Diana Abagyan, Alejandro R. Salamanca, Andres Felipe Cruz-Salinas, Kris Cao, Hangyu Lin, Acyr Locatelli, Marzieh Fadaee, Ahmet Üstün, Sara Hooker. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Including code in the pre-training data mixture, even for models not specifically designed for code, has become a common practice in LLMs pre-training. While there has been anecdotal consensus among practitioners that code data plays a vital role in general LLMs' performance, there is only limited work analyzing the precise impact of code on non-code tasks. In this work, we systematically investigate the impact of code data on general performance. We ask “what is the impact of code data used in pre-training on a large variety of downstream tasks beyond code generation”. We conduct extensive ablations and evaluate across a broad range of natural language reasoning tasks, world knowledge tasks, code benchmarks, and LLM-as-a-judge win-rates for models with sizes ranging from 470M to 2.8B parameters. Across settings, we find a consistent results that code is a critical building block for generalization far beyond coding tasks and improvements to code quality have an outsized impact across all tasks. In particular, compared to text-only pre-training, the addition of code results in up to relative increase of 8.2% in natural language (NL) reasoning, 4.2% in world knowledge, 6.6% improvement in generative win-rates, and a 12x boost in code performance respectively. Our work suggests investments in code quality and preserving code during pre-training have positive impacts.

One of the most profound challenges of modern machine learning is performing well on the long-tail of rare and underrepresented features. Large general-purpose models are trained for many tasks, but work best on high-frequency use cases. After training, it is hard to adapt a model to perform well on specific use cases underrepresented in the training corpus. Relying on prompt engineering or few-shot examples to maximize the output quality on a particular test case can be frustrating, as models can be highly sensitive to small changes, react in unpredicted ways or rely on a fixed system prompt for maintaining performance. In this work, we ask: Can we optimize our training protocols to both improve controllability and performance on underrepresented use cases at inference time? We revisit the divide between training and inference techniques to improve long-tail performance while providing users with a set of control levers the model is trained to be responsive to. We create a detailed taxonomy of data characteristics and task provenance to explicitly control generation attributes and implicitly condition generations at inference time. We fine-tune a base model to infer these markers automatically, which makes them optional at inference time. This principled and flexible approach yields pronounced improvements in performance on examples from the long tail of the training distribution. Overall, we observe lifts of 5.7% across all tasks. However, treasure markers are particularly effective at finding difficult to obtain gains in the long-tail. We observe relative lifts of up to 14.1% on underrepresented tasks like CodeRepair and absolute improvements of 35.3% on length instruction following evaluations.

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion.Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we found one provider testing 27 private variants before making one model public at the second position on the leaderboard. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. The top two providers have individually received an estimated 19.2% and 20.4% of all data on the arena. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. With conservative estimates, we show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on ArenaHard, a test set from the arena distribution.Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field.

Abstract Instruction tuning enhances large language models (LLMs) by aligning them with human preferences across diverse tasks. Traditional approaches to create instruction tuning datasets face serious challenges for low-resource languages due to their dependence on data annotation. This work introduces a novel method, Multilingual Reverse Instructions (MURI), which generates high-quality instruction tuning datasets for low-resource languages without requiring human annotators or pre-existing multilingual models. Utilizing reverse instructions and a translation pipeline, MURI produces instruction-output pairs from existing human-written texts in low-resource languages. This method ensures cultural relevance and diversity by sourcing texts from different native domains and applying filters to eliminate inappropriate content. Our dataset, MURI-IT, includes more than 2 million instruction-output pairs across 200 languages. Evaluation by native speakers and fine-tuning experiments with mT5 models demonstrate the approach’s effectiveness for both NLU and open-ended generation. We publicly release datasets and models at https://github.com/akoksal/muri.

