 1966)-to support researchers in modeling and benchmarking the language. It contains 2,351 words in the dialect paired with their meanings described in Standard German. We then use this dataset to answer the following research questions: (1) Can state-of-the-art large language models (LLMs) generate definitions for dialect words? (2) Can LLMs generate words in Meenzerisch, given their definitions? Our experiments show that LLMs can do neither: the best model for definitions reaches only 6.27% accuracy and the best word generation model's accuracy is 1.51%. We then conduct two additional experiments in order to see if accuracy is improved by few-shot learning and by extracting rules from the training set, which are then passed to the LLM. While those approaches are able to improve the results, accuracy remains below 10%. This highlights that additional resources and an intensification of research efforts focused on German dialects are desperately needed.

Minh Duc Bui, Kyung Eun Park, Goran Glavaš et al.

Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.

Minh Duc Bui, Xenia Heilmann, Mattia Cerrato et al.

Prior work evaluates code generation bias primarily through simple conditional statements, which represent only a narrow slice of real-world programming and reveal solely overt, explicitly encoded bias. We demonstrate that this approach dramatically underestimates bias in practice by examining a more realistic task: generating machine learning (ML) pipelines. Testing both code-specialized and general-instruction large language models, we find that generated pipelines exhibit significant bias during feature selection. Sensitive attributes appear in 87.7% of cases on average, despite models demonstrably excluding irrelevant features (e.g., including "race" while dropping "favorite color" for credit scoring). This bias is substantially more prevalent than that captured by conditional statements, where sensitive attributes appear in only 59.2% of cases. These findings are robust across prompt mitigation strategies, varying numbers of attributes, and different pipeline difficulty levels. Our results challenge simple conditionals as valid proxies for bias evaluation and suggest current benchmarks underestimate bias risk in practical deployments.

Sagi Shaier, Kevin Bennett, Lawrence E Hunter et al.

The differences between cloze-task language model (LM) probing with 1) expert-made templates and 2) naturally-occurring text have often been overlooked. Here, we evaluate 16 different LMs on 10 probing English datasets -- 4 template-based and 6 template-free -- in general and biomedical domains to answer the following research questions: (RQ1) Do model rankings differ between the two approaches? (RQ2) Do models' absolute scores differ between the two approaches? (RQ3) Do the answers to RQ1 and RQ2 differ between general and domain-specific models? Our findings are: 1) Template-free and template-based approaches often rank models differently, except for the top domain-specific models. 2) Scores decrease by up to 42% Acc@1 when comparing parallel template-free and template-based prompts. 3) Perplexity is negatively correlated with accuracy in the template-free approach, but, counter-intuitively, they are positively correlated for template-based probing. 4) Models tend to predict the same answers frequently across prompts for template-based probing, which is less common when employing template-free techniques.

Sagi Shaier, Lawrence E Hunter, Katharina von der Wense

Prior work has uncovered a set of common problems in state-of-the-art context-based question answering (QA) systems: a lack of attention to the context when the latter conflicts with a model's parametric knowledge, little robustness to noise, and a lack of consistency with their answers. However, most prior work focus on one or two of those problems in isolation, which makes it difficult to see trends across them. We aim to close this gap, by first outlining a set of -- previously discussed as well as novel -- desiderata for QA models. We then survey relevant analysis and methods papers to provide an overview of the state of the field. The second part of our work presents experiments where we evaluate 15 QA systems on 5 datasets according to all desiderata at once. We find many novel trends, including (1) systems that are less susceptible to noise are not necessarily more consistent with their answers when given irrelevant context; (2) most systems that are more susceptible to noise are more likely to correctly answer according to a context that conflicts with their parametric knowledge; and (3) the combination of conflicting knowledge and noise can reduce system performance by up to 96%. As such, our desiderata help increase our understanding of how these models work and reveal potential avenues for improvements.

George Arthur Baker, Ankush Raut, Sagi Shaier et al.

