Carolin Holtermann

🧠 Which knowledge module do I need and how much of it ⁉️ 🌟 Check out our latest paper, "What the Weight?! A Unified Framework for Zero-Shot Knowledge Composition", which has been officially accepted for Findings of the EACL Conference! arxiv.org/pdf/2401.12756…

A huge thank you to all the amazing authors of this work!! ✨@minhducbui_nlp @KaitlynZhou @vjhofmann @kelina1124 @anne_lauscher

🚨 New paper alert: A new generation of LLMs can now process speech natively. This could expand access for millions excluded by text interfaces, but our research shows a cost: demographic cues in speaker voice can trigger stereotypical model responses. 🎙️⚖️

Link to paper: arxiv.org/abs/2603.22260 Link to website: tai-hamburg.github.io/papers/diverse…

🚨The result is a self-reinforcing gap: systems optimize for frequent users, while growing more distant from those they were supposed to reach. Fairness and accessibility have to be tackled together.

🛠️ Study 3: Mitigation exists. Discrimination tracks continuously with vocal pitch. Systematically lowering pitch reduced bias to statistically insignificant levels across all tested models.

📣 CFP: Eval4SD @ KONVENS 2026, Hamburg (Sept 14–17) 1st workshop on evaluating LLMs for specialized domains. We invite work on benchmarking, replication, and evaluation methodology across diverse fields. Deadline: Jul 3, 2026, 23:59 CEST eval4sd.github.io

✨Preprint Alert ✨ Thrilled to share that our paper has been accepted at the #EMNLP2025 main conference! 🚨 We uncover harmful stereotypical biases in LLMs toward German dialect speakers. 🚨 Check out the paper: arxiv.org/abs/2509.13835 Many thanks to all the great authors! 🙌🏼

Want to train a *multilingual* LVLM but not sure how? Or looking for a strong model to use? Presenting "Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model"! Arxiv: arxiv.org/abs/2501.05122 HF Collection: huggingface.co/collections/Wu…

📣 I am looking for a new Ph.D. student! 📣 The position is part of a @DASHHPhD project on (multimodal) #LLMs for laser operations with me (@DS_research_HH @unihh) and @HTuennermann from @desy. Deadline: October 20th! Feel free to reach out! 🚀🛸💥 dashh.org/application/ph…

🎉 We are happy to announce that the Data Science Lab Hamburg has 4 papers accepted at the #acl2024 in Bangkok🇹🇭! @KatiaShutova @anne_lauscher @CarolinHolterm @FrohmannM @manuellardelli @peppeatta @paul_rottger

📢Hi X Community! We are the #datascience research group led by @anne_lauscher in Hamburg! We are doing #NLProc research relating to various topics such as #LLMs, multimodality, #EthNLP, etc. Follow our new account if you are interested in latest results, insights, and more! 📢

🎉 Our paper "ScaLearn: Simple and Highly Parameter-Efficient Task Transfer by Learning to Scale" has been accepted to #ACL2024 findings! Congrats to my amazing co-authors @CarolinHolterm, Shahed Masoudian, @anne_lauscher, Navid Rekabsaz!

I recently gave a talk on fairness & inclusion in #LLMs at the annual meeting of @IT_Frauen of @informatikradar. Here, I present insights from our recent work lead by @CarolinHolterm & @paul_rottger. Thanks to @ada_gelie for the picture! #NLProc

@rasbt Very cool collection! We have a paper at Findings of the EACL 2024 where we present a unified framework for zero-shot composition of knowledge modules using parameter averaging and output ensembling: arxiv.org/pdf/2401.12756…

Weight averaging and model merging for LLMs seem to be the most interesting themes in 2024 so far. What are the benefits? Combining multiple models (or checkpoints) into a single one can improve training convergence, overall performance, and also robustness. I will probably do a deeper dive in the upcoming weeks, but here are at least 3 interesting papers. To get started, here's a selection of papers on this learning trajectory (no pun intended).

🔥 A special thanks to my great co-authors Markus Frohmann @markus5831 , Navid Rekabsaz @navidrekabsaz and Anne Lauscher @anne_lauscher ! 👩🏽‍💻Github: github.com/UhhDS/WhatTheW… 📃Arxiv: arxiv.org/abs/2401.12756 Looking forward to see you at the EACL!!

🧠 Which knowledge module do I need and how much of it ⁉️ 🌟 Check out our latest paper, "What the Weight?! A Unified Framework for Zero-Shot Knowledge Composition", which has been officially accepted for Findings of the EACL Conference! arxiv.org/pdf/2401.12756…

📅We're excited to present our Findings paper in a poster Session at the EACL Main Conference and additionally in an oral presentation at the MOOMIN (moomin-workshop.github.io/cfp) Workshop! @eaclmeeting

(2/2) Focusing on the scenario of domain knowledge and adapter layers, our framework provides a systematic unification of concepts, allowing us to conduct the first comprehensive benchmarking study of various zero-shot knowledge composition strategies.

(1/2)🔍In this work, we delve into the possibilities of combining knowledge by proposing a novel framework for zero-shot module composition, which encompasses existing and some novel variations for selecting, weighting, and combining parameter modules under a unified notion.