arXiv subjects

Walter Daelemans

Publications and source records attributed to Walter Daelemans.

At least 19 recordsLinked to original sources

arXiv2026

Benchmarking the Benchmarks: Testing the Predictive Validity of Commonsense Benchmarks

Predicting LLM's capabilities on real-world tasks is essential, yet the extent to which performance on commonsense benchmarks predicts downstream performance remains underspecified. To establish the practical usability of widely adopted commonsense benchmarks, we evaluate 23 models from six families on four established commonsense benchmarks, four reworked variants, three non-commonsense controls, and eight downstream tasks requiring implicit social, pragmatic, temporal, or physical reasoning. We compare model rankings, compute controlled correlations, and use leave-one-family-out cross-validation to assess the criterion validity of commonsense benchmarks. Our results show that revised benchmarks largely preserve original model rankings and do not improve downstream predictive power. Commonsense benchmarks show consistent cross-family predictive validity for only a narrow subset of downstream tasks, with smaller or metric-specific gains elsewhere. Overall, standardized commonsense benchmarks provide task-dependent rather than broad evidence of downstream commonsense competence.

cs.CL↗

arXiv2026

Decomposing Factual Sycophancy in Language Models: How Size and Instruction Tuning Shape Robustness

Factual sycophancy occurs when a language model abandons a correct, verifiable answer under social pressure. Because a flip occurs only when pressure toward a false answer exceeds the model's neutral preference for the truth, flip rates conflate two mechanisms: the strength of that baseline preference (truth margin), and how far pressure shifts it (manipulation sensitivity). We decompose factual sycophancy into these channels and use them to separate the effects of size and instruction tuning across 56 open-weight models spanning 0.3B-32B parameters and 13 manipulation types. We find that vulnerability is governed mainly by size, but instruction tuning changes how size acts: small instruction-tuned models can become less robust, whereas large instruction-tuned models usually become more robust. Instruction tuning primarily increases truth margin, but its behavioral effect depends on manipulation type. Scaling also changes the two channels differently: base models gain margin but become mildly more manipulation-sensitive, whereas instruction-tuned models gain margin faster and become less sensitive. Factual sycophancy is therefore not a single scalar property. Evaluations should report channel-specific, manipulation-specific, and size-conditioned robustness rather than flip rates alone.

cs.CL↗

arXiv2025

An Agentic AI Framework for Training General Practitioner Student Skills

Advancements in large language models offer strong potential for enhancing virtual simulated patients (VSPs) in medical education by providing scalable alternatives to resource-intensive traditional methods. However, current VSPs often struggle with medical accuracy, consistent roleplaying, scenario generation for VSP use, and educationally structured feedback. We introduce an agentic framework for training general practitioner student skills that unifies (i) configurable, evidence-based vignette generation, (ii) controlled persona-driven patient dialogue with optional retrieval grounding, and (iii) standards-based assessment and feedback for both communication and clinical reasoning. We instantiate the framework in an interactive spoken consultation setting and evaluate it with medical students ($\mathbf{N{=}14}$). Participants reported realistic and vignette-faithful dialogue, appropriate difficulty calibration, a stable personality signal, and highly useful example-rich feedback, alongside excellent overall usability. These results support agentic separation of scenario control, interaction control, and standards-based assessment as a practical pattern for building dependable and pedagogically valuable VSP training tools.

cs.CL↗

arXiv2025

Do You Get the Hint? Benchmarking LLMs on the Board Game Concept

Large language models (LLMs) have achieved striking successes on many benchmarks, yet recent studies continue to expose fundamental weaknesses. In this paper, we introduce Concept, a simple word-guessing board game, as a benchmark for probing abductive reasoning. Our results show that this game, easily solved by humans (with a success rate of over 90\%), is still very challenging for state-of-the-art LLMs (no model exceeds 40\% success rate). Specifically, we observe that LLMs struggle with interpreting other players' strategic intents, and with correcting initial hypotheses given sequential information updates. In addition, we extend the evaluation across multiple languages, and find that the LLM performance drops further in lower-resource languages (Dutch, French, and Spanish) compared to English.

cs.CL↗

arXiv2025

One Size Does Not Fit All: Exploring Variable Thresholds for Distance-Based Multi-Label Text Classification