Previous work finds that recent long-context language models fail to make equal use of information in the middle of their inputs, preferring pieces of information located at the tail ends which creates an undue bias in situations where we would like models to be equally capable of using different parts of the input. Thus far, the problem has mainly only been considered in settings with single pieces of critical information, leading us to question what happens when multiple necessary pieces of information are spread out over the inputs. Here, we demonstrate the effects of the "lost in the middle" problem in the multi-hop question answering setting -- in which multiple reasoning "hops" over disconnected documents are required -- and show that performance degrades not only with respect to the distance of information from the edges of the context, but also between pieces of information. Additionally, we experiment with means of alleviating the problem by reducing superfluous document contents through knowledge graph triple extraction and summarization, and prompting models to reason more thoroughly using chain-of-thought prompting.

Enora Rice, Ali Marashian, Hannah Haynie et al.

Cross-lingual transfer learning is an invaluable tool for overcoming data scarcity, yet selecting a suitable transfer language remains a challenge. The precise roles of linguistic typology, training data, and model architecture in transfer language choice are not fully understood. We take a holistic approach, examining how both dataset-specific and fine-grained typological features influence transfer language selection for part-of-speech tagging, considering two different sources for morphosyntactic features. While previous work examines these dynamics in the context of bilingual biLSTMS, we extend our analysis to a more modern transfer learning pipeline: zero-shot prediction with pretrained multilingual models. We train a series of transfer language ranking systems and examine how different feature inputs influence ranker performance across architectures. Word overlap, type-token ratio, and genealogical distance emerge as top features across all architectures. Our findings reveal that a combination of typological and dataset-dependent features leads to the best rankings, and that good performance can be obtained with either feature group on its own.

Ali Marashian, Enora Rice, Luke Gessler et al.

Many of the world's languages have insufficient data to train high-performing general neural machine translation (NMT) models, let alone domain-specific models, and often the only available parallel data are small amounts of religious texts. Hence, domain adaptation (DA) is a crucial issue faced by contemporary NMT and has, so far, been underexplored for low-resource languages. In this paper, we evaluate a set of methods from both low-resource NMT and DA in a realistic setting, in which we aim to translate between a high-resource and a low-resource language with access to only: a) parallel Bible data, b) a bilingual dictionary, and c) a monolingual target-domain corpus in the high-resource language. Our results show that the effectiveness of the tested methods varies, with the simplest one, DALI, being most effective. We follow up with a small human evaluation of DALI, which shows that there is still a need for more careful investigation of how to accomplish DA for low-resource NMT.

Adam Wiemerslage, Katharina von der Wense

Self-supervised objectives have driven major advances in NLP by leveraging large-scale unlabeled data, but such resources are scarce for many of the world's languages. Surprisingly, they have not been explored much for character-level tasks, where smaller amounts of data have the potential to be beneficial. We investigate the effectiveness of self-supervised auxiliary tasks for morphological inflection -- a character-level task highly relevant for language documentation -- in extremely low-resource settings, training encoder-decoder transformers for 19 languages and 13 auxiliary objectives. Autoencoding yields the best performance when unlabeled data is very limited, while character masked language modeling (CMLM) becomes more effective as data availability increases. Though objectives with stronger inductive biases influence model predictions intuitively, they rarely outperform standard CMLM. However, sampling masks based on known morpheme boundaries consistently improves performance, highlighting a promising direction for low-resource morphological modeling.

Stéphane Aroca-Ouellette, Miguel Aroca-Ouellette, Katharina von der Wense et al.

In collaborative tasks, autonomous agents fall short of humans in their capability to quickly adapt to new and unfamiliar teammates. We posit that a limiting factor for zero-shot coordination is the lack of shared task abstractions, a mechanism humans rely on to implicitly align with teammates. To address this gap, we introduce HA$^2$: Hierarchical Ad Hoc Agents, a framework leveraging hierarchical reinforcement learning to mimic the structured approach humans use in collaboration. We evaluate HA$^2$ in the Overcooked environment, demonstrating statistically significant improvement over existing baselines when paired with both unseen agents and humans, providing better resilience to environmental shifts, and outperforming all state-of-the-art methods.

Aaron Gluck, Katharina von der Wense, Maria Leonor Pacheco

Automated definition generation systems have been proposed to support vocabulary expansion for language learners. The main barrier to the success of these systems is that learners often struggle to understand definitions due to the presence of potentially unfamiliar words and grammar, particularly when non-standard language is involved. To address these challenges, we propose CLIX, the task of Cross-Lingual explanations of Idiomatic eXpressions. We explore the capabilities of current NLP models for this task, and observe that while it remains challenging, large language models show promise. Finally, we perform a detailed error analysis to highlight the key challenges that need to be addressed before we can reliably incorporate these systems into educational tools.

