🏆 full house at the IJCAI-JAIR Best Paper Award #talk delivered by Verna Dankers on Compositionality Decomposed: How do Neural Networks Generalise?! #IJCAI2025

Thanks Jelle!! I'll make sure to prominently feature UvA in the talk tomorrow 😎 all discussions for Compositionality Decomposed were held back at Nikhef!

Congratulations to the winners of the 2025 IJCAI–JAIR Prize for their paper “Compositionality Decomposed: How Do Neural Networks Generalise?” — Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni! Presented by Edith Elkind, Northwestern University arxiv.org/abs/1908.08351
#IJCAI2025

Proud to accept a 5y outstanding paper award
@ijcai.org 🏆 from the Journal of AI Research for the impact Compositionality Decomposed has had, on behalf of the team (Dieuwke Hupkes, Elia Bruni & Mathijs Mul)!🧡 Come to room 513 on [email protected] to learn about rethinking compgen evaluation in the LLM era🤖

Had a blast at #ACL2025 connecting with new/familiar faces, explaining the many affiliations on my badge, chatting about memorisation--generalisation, and visiting stunning Schönbrunn. A shoutout to the co-organisers & speakers of the successful @l2m2workshop.bsky.social 🧡 I learnt a lot from you!

Main takeaway: students excel both by mimicking teachers and deviating from them & be careful with distillation, students may inherit teachers’ pros and cons. Work done during a Microsoft internship. Find me 👋🏼 Weds at 11 in poster session 4 arxiv.org/pdf/2502.01491 (5/5)

...identifying scenarios in which students *outperform* teachers through SeqKD’s amplified denoising effects. Lastly, we establish that through AdaptiveSeqKD (briefly finetuning your teacher prior to distillation with HQ data) you can strongly decrease memorization and hallucinations. (4/5)

Students showed increases in verbatim memorization, large increases in extractive memorization, and also hallucinated much more than baselines! We go beyond average-case performance through additional analyses of how T, B, and S perform on data subgroups... (3/5)

SeqKD is still widely applied in NMT to obtain strong & small deployable systems, but your student models do not only inherit good things from teachers. In our short paper, we contrast baselines trained on the original corpus to students (trained on teacher-generated targets). (2/5)

Thrilled to be in Vienna to learn about all the papers & catch up with NLP friends 🇦🇹 You'll find me at the ACL mentorship session (Mon 2PM), at our @l2m2workshop.bsky.social (Fri) and at my poster (Wed 11AM)! I'll present work w/ @vyraun.bsky.social on memorization in NMT under SeqKD #ACL2025 (1/5)

I miss Edinburgh and its wonderful people already!! Thanks to @tallinzen.bsky.social and @edoardo-ponti.bsky.social for inspiring discussions during the viva! I'm now exchanging Arthur's Seat for Mont Royal to join @sivareddyg.bsky.social's wonderful lab @mila-quebec.bsky.social 🤩

Models like DeepSeek-R1 🐋 mark a fundamental shift in how LLMs approach complex problems. In our preprint on R1 Thoughtology, we study R1’s reasoning chains across a variety of tasks; investigating its capabilities, limitations, and behaviour.
🔗: mcgill-nlp.github.io/thoughtology/

👋

Super excited about the First L2M2 (Large Language Model Memorization) workshop at #ACL2025 in Vienna @l2m2workshop.bsky.social! 🥳 Submit via ARR in February, or directly to the workshop in March! Archival and non-archival contributions are welcome. 🗒️ sites.google.com/view/memoriz...

thinking of calling this "The Illusion Illusion"

(more examples below)