Distance-based unsupervised text classification is a method within text classification that leverages the semantic similarity between a label and a text to determine label relevance. This method provides numerous benefits, including fast inference and adaptability to expanding label sets, as opposed to zero-shot, few-shot, and fine-tuned neural networks that require re-training in such cases. In multi-label distance-based classification and information retrieval algorithms, thresholds are required to determine whether a text instance is "similar" to a label or query. Similarity between a text and label is determined in a dense embedding space, usually generated by state-of-the-art sentence encoders. Multi-label classification complicates matters, as a text instance can have multiple true labels, unlike in multi-class or binary classification, where each instance is assigned only one label. We expand upon previous literature on this underexplored topic by thoroughly examining and evaluating the ability of sentence encoders to perform distance-based classification. First, we perform an exploratory study to verify whether the semantic relationships between texts and labels vary across models, datasets, and label sets by conducting experiments on a diverse collection of realistic multi-label text classification (MLTC) datasets. We find that similarity distributions show statistically significant differences across models, datasets and even label sets. We propose a novel method for optimizing label-specific thresholds using a validation set. Our label-specific thresholding method achieves an average improvement of 46% over normalized 0.5 thresholding and outperforms uniform thresholding approaches from previous work by an average of 14%. Additionally, the method demonstrates strong performance even with limited labeled examples.

cs.CL↗

arXiv2025

MTEB-NL and E5-NL: Embedding Benchmark and Models for Dutch

Recently, embedding resources, including models, benchmarks, and datasets, have been widely released to support a variety of languages. However, the Dutch language remains underrepresented, typically comprising only a small fraction of the published multilingual resources. To address this gap and encourage the further development of Dutch embeddings, we introduce new resources for their evaluation and generation. First, we introduce the Massive Text Embedding Benchmark for Dutch (MTEB-NL), which includes both existing Dutch datasets and newly created ones, covering a wide range of tasks. Second, we provide a training dataset compiled from available Dutch retrieval datasets, complemented with synthetic data generated by large language models to expand task coverage beyond retrieval. Finally, we release a series of E5-NL models compact yet efficient embedding models that demonstrate strong performance across multiple tasks. We make our resources publicly available through the Hugging Face Hub and the MTEB package.

cs.CL↗

arXiv2025

WinoWhat: A Parallel Corpus of Paraphrased WinoGrande Sentences with Common Sense Categorization

In this study, we take a closer look at how Winograd schema challenges can be used to evaluate common sense reasoning in LLMs. Specifically, we evaluate generative models of different sizes on the popular WinoGrande benchmark. We release WinoWhat, a new corpus, in which each instance of the WinoGrande validation set is paraphrased. Additionally, we evaluate the performance on the challenge across five common sense knowledge categories, giving more fine-grained insights on what types of knowledge are more challenging for LLMs. Surprisingly, all models perform significantly worse on WinoWhat, implying that LLM reasoning capabilities are overestimated on WinoGrande. To verify whether this is an effect of benchmark memorization, we match benchmark instances to LLM trainingdata and create two test-suites. We observe that memorization has a minimal effect on model performance on WinoGrande.

cs.CL↗

arXiv2024

BEIR-NL: Zero-shot Information Retrieval Benchmark for the Dutch Language

Zero-shot evaluation of information retrieval (IR) models is often performed using BEIR; a large and heterogeneous benchmark composed of multiple datasets, covering different retrieval tasks across various domains. Although BEIR has become a standard benchmark for the zero-shot setup, its exclusively English content reduces its utility for underrepresented languages in IR, including Dutch. To address this limitation and encourage the development of Dutch IR models, we introduce BEIR-NL by automatically translating the publicly accessible BEIR datasets into Dutch. Using BEIR-NL, we evaluated a wide range of multilingual dense ranking and reranking models, as well as the lexical BM25 method. Our experiments show that BM25 remains a competitive baseline, and is only outperformed by the larger dense models trained for retrieval. When combined with reranking models, BM25 achieves performance on par with the best dense ranking models. In addition, we explored the impact of translation on the data by back-translating a selection of datasets to English, and observed a performance drop for both dense and lexical methods, indicating the limitations of translation for creating benchmarks. BEIR-NL is publicly available on the Hugging Face hub.

