Luna De Bruyne

Victor De Marez, Luna De Bruyne, Walter Daelemans

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.

Eleni Gkovedarou, Joke Daems, Luna De Bruyne

As the demand for inclusive language increases, concern has grown over the susceptibility of machine translation (MT) systems to reinforce gender stereotypes. This study investigates gender bias in two commercial MT systems, Google Translate and DeepL, focusing on the understudied English-to-Greek language pair. We address three aspects of gender bias: i) male bias, ii) occupational stereotyping, and iii) errors in anti-stereotypical translations. Additionally, we explore the potential of prompted GPT-4o as a bias mitigation tool that provides both gender-explicit and gender-neutral alternatives when necessary. To achieve this, we introduce GendEL, a manually crafted bilingual dataset of 240 gender-ambiguous and unambiguous sentences that feature stereotypical occupational nouns and adjectives. We find persistent gender bias in translations by both MT systems; while they perform well in cases where gender is explicitly defined, with DeepL outperforming both Google Translate and GPT-4o in feminine gender-unambiguous sentences, they are far from producing gender-inclusive or neutral translations when the gender is unspecified. GPT-4o shows promise, generating appropriate gendered and neutral alternatives for most ambiguous cases, though residual biases remain evident.

Ine Gevers, Victor De Marez, Luna De Bruyne et al.

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.

Luna De Bruyne, Pepa Atanasova, Isabelle Augenstein

Emotion lexica are commonly used resources to combat data poverty in automatic emotion detection. However, vocabulary coverage issues, differences in construction method and discrepancies in emotion framework and representation result in a heterogeneous landscape of emotion detection resources, calling for a unified approach to utilising them. To combat this, we present an extended emotion lexicon of 30,273 unique entries, which is a result of merging eight existing emotion lexica by means of a multi-view variational autoencoder (VAE). We showed that a VAE is a valid approach for combining lexica with different label spaces into a joint emotion label space with a chosen number of dimensions, and that these dimensions are still interpretable. We tested the utility of the unified VAE lexicon by employing the lexicon values as features in an emotion detection model. We found that the VAE lexicon outperformed individual lexica, but contrary to our expectations, it did not outperform a naive concatenation of lexica, although it did contribute to the naive concatenation when added as an extra lexicon. Furthermore, using lexicon information as additional features on top of state-of-the-art language models usually resulted in a better performance than when no lexicon information was used.