Yesterday, Imanol Pradales, the Lehendakari, together with Ibone Bengoetxea, First Vice Lehendakari and Minister of Culture and Language Policy, and Juan Ignacio Pérez Iglesias, Minister of Science, University and Innovation, visited our HiTZ Centre. www.irekia.euskadi.eus/en/news/1031...

Atzo goizean, Imanol Pradales lehendakaria, Ibone Bengoetxea lehen lehendakariorde eta Kultura eta Hizkuntza Politikako sailburua, eta @jipereziglesias.bsky.social @zubsaila.bsky.social sailburua, HiTZ Zentroan izan ziren. www.irekia.euskadi.eus/eu/news/1031...

Key takeaway: Adding simple structure at inference-time, through image crops and text segments, is a powerful, training-free way to improve Vision-Language Compositionality performance. Joint work with @Ander Salaberria @eagirre.bsky.social @gazkune.bsky.social @hitz-zentroa.bsky.social

Our analysis shows that: 1. There is room to improve the quality of extracted text segments. 2. Our method achieves significant performance gains in Winoground's non-trivial instances. 3. Isolated image crops can lose size and quantity information, leaving room for improvement.

Why are image crops crucial? 🤔 We found that simply adding text segments isn't enough. The biggest performance gains come when text segments are paired with image crops, proving the power of serial image computing.

Our approach is straightforward yet effective: 1. Divide the image into smaller crops. 2. Extract text segments capturing objects, attributes and relations. 3. Use the VLM to find image crops that best fit the text segments. 4. Aggregate matching similarities for the final score.

#newHitzPaper Can a simple inference-time approach unlock better Vision-Language Compositionality?🤯 Our latest paper shows how adding structure at inference significantly boosts performance in popular dual-encoder VLMs on different datasets. Read more: arxiv.org/abs/2506.09691

While the experiments were not complicated, they required the collaboration of amazing co-authors, many compute hours, and of course, the impressive collaboration of the Basque community that was involved in manually assessing the models on an arena style evaluation. Thank you!

In this work we face the challenge of developing instruct models for Basque, a low-resource language. Continue pretraining base models is intuitive, but what about instructed models? We analyze systematically all different approaches to find the best solution. 2/3

Do you know that you can continue pretraining Instructed LLMs without losing their instruction following capabilities? We did so to teach Basque to Llama models with promising results! Interestingly, you only need English instructions and target language corpora 🤯 1/3

Ostegunero, HiTZ zentroko kideak biltzen gara gure ikerketen berri emateko HiTZ mintegian. Aste honetan, bi tesi proiektu aurkeztu dira: Irune Zubiagak" Learning to Judge: Automated Multilingual Evaluation of LLM-Generated Text" eta @blanca-calvo-figue.bsky.social-ek "Critical Questions Generation"

@oierldl.bsky.social and I are hiring for a post-doc position starting in September within the LUMINOUS project, where we focus on multi-modal LLMs. If you are interested or have questions, feel free to get in contact with us or apply directly at the link below. link: lnkd.in/d-Zw6Ysh

Apirilaren 30ean, EHUren Zientzia eta Teknologia Fakultateko Paraninfoan, Xabier Arregik "EUSKARAREN ERRONKAK UNIBERTSO DIGITALEAN: Adimen Artifiziala" izeneko mahai-inguruan parte hartuko du.

Tomorrow we will have @ebugliarello.bsky.social in our webinar series at @hitz-zentroa.bsky.social . He will talk about vision-language models in the multilingual and multicultural context, exploring evaluation, training and generation. Info here: www.hitz.eus/webinars

📊 Ebaluatoiaren lehen 5 egunetako datuak! 📊 775+ erabiltzaile eta 6000+ bidalketa! 🚀 Mila esker guztioi! 💕 Erronka: 20000 bidalketa lortzea apirilaren 2a baino lehen! 🕒 Sartu ebaluatoia.hitz.eus eta egin zure galdera!

Euskal Herriko Unibertsitateak #HITZ bezalako erreferentziazko ikerketa zentroak izateak harrotasunez betetzen gaitu. Eneko Agirre zuzendaria eta bere taldekideekin izan gara goizean. Jarrai dezazuela bide berriak urratzen hizkuntza eta ahotsa ardatz duen AAren inguruan, @hitz-zentroa.bsky.social.

📢#ebaluatoia Euskarazko txatbota hobetzen laguntzeko deia egin du @hitz-zentroa.bsky.social -ak. Martxoaren 19tik apirilaren 2ra bitartean galdetu #Latxa ri eta baloratu erantzuna. Eta parte hartu: www.ehu.eus/eu/web/campu... Umorea ere baduela konprobatu dugu! Begira👇

Adimen artifizialeko adituen bila zabiltza? HiTZ zentroan badituzu hamaika emakume! #M8 #martxoak8

🧙‍♂️ New paper 🧙‍♀️: Presenting Wicked: a simple automated method to make MCQA benchmarks more challenging. Wicked shook up 18 open-weight LLMs on 6 benchmarks, with up to 19.7% performance drop with direct prompting 🤯 Paper: shorturl.at/1CGq0 Code: shorturl.at/n2nCU

Iñigo Alonsok “Improving Fidelity and Table Representation in Table Understanding and Table-to-Text Generation” doktore-tesia defendatu du gaur. Zorionak, Iñigo!