Minh Duc Bui, Fabian David Schmidt, Goran Glavaš et al.

Compared to standard language model (LM) pretraining (i.e., from scratch), Knowledge Distillation (KD) entails an additional forward pass through a teacher model that is typically substantially larger than the target student model. As such, KD in LM pretraining materially slows down throughput of pretraining instances vis-a-vis pretraining from scratch. Scaling laws of LM pretraining suggest that smaller models can close the gap to larger counterparts if trained on more data (i.e., processing more tokens)-and under a fixed computation budget, smaller models are able be process more data than larger models. We thus hypothesize that KD might, in fact, be suboptimal to pretraining from scratch for obtaining smaller LMs, when appropriately accounting for the compute budget. To test this, we compare pretraining from scratch against several KD strategies for masked language modeling (MLM) in a fair experimental setup, with respect to amount of computation as well as pretraining data. Downstream results on GLUE, however, do not confirm our hypothesis: while pretraining from scratch performs comparably to ordinary KD under a fixed computation budget, more sophisticated KD strategies, namely TinyBERT (Jiao et al., 2020) and MiniLM (Wang et al., 2023), outperform it by a notable margin. We further find that KD yields larger gains over pretraining from scratch when the data must be repeated under the fixed computation budget.

Maria Valentini, Téa Wright, Ali Marashian et al.

To address an important gap in creating children's stories for vocabulary enrichment, we investigate the automatic evaluation of how well stories convey the semantics of target vocabulary words, a task with substantial implications for generating educational content. We motivate this task, which we call measuring contextual informativeness in children's stories, and provide a formal task definition as well as a dataset for the task. We further propose a method for automating the task using a large language model (LLM). Our experiments show that our approach reaches a Spearman correlation of 0.4983 with human judgments of informativeness, while the strongest baseline only obtains a correlation of 0.3534. An additional analysis shows that the LLM-based approach is able to generalize to measuring contextual informativeness in adult-directed text, on which it also outperforms all baselines.

Minh Duc Bui, Katharina von der Wense

Current natural language processing (NLP) research tends to focus on only one or, less frequently, two dimensions - e.g., performance, privacy, fairness, or efficiency - at a time, which may lead to suboptimal conclusions and often overlooking the broader goal of achieving trustworthy NLP. Work on adapter modules (Houlsby et al., 2019; Hu et al., 2021) focuses on improving performance and efficiency, with no investigation of unintended consequences on other aspects such as fairness. To address this gap, we conduct experiments on three text classification datasets by either (1) finetuning all parameters or (2) using adapter modules. Regarding performance and efficiency, we confirm prior findings that the accuracy of adapter-enhanced models is roughly on par with that of fully finetuned models, while training time is substantially reduced. Regarding fairness, we show that adapter modules result in mixed fairness across sensitive groups. Further investigation reveals that, when the standard fine-tuned model exhibits limited biases, adapter modules typically do not introduce extra bias. On the other hand, when the finetuned model exhibits increased bias, the impact of adapter modules on bias becomes more unpredictable, introducing the risk of significantly magnifying these biases for certain groups. Our findings highlight the need for a case-by-case evaluation rather than a one-size-fits-all judgment.

Mario Sanz-Guerrero, Katharina von der Wense

Large language models (LLMs) are powerful zero- and few-shot learners. However, when predicting over a set of candidate options, LLMs suffer from label biases, and existing calibration methods overlook biases arising from multi-token class labels. We tackle an issue we call label length bias, where labels of different lengths are treated inconsistently, even after standard length normalization. To mitigate it, we propose normalized contextual calibration (NCC), an effective method that normalizes and calibrates predictions at the full-label level. NCC achieves statistically significant improvements over prior approaches across multiple datasets and models, with gains of up to 10% F1. Moreover, NCC extends bias mitigation to broader tasks such as multiple-choice question answering. Our analysis shows that, when combined with in-context learning, NCC is less sensitive to few-shot example selection, requires fewer examples for competitive performance, and produces more reliable confidence estimates. These findings highlight the importance of mitigating full-label biases to improve the performance and robustness of LLM-based methods, particularly in real-world applications where class labels naturally consist of multiple tokens.

