Papers by Ahmet Üstün
When Personalization Meets Reality: A Multi-Faceted Analysis of Personalized Preference Learning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback assumes homogeneous preferences across users . personalization can introduce up to 20% safety misalignment . |
| Approach: | They propose a framework to assess personalized preference learning by tailoring preferences for users . they compare eight personalization methods across three preference datasets . |
| Outcome: | The proposed framework measures performance, fairness, unintended effects, adaptability across preferences . performance differences between personalization methods could reach 36% when users strongly disagree . |
RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Preference optimization is a widely adopted post-training technique to align large language models with human preferences. |
| Approach: | They propose a method for generating multilingual feedback data to balance data coverage. |
| Outcome: | The proposed method achieves 54.4% win-rate against current state-of-the-art multilingual LLM in its parameter class and 69.5% win- rate or higher against widely used models like Gemma, Mistral and Llama 3. |
Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs (2024.acl-long)
Copied to clipboard
Arash Ahmadian, Chris Cremer, Matthias Gallé, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, Sara Hooker
| Challenge: | Proximal Policy Optimization (PPO) is used for RLHF but requires high computational cost and sensitive hyperparameter tuning. |
| Approach: | They propose to use Proximal Policy Optimization to align large language models to human preferences. |
| Outcome: | The proposed method preserves and even increases performance while preserving the motivational principles that led to the development of PPO. |
Multilingual Unsupervised Neural Machine Translation with Denoising Adapters (2021.emnlp-main)
Copied to clipboard
| Challenge: | Multilingual unsupervised machine translation is a computationally expensive and hard to tune approach . auxiliary parallel data is used to train translation systems from monolingual data . |
| Approach: | They propose to use auxiliary parallel language pairs to train unsupervised machine translations . they propose to add auxiliary languages to pre-trained mBART-50 models with denoising adapters . |
| Outcome: | The proposed approach is on-par with back-translation and allows adding unseen languages incrementally. |
Hyper-X: A Unified Hypernetwork for Multi-Task Multilingual Transfer (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing multilingual models cannot fully leverage training data when it is available in different task-language combinations. |
| Approach: | They propose a single hypernetwork that unifies multi-task and multilingual learning with efficient adaptation. |
| Outcome: | The proposed model achieves the best or competitive gain when a mixture of multiple resources is available while being significantly more efficient than existing models. |
When does Parameter-Efficient Transfer Learning Work for Machine Translation? (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prior work indicates that parametric fine-tuning methods may not work as well for machine translation (MT). |
| Approach: | They propose to use parameter-efficient fine-tuning methods to adapt large pre-trained models while only tuning a small number of parameters. |
| Outcome: | The proposed methods outperform full fine-tuning for many downstream tasks when the parameter budget corresponds to 10% of the model parameters. |
Massive Choice, Ample Tasks (MaChAmp): A Toolkit for Multi-task Learning in NLP (2021.eacl-demos)
Copied to clipboard
| Challenge: | Multi-task learning (MTL) has become a standard repertoire in natural language processing (NLP) it enables neural networks to learn tasks in parallel while leveraging the benefits of sharing parameters. |
| Approach: | They propose a toolkit for fine-tuning contextualized embeddings in multi-task settings. |
| Outcome: | The proposed toolkit supports a variety of natural language processing tasks . it enables neural networks to learn tasks in parallel while leveraging the benefits of sharing parameters. |
Nexus: Adaptive Upcycling to Efficiently Pretrain Mixture of Experts (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Nevertheless, training from scratch on trillions of tokens remains expensive that most users can only finetune these models. |
| Approach: | They propose to reuse parameters of dense models for the MoE layers with a router that can integrate new experts into existing trained models without hurting performance on previous domains. |
| Outcome: | The proposed router can integrate new experts into existing trained models without hurting the performance on previous domains. |