Arash Ahmadian, Chris Cremer, Matthias Gallé, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, Sara Hooker. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Shivalika Singh, Freddie Vargus, Daniel D’souza, Börje F. 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. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Ahmet Üstün, Viraat Aryabumi, Zheng Yong, Wei-Yin Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Freddie Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzieh Fadaee, Julia Kreutzer, Sara Hooker. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Preference optimization techniques have become a standard final stage for training state-ofart large language models (LLMs).However, despite widespread adoption, the vast majority of work to-date has focused on a small set of high-resource languages like English and Chinese.This captures a small fraction of the languages in the world, but also makes it unclear which aspects of current state-of-the-art research transfer to a multilingual setting.In this work, we perform an exhaustive study to achieve a new state of the art in aligning multilingual LLMs.We introduce a novel, scalable method for generating high-quality multilingual feedback data to balance data coverage.We establish the benefits of cross-lingual transfer and increased dataset size in preference training.Our preference-trained model achieves a 54.4% win-rate against Aya 23 8B, the current state-of-the-art multilingual LLM in its parameter class, and a 69.5% win-rate or higher against widely used models like Gemma, Mistral and Llama 3. As a result of our efforts, we expand the frontier of alignment techniques to 23 languages, covering approximately half of the world's population.

The Mixture of Experts (MoE) is a widely known neural architecture where an ensemble of specialized sub-models optimizes overall performance with a constant computational cost. However, conventional MoEs pose challenges at scale due to the need to store all experts in memory. In this paper, we push MoE to the limit. We propose extremely parameter-efficient MoE by uniquely combining MoE architecture with lightweight experts.Our MoE architecture outperforms standard parameter-efficient fine-tuning (PEFT) methods and is on par with full fine-tuning by only updating the lightweight experts -- less than 1\% of an 11B parameters model. Furthermore, our method generalizes to unseen tasks as it does not depend on any prior task knowledge. Our research underscores the versatility of the mixture of experts architecture, showcasing its ability to deliver robust performance even when subjected to rigorous parameter constraints.

Mixture of Experts (MoE) framework has become a popular architecture for large language models due to its superior performance compared to dense models. However, training MoEs from scratch in a large-scale regime is prohibitively expensive. Previous work addresses this challenge by independently training multiple dense expert models and using them to initialize an MoE. In particular, state-of-the-art approaches initialize MoE layers using experts' feed-forward parameters while merging all other parameters, limiting the advantages of the specialized dense models when upcycling them as MoEs. We propose BAM (Branch-Attend-Mix), a simple yet effective improvement to MoE training. BAM makes full use of specialized dense models by not only using their feed-forward network (FFN) to initialize the MoE layers but also leveraging experts' attention weights fully by leveraging them as mixture-of-attention (MoA) layers. We explore two methods for upcycling MoA layers: 1) initializing separate attention experts from dense models including key, value, and query matrices; and 2) initializing only Q projections while sharing key-value pairs across all experts to facilitate efficient inference. Our experiments using seed models ranging from 590 million to 2 billion parameters show that our approach outperforms state-of-the-art approaches under the same data and compute budget in both perplexity and downstream tasks evaluations, confirming the effectiveness of BAM.

Emergent properties have been widely adopted as a term to describe behavior not present in smaller models but observed in larger models (Wei et al., 2022a). Recent work suggests that the trade-off incurred by quantization is also an emergent property, with sharp drops in performance in models over 6B parameters. In this work, we ask _are quantization cliffs in performance solely a factor of scale?_ Against a backdrop of increased research focus on why certain emergent properties surface at scale, this work provides a useful counter-example. We posit that it is possible to optimize for a quantization friendly training recipe that suppresses large activation magnitude outliers. Here, we find that outlier dimensions are not an inherent product of scale, but rather sensitive to the optimization conditions present during pre-training. This both opens up directions for more efficient quantization, and poses the question of whether other emergent properties are inherent or can be altered and conditioned by optimization and architecture design choices. We successfully quantize models ranging in size from 410M to 52B with minimal degradation in performance.

