Mario Sanz

PhD student in #NLProc at @NALACUJGU

[Mario Sanz @_mariosanz]

Mind the gap when evaluating LLMs with multiple-choice QA 🚨

In our #EMNLP2025 paper, we show that a tiny space tokenization can shift accuracy by up to 11% – and even reshuffle leaderboards.

Big thanks to my great co-authors @minhducbui_nlp & @kelina1124!

[NALA @NALACUJGU]

🧐 Evaluating your LLM with multiple-choice question answering?

🧵 A tiny space in the prompt can make accuracy jump by 11% – and even reshuffle model rankings.

#EMNLP2025 #NLP #AI #LLM #Evaluation

[Minh Duc Bui @minhducbui_nlp]

Your dialect could change how AI perceives you. 🗣️ In our #EMNLP2025 paper, we uncover systematic German dialect bias in leading LLMs.

Grateful to my amazing collaborators who made this work possible: @CarolinHolterm* @vjhofmann @anne_lauscher @kelina1124 🙌

[NALA @NALACUJGU]

"You speak Bavarian? Then you must be uneducated and closed-minded!"

🤯 Not your opinion? Good. But it might be your LLM's!

🧵 In our #EMNLP2025 paper we uncover concerning dialect bias in recent LLMs - including GPT-5.

#AI #Bias #Dialect #Fairness #LLM #NLProc #Safety

[NALA @NALACUJGU]

Great news from the @NALACUJGU Group: we’ll be presenting 7(!) papers at #EMNLP2025! 🙌 Stay tuned, we’ll be sharing summaries of all papers soon!

[Minh Duc Bui @minhducbui_nlp]

🏆 Our paper has received the Outstanding Paper Award at @naaclmeeting! 🎉 Many thanks to my co-authors @kelina1124 and @anne_lauscher!

We introduce Multi3Hate, a novel multimodal and multilingual parallel hate speech dataset annotated by a multicultural set of annotators.

[Informática UCM @informaticaucm]

Mario Sanz, estudiante de GII, primer premio nacional Laboral Kutxa "Transformación de las finanzas para la sociedad" por su TFG en el que aplicaba IA explicable y modelos de lenguaje grandes al riesgo de crédito. https://t.co/LHVBMz3MBv

[Informática UCM @informaticaucm]

Tercer premio:
Mario Sanz Guerrero
Evaluación del rendimiento de modelos de riesgo crediticio con algoritmos de boosting y transfer learning sobre modelos grandes de lenguaje