Abteen Ebrahimi, Adam Wiemerslage, Katharina von der Wense

We present NN-Rank, an algorithm for ranking source languages for cross-lingual transfer, which leverages hidden representations from multilingual models and unlabeled target-language data. We experiment with two pretrained multilingual models and two tasks: part-of-speech tagging (POS) and named entity recognition (NER). We consider 51 source languages and evaluate on 56 and 72 target languages for POS and NER, respectively. When using in-domain data, NN-Rank beats state-of-the-art baselines that leverage lexical and linguistic features, with average improvements of up to 35.56 NDCG for POS and 18.14 NDCG for NER. As prior approaches can fall back to language-level features if target language data is not available, we show that NN-Rank remains competitive using only the Bible, an out-of-domain corpus available for a large number of languages. Ablations on the amount of unlabeled target data show that, for subsets consisting of as few as 25 examples, NN-Rank produces high-quality rankings which achieve 92.8% of the NDCG achieved using all available target data for ranking.

Hossain Shaikh Saadi, Minh Duc Bui, Mario Sanz-Guerrero et al.

This paper presents the JGU Mainz submission to the WMT25 Shared Task on LLMs with Limited Resources for Slavic Languages: Machine Translation and Question Answering, focusing on Ukrainian, Upper Sorbian, and Lower Sorbian. For each language, we jointly fine-tune a Qwen2.5-3B-Instruct model for both tasks with parameter-efficient finetuning. Our pipeline integrates additional translation and multiple-choice question answering (QA) data. For Ukrainian QA, we further use retrieval-augmented generation. We also apply ensembling for QA in Upper and Lower Sorbian. Experiments show that our models outperform the baseline on both tasks.

Mario Sanz-Guerrero, Minh Duc Bui, Katharina von der Wense

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string "Answer:" to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy -- tokenizing the space together with the answer letter -- as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model's confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.

Minh Duc Bui, Carolin Holtermann, Valentin Hofmann et al.

Dialects represent a significant component of human culture and are found across all regions of the world. In Germany, more than 40% of the population speaks a regional dialect (Adler and Hansen, 2022). However, despite cultural importance, individuals speaking dialects often face negative societal stereotypes. We examine whether such stereotypes are mirrored by large language models (LLMs). We draw on the sociolinguistic literature on dialect perception to analyze traits commonly associated with dialect speakers. Based on these traits, we assess the dialect naming bias and dialect usage bias expressed by LLMs in two tasks: an association task and a decision task. To assess a model's dialect usage bias, we construct a novel evaluation corpus that pairs sentences from seven regional German dialects (e.g., Alemannic and Bavarian) with their standard German counterparts. We find that: (1) in the association task, all evaluated LLMs exhibit significant dialect naming and dialect usage bias against German dialect speakers, reflected in negative adjective associations; (2) all models reproduce these dialect naming and dialect usage biases in their decision making; and (3) contrary to prior work showing minimal bias with explicit demographic mentions, we find that explicitly labeling linguistic demographics--German dialect speakers--amplifies bias more than implicit cues like dialect usage.

Enora Rice, Katharina von der Wense, Alexis Palmer

Computational morphology has the potential to support language documentation through tasks like morphological segmentation and the generation of Interlinear Glossed Text (IGT). However, our research outputs have seen limited use in real-world language documentation settings. This position paper situates the disconnect between computational morphology and language documentation within a broader misalignment between research and practice in NLP and argues that the field risks becoming decontextualized and ineffectual without systematic integration of User-Centered Design (UCD). To demonstrate how principles from UCD can reshape the research agenda, we present a case study of GlossLM, a state-of-the-art multilingual IGT generation model. Through a small-scale user study with three documentary linguists, we find that despite strong metric based performance, the system fails to meet core usability needs in real documentation contexts. These insights raise new research questions around model constraints, label standardization, segmentation, and personalization. We argue that centering users not only produces more effective tools, but surfaces richer, more relevant research directions