cs.CL↗

arXiv2024

Bilingual BSARD: Extending Statutory Article Retrieval to Dutch

Statutory article retrieval plays a crucial role in making legal information more accessible to both laypeople and legal professionals. Multilingual countries like Belgium present unique challenges for retrieval models due to the need for handling legal issues in multiple languages. Building on the Belgian Statutory Article Retrieval Dataset (BSARD) in French, we introduce the bilingual version of this dataset, bBSARD. The dataset contains parallel Belgian statutory articles in both French and Dutch, along with legal questions from BSARD and their Dutch translation. Using bBSARD, we conduct extensive benchmarking of retrieval models available for Dutch and French. Our benchmarking setup includes lexical models, zero-shot dense models, and fine-tuned small foundation models. Our experiments show that BM25 remains a competitive baseline compared to many zero-shot dense models in both languages. We also observe that while proprietary models outperform open alternatives in the zero-shot setting, they can be matched or surpassed by fine-tuning small language-specific models. Our dataset and evaluation code are publicly available.

cs.CL↗

arXiv2024

Bag of Lies: Robustness in Continuous Pre-training BERT

This study aims to acquire more insights into the continuous pre-training phase of BERT regarding entity knowledge, using the COVID-19 pandemic as a case study. Since the pandemic emerged after the last update of BERT's pre-training data, the model has little to no entity knowledge about COVID-19. Using continuous pre-training, we control what entity knowledge is available to the model. We compare the baseline BERT model with the further pre-trained variants on the fact-checking benchmark Check-COVID. To test the robustness of continuous pre-training, we experiment with several adversarial methods to manipulate the input data, such as training on misinformation and shuffling the word order until the input becomes nonsensical. Surprisingly, our findings reveal that these methods do not degrade, and sometimes even improve, the model's downstream performance. This suggests that continuous pre-training of BERT is robust against misinformation. Furthermore, we are releasing a new dataset, consisting of original texts from academic publications in the LitCovid repository and their AI-generated false counterparts.

cs.CL↗

arXiv2024

PersonalityChat: Conversation Distillation for Personalized Dialog Modeling with Facts and Traits

The new wave of Large Language Models (LLM) has offered an efficient tool to curate sizeable conversational datasets. So far studies have mainly focused on task-oriented or generic open-domain dialogs, and have not fully explored the ability of LLMs in following complicated prompts. In this work, we focus on personalization, and employ LLMs to curate a dataset which is difficult and costly to crowd-source: PersonalityChat is a synthetic conversational dataset based upon the popular PersonaChat dataset, but conditioned on both personas and (Big-5) personality traits. Evaluating models fine-tuned on this dataset, we show that the personality trait labels can be used for trait-based personalization of generative dialogue models. We also perform a head-to-head comparison between PersonalityChat and PersonaChat, and show that training on the distilled dataset results in more fluent and coherent dialog agents in the small-model regime.

cs.CL↗

arXiv2022

Open-Domain Dialog Evaluation using Follow-Ups Likelihood

Automatic evaluation of open-domain dialogs remains an unsolved problem. Moreover, existing methods do not correlate strongly with human annotations. This paper presents a new automated evaluation method using follow-ups: we measure the probability that a language model will continue the conversation with a fixed set of follow-ups (e.g., not really relevant here, what are you trying to say). When compared against twelve existing methods, our new evaluation achieves the highest correlation with human evaluations.

cs.CL↗

arXiv2022

CoNTACT: A Dutch COVID-19 Adapted BERT for Vaccine Hesitancy and Argumentation Detection

We present CoNTACT: a Dutch language model adapted to the domain of COVID-19 tweets. The model was developed by continuing the pre-training phase of RobBERT (Delobelle, 2020) by using 2.8M Dutch COVID-19 related tweets posted in 2021. In order to test the performance of the model and compare it to RobBERT, the two models were tested on two tasks: (1) binary vaccine hesitancy detection and (2) detection of arguments for vaccine hesitancy. For both tasks, not only Twitter but also Facebook data was used to show cross-genre performance. In our experiments, CoNTACT showed statistically significant gains over RobBERT in all experiments for task 1. For task 2, we observed substantial improvements in virtually all classes in all experiments. An error analysis indicated that the domain adaptation yielded better representations of domain-specific terminology, causing CoNTACT to make more accurate classification decisions.