German Rigau zuzendariordea Parisen dago "Language Technologies for All" konferentzian parte hartzen. www.lt4all2025.eu

Bisita izan dugu! Eusko Jaurlaritzako Hizkuntza Politikarako Sailburuordetzako talde zabalarekin, Aitor Aldasoro sailburuordea, Josune Irabien eta Sonia Rodriguez barne, arratsalde ederra pasa genuen zentroa erakusten eta Ikergaitu proiektuko azken emaitzak erakusten https://www.hitz.eus/iker-gaitu/

Iker Garcia-Ferrerok "Cross-Lingual Transfer for Low-Resource Natural Language Processing" doktore-tesia defendatu du gaur. Zorionak, Iker!

Gogoratu otsailaren 17rarte aukera duzuela euskarazko txatbotak ebaluatu eta ikerkuntzan laguntzeko: ebaluatoia.hitz.eus Sartu eta parte hartu, erraza eta dibertigarria izateaz gain, sariak ere badaude!

Atzo, otsailak 11, ETB1eko Egun On saioan aritu zen @eagirre.bsky.social HiTZeko zuzendaria Ebaluatoiari buruz hizketan. Hemen topatu dezakezue saio osoaren grabazioa, bere parte-hartzea 10:46 inguruan izan zen (grabazioaren ~1:46): www.eitb.eus/eu/nahieran/...

#emakumeakZientzian egunero!

Euskarazko txatbot publikoak hobetzeko zure laguntza behar dugu! https://ebaluatoia.hitz.eus Sartu, galdetu eta erantzun hoberena aukeratu. Erraza eta dibertigarria, lehiaketa eta sariak daude!

HiTZen hainbat txatbot lehian jarri ditugu, hipotesi zientifikoak baieztatzeko asmoz. Denetarik dago, txatbot onak, oso onak eta txarrak ere. Otsailaren 17 arte baietz 20 mila erantzun lortu! https://ebaluatoia.hitz.eus

Iker de la Iglesiak medikuntzako argumentuen ebaluazio automatikoari buruzko posterra aurkeztu du #COLING2025 konferentzian

We are in #COLING2025!

Paula has presented this paper about efficient ways of augmenting Language Models with visual knowledge at COLING 2025. Check it!

Paula Ontalvillak "Improving the Efficiency of Visually Augmented Language Models" lana aurkeztu du COLING konferentzian

Gaur 18:30etan, Euskadi Irratiko 'Baipasa' irratsaioan, Ane Garcia eta Josu Goikoetxea HiTZ kideek hizkuntzaren prozesamendua, medikuntza eta euskera izango dituzte hizpide.

BiVLC: Extending Vision-Language Compositionality Evaluation with Text-to-Image Retrieval by Imanol Miranda, Ander Salaberria, @eagirre.bsky.social, @gazkune.bsky.social neurips.cc/virtual/2024...

BertaQA: How Much Do Language Models Know About Local Culture? by @juletxara.bsky.social @gazkune.bsky.social @asoroa.bsky.social Oier Lopez de Lacalle and Mikel Artetxe nips.cc/virtual/2024...

We presented two posters at #NeurIPS2024! @gazkune.bsky.social @juletxara.bsky.social and Imanol Miranda

If you are interested in LLMs' knowledge about local cultures, check this paper by @juletxara.bsky.social and come to discuss with us in Vancouver #NeurIPS2024 arxiv.org/abs/2406.07302

Very proud of the work done by Imanol Miranda in @hitz-zentroa.bsky.social Check the camera-ready version for new experiments with very interesting results and see you in Vancouver! #NeurIPS2024 arxiv.org/abs/2406.09952

🚀 Introducing INCLUDE 🌍: A multilingual LLM evaluation benchmark spanning 44 languages! Contains *newly-collected* data, prioritizing *regional knowledge*. Setting the stage for truly global AI evaluation. Ready to see how your model measures up? #AI #Multilingual #LLM #NLProc

Paula Ontalvilla, Aitor Ormazabal, Gorka Azkune Improving the Efficiency of Visually Augmented Language Models https://arxiv.org/abs/2409.11148

I couldn't be happier! Preparing the trip to #NeurIPS2024 we received the notification of acceptance of our paper in #COLING2025 with Paula Ontalvilla and Aitor Ormazabal. Paula has just started with her PhD at @hitz-zentroa.bsky.social and she is already publishing at this level!