How Does Quantization Affect Multilingual LLMs? (2024.findings-emnlp)
Copied to clipboard
Kelly Marchisio, Saurabh Dash, Hongyu Chen, Dennis Aumiller, Ahmet Üstün, Sara Hooker, Sebastian Ruder
| Challenge: | Quantization is widely used to improve inference speed and deployment of large language models. |
| Approach: | They conduct a thorough analysis of quantized multilingual LLMs . they find language disparately affected by quantization, non-Latin script languages worst . authors urge consideration of multilingual performance as evaluation criterion for efficient models . |
| Outcome: | The results show that quantization has harmful effects on human evaluation . language performance is disparately affected by quantization, the authors say . |
Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning (2024.acl-long)
Copied to clipboard
Shivalika Singh, Freddie Vargus, Daniel D’souza, Börje Karlsson, Abinaya Mahendiran, Wei-Yin Ko, Herumb Shandilya, Jay Patel, Deividas Mataciunas, Laura O’Mahony, Mike Zhang, Ramith Hettiarachchi, Joseph Wilson, Marina Machado, Luisa Moura, Dominik Krzemiński, Hakimeh Fadaei, Irem Ergun, Ifeoma Okoh, Aisha Alaagib, Oshan Mudannayake, Zaid Alyafeai, Vu Chien, Sebastian Ruder, Surya Guthikonda, Emad Alghamdi, Sebastian Gehrmann, Niklas Muennighoff, Max Bartolo, Julia Kreutzer, Ahmet Üstün, Marzieh Fadaee, Sara Hooker
| Challenge: | Existing datasets in the English language are mostly in the realm of instruction fine-tuning . aya dataset, the Aya Collection, and the AYa Evaluation Suite are key resources . |
| Approach: | They aim to build a human-curated instruction-following dataset spanning 65 languages . they work with fluent speakers of languages from around the world to collect natural instances of instructions and completions . |
| Outcome: | The goal is to build a human-curated instruction-following dataset spanning 65 languages. |
On the Difficulty of Translating Free-Order Case-Marking Languages (2021.tacl-1)
Copied to clipboard
| Challenge: | Free-order case-marking languages are more difficult to model than fixed-order languages . however, in medium- and low-resource settings, the overall NMT quality of fixed-or-fixed-order language pairs remains unmatched. |
| Approach: | They investigate whether certain languages are harder to model than others by adding case marking to their models. |
| Outcome: | The proposed models show that word order flexibility in the source language leads to a small loss of NMT quality even though the core verb arguments become impossible to disambiguate in sentences without semantic cues. |
One Tokenizer To Rule Them All: Emergent Language Plasticity via Multilingual Tokenizers (2026.acl-long)
Copied to clipboard
Diana Abagyan, Alejandro R. Salamanca, Andres Felipe Cruz-Salinas, Kris Cao, Hangyu Lin, Acyr Locatelli, Marzieh Fadaee, Ahmet Üstün, Sara Hooker
| Challenge: | Existing approaches to train multilingual large language models for many languages at once are limited due to limited model capacity, scarce high-quality data, and compute constraints. |
| Approach: | They propose to use a universal tokenizer to improve language plasticity and adaptability to new languages by up to 20%. |
| Outcome: | The proposed tokenizer improves language plasticity and improves plasticity towards languages that are completely unseen in the tokenizer and pretraining, by up to 5% win rate gain. |
From Masked Language Modeling to Translation: Non-English Auxiliary Tasks Improve Zero-shot Spoken Language Understanding (2021.naacl-main)
Copied to clipboard
Rob van der Goot, Ibrahim Sharaf, Aizhan Imankulova, Ahmet Üstün, Marija Stepanović, Alan Ramponi, Siti Oryza Khairunnisa, Mamoru Komachi, Barbara Plank
| Challenge: | Lack of publicly available evaluation data for low-resource languages limits progress in SLU . despite advances in neural modeling for slot and intent detection, datasets for SLU remain limited. |
| Approach: | They propose a joint learning approach with English SLU training data and non-English auxiliary tasks from raw text, syntax and translation for transfer. |
| Outcome: | The proposed model can learn English SLU training data and non-English auxiliary tasks from raw text, syntax and translation for transfer. |
UDapter: Language Adaptation for Truly Universal Dependency Parsing (2020.emnlp-main)
Copied to clipboard
| Challenge: | Cross-language interference and restrained model capacity remain major obstacles in multilingual dependency parsing. |
| Approach: | They propose a multilingual task adaptation approach based on contextual parameter generation and adapter modules that learn adapters via language embeddings while sharing model parameters across languages. |
| Outcome: | The proposed approach outperforms strong monolingual and multilingual baselines on most languages on high-resource and low-resourced (zero-shot) languages. |