cs.CL↗

arXiv2022

Cyberbullying Classifiers are Sensitive to Model-Agnostic Perturbations

A limited amount of studies investigates the role of model-agnostic adversarial behavior in toxic content classification. As toxicity classifiers predominantly rely on lexical cues, (deliberately) creative and evolving language-use can be detrimental to the utility of current corpora and state-of-the-art models when they are deployed for content moderation. The less training data is available, the more vulnerable models might become. This study is, to our knowledge, the first to investigate the effect of adversarial behavior and augmentation for cyberbullying detection. We demonstrate that model-agnostic lexical substitutions significantly hurt classifier performance. Moreover, when these perturbed samples are used for augmentation, we show models become robust against word-level perturbations at a slight trade-off in overall task performance. Augmentations proposed in prior work on toxicity prove to be less effective. Our results underline the need for such evaluations in online harm areas with small corpora. The perturbed data, models, and code are available for reproduction at https://github.com/cmry/augtox

cs.CL↗

arXiv2021

ConveRT for FAQ Answering

Knowledgeable FAQ chatbots are a valuable resource to any organization. While powerful and efficient retrieval-based models exist for English, it is rarely the case for other languages for which the same amount of training data is not available. In this paper, we propose a novel pre-training procedure to adapt ConveRT, an English conversational retriever model, to other languages with less training data available. We apply it for the first time to the task of Dutch FAQ answering related to the COVID-19 vaccine. We show it performs better than an open-source alternative in both a low-data regime and a high-data regime.

cs.CL↗

arXiv2021

MFAQ: a Multilingual FAQ Dataset

In this paper, we present the first multilingual FAQ dataset publicly available. We collected around 6M FAQ pairs from the web, in 21 different languages. Although this is significantly larger than existing FAQ retrieval datasets, it comes with its own challenges: duplication of content and uneven distribution of topics. We adopt a similar setup as Dense Passage Retrieval (DPR) and test various bi-encoders on this dataset. Our experiments reveal that a multilingual model based on XLM-RoBERTa achieves the best results, except for English. Lower resources languages seem to learn from one another as a multilingual model achieves a higher MRR than language-specific ones. Our qualitative analysis reveals the brittleness of the model on simple word changes. We publicly release our dataset, model and training script.

cs.CL↗

arXiv2021

Teach Me What to Say and I Will Learn What to Pick: Unsupervised Knowledge Selection Through Response Generation with Pretrained Generative Models

Knowledge Grounded Conversation Models (KGCM) are usually based on a selection/retrieval module and a generation module, trained separately or simultaneously, with or without having access to a gold knowledge option. With the introduction of large pre-trained generative models, the selection and generation part have become more and more entangled, shifting the focus towards enhancing knowledge incorporation (from multiple sources) instead of trying to pick the best knowledge option. These approaches however depend on knowledge labels and/or a separate dense retriever for their best performance. In this work we study the unsupervised selection abilities of pre-trained generative models (e.g. BART) and show that by adding a score-and-aggregate module between encoder and decoder, they are capable of learning to pick the proper knowledge through minimising the language modelling loss (i.e. without having access to knowledge labels). Trained as such, our model - K-Mine - shows competitive selection and generation performance against models that benefit from knowledge labels and/or separate dense retriever.

cs.CL↗

arXiv2020

Character-level Transformer-based Neural Machine Translation

Neural machine translation (NMT) is nowadays commonly applied at the subword level, using byte-pair encoding. A promising alternative approach focuses on character-level translation, which simplifies processing pipelines in NMT considerably. This approach, however, must consider relatively longer sequences, rendering the training process prohibitively expensive. In this paper, we discuss a novel, Transformer-based approach, that we compare, both in speed and in quality to the Transformer at subword and character levels, as well as previously developed character-level models. We evaluate our models on 4 language pairs from WMT'15: DE-EN, CS-EN, FI-EN and RU-EN. The proposed novel architecture can be trained on a single GPU and is 34% percent faster than the character-level Transformer; still, the obtained results are at least on par with it. In addition, our proposed model outperforms the subword-level model in FI-EN and shows close results in CS-EN. To stimulate further research in this area and close the gap with subword-level NMT, we make all our code and models publicly available.

cs.CL↗