Mario Sanz-Guerrero, Katharina von der Wense

In-context learning (ICL) has transformed the use of large language models (LLMs) for NLP tasks, enabling few-shot learning by conditioning on labeled examples without finetuning. Despite its effectiveness, ICL is prone to errors, especially for challenging examples. With the goal of improving the performance of ICL, we propose corrective in-context learning (CICL), an approach that incorporates a model's incorrect predictions alongside ground truth corrections into the prompt, aiming to enhance classification accuracy through self-correction. However, contrary to our hypothesis, extensive experiments on text classification tasks demonstrate that CICL consistently underperforms standard ICL, with performance degrading as the proportion of corrections in the prompt increases. Our findings indicate that CICL introduces confusion by disrupting the model's task understanding, rather than refining its predictions. Additionally, we observe that presenting harder examples in standard ICL does not improve performance, suggesting that example difficulty alone may not be a reliable criterion for effective selection. By presenting these negative results, we provide important insights into the limitations of self-corrective mechanisms in LLMs and offer directions for future research.

Sagi Shaier, George Arthur Baker, Chiranthan Sridhar et al.

Language models (LMs) have excelled in various broad domains. However, to ensure their safe and effective integration into real-world educational settings, they must demonstrate proficiency in specific, granular areas of knowledge. Existing cloze-style benchmarks, commonly used to evaluate LMs' knowledge, have three major limitations. They: 1) do not cover the educational domain; 2) typically focus on low-complexity, generic knowledge or broad domains, which do not adequately assess the models' knowledge in specific subjects; and 3) often rely on templates that can bias model predictions. Here, we introduce MALAMUTE, a multilingual, template-free, and highly granular probing dataset comprising expert-written, peer-reviewed probes from 71 university-level textbooks across three languages (English, Spanish, and Polish). MALAMUTE is the first education-based cloze-style dataset. It covers eight domains, each with up to 14 subdomains, further broken down into concepts and concept-based prompts, totaling 33,361 university curriculum concepts and 116,887 prompts. MALAMUTE's fine granularity, educational focus, and inclusion of both sentence-level and paragraph-level prompts make it an ideal tool for evaluating LMs' course-related knowledge. Our evaluation of masked and causal LMs on MALAMUTE shows that despite overall proficiency, they have significant gaps in knowledge when examined closely on specific subjects, hindering their safe use in classrooms and underscoring the need for further development.

Sagi Shaier, Lawrence E Hunter, Katharina von der Wense

Natural language processing has seen rapid progress over the past decade. Due to the speed of developments, some practices get established without proper evaluation. Considering one such case and focusing on reading comprehension, we ask our first research question: 1) How does the order of inputs -- i.e., question and context -- affect model performance? Additionally, given recent advancements in input emphasis, we ask a second research question: 2) Does emphasizing either the question, the context, or both enhance performance? Experimenting with 9 large language models across 3 datasets, we find that presenting the context before the question improves model performance, with an accuracy increase of up to $31\%$. Furthermore, emphasizing the context yields superior results compared to question emphasis, and in general, emphasizing parts of the input is particularly effective for addressing questions that models lack the parametric knowledge to answer. Experimenting with both prompt-based and attention-based emphasis methods, we additionally find that the best method is surprisingly simple: it only requires concatenating a few tokens to the input and results in an accuracy improvement of up to $36\%$, allowing smaller models to outperform their significantly larger counterparts.

Enora Rice, Ali Marashian, Luke Gessler et al.

Canonical morphological segmentation is the process of analyzing words into the standard (aka underlying) forms of their constituent morphemes. This is a core task in language documentation, and NLP systems have the potential to dramatically speed up this process. But in typical language documentation settings, training data for canonical morpheme segmentation is scarce, making it difficult to train high quality models. However, translation data is often much more abundant, and, in this work, we present a method that attempts to leverage this data in the canonical segmentation task. We propose a character-level sequence-to-sequence model that incorporates representations of translations obtained from pretrained high-resource monolingual language models as an additional signal. Our model outperforms the baseline in a super-low resource setting but yields mixed results on training splits with more data. While further work is needed to make translations useful in higher-resource settings, our model shows promise in severely resource-constrained settings.