Massively multilingual models are promising for transfer learning across tasks and languages.However, existing methods are unable to fully leverage training data when it is available in different task-language combinations.To exploit such heterogeneous supervision, we propose Hyper-X, a single hypernetwork that unifies multi-task and multilingual learning with efficient adaptation.This model generates weights for adapter modules conditioned on both tasks and language embeddings.By learning to combine task and language-specific knowledge, our model enables zero-shot transfer for unseen languages and task-language combinations.Our experiments on a diverse set of languages demonstrate that Hyper-X achieves the best or competitive gain when a mixture of multiple resources is available, while being on par with strong baselines in the standard scenario.Hyper-X is also considerably more efficient in terms of parameters and resources compared to methods that train separate adapters.Finally, Hyper-X consistently produces strong results in few-shot scenarios for new languages, showing the versatility of our approach beyond zero-shot transfer.1 NER en Pre-trained Model Pre-trained Model Fine-tuned Model ar tr Fine-tuned Model ar tr POS en Single-Task Pre-trained Model Fine-tuned Model ar tr ar tr NER POS en en Multi-Task Pre-trained Model Fine-tuned Model tr NER POS ar tr NER ar POS en en Mixed-Language Multi-Task Anne Lauscher, Vinit Ravishankar, Ivan Vulić, and Goran Glavaš.2020a.From Zero to Hero: On the Limitations of Zero-Shot

Parameter-efficient fine-tuning methods (PEFTs) offer the promise of adapting large pre-trained models while only tuning a small number of parameters.They have been shown to be competitive with full model fine-tuning for many downstream tasks.However, prior work indicates that PEFTs may not work as well for machine translation (MT), and there is no comprehensive study showing when PEFTs work for MT.We conduct a comprehensive empirical study of PEFTs for MT, considering (1) various parameter budgets, (2) a diverse set of language-pairs, and (3) different pre-trained models.We find that 'adapters', in which small feed-forward networks are added after every layer, are indeed on par with full model fine-tuning when the parameter budget corresponds to 10% of total model parameters.Nevertheless, as the number of tuned parameters decreases, the performance of PEFTs decreases.The magnitude of this decrease depends on the language pair, with PEFTs particularly struggling for distantly related language-pairs.We find that using PEFTs with a larger pre-trained model outperforms full fine-tuning with a smaller model, and for smaller training data sizes, PEFTs outperform full fine-tuning for the same pre-trained model. 1

Abstract Recent advances in multilingual language modeling have brought the idea of a truly universal parser closer to reality. However, such models are still not immune to the “curse of multilinguality”: Cross-language interference and restrained model capacity remain major obstacles. To address this, we propose a novel language adaptation approach by introducing contextual language adapters to a multilingual parser. Contextual language adapters make it possible to learn adapters via language embeddings while sharing model parameters across languages based on contextual parameter generation. Moreover, our method allows for an easy but effective integration of existing linguistic typology features into the parsing model. Because not all typological features are available for every language, we further combine typological feature prediction with parsing in a multi-task model that achieves very competitive parsing performance without the need for an external prediction system for missing features. The resulting parser, UDapter, can be used for dependency parsing as well as sequence labeling tasks such as POS tagging, morphological tagging, and NER. In dependency parsing, it outperforms strong monolingual and multilingual baselines on the majority of both high-resource and low-resource (zero-shot) languages, showing the success of the proposed adaptation approach. In sequence labeling tasks, our parser surpasses the baseline on high resource languages, and performs very competitively in a zero-shot setting. Our in-depth analyses show that adapter generation via typological features of languages is key to this success.1

We consider the problem of multilingual unsupervised machine translation, translating to and from languages that only have monolingual data by using auxiliary parallel language pairs.For this problem the standard procedure so far to leverage the monolingual data is back-translation, which is computationally costly and hard to tune.In this paper we propose instead to use denoising adapters, adapter layers with a denoising objective, on top of pre-trained mBART-50.In addition to the modularity and flexibility of such an approach we show that the resulting translations are on-par with back-translating as measured by BLEU, and furthermore it allows adding unseen languages incrementally.

Rob van der Goot, Ibrahim Sharaf, Aizhan Imankulova, Ahmet Üstün, Marija Stepanović, Alan Ramponi, Siti Oryza Khairunnisa, Mamoru Komachi, Barbara Plank. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Abstract We investigate the usage of semantic information for morphological segmentation since words that are derived from each other will remain semantically related. We use mathematical models such as maximum likelihood estimate (MLE) and maximum a posteriori estimate (MAP) by incorporating semantic information obtained from dense word vector representations. Our approach does not require any annotated data which make it fully unsupervised and require only a small amount of raw data together with pretrained word embeddings for training purposes. The results show that using dense vector representations helps in morphological segmentation especially for low-resource languages. We present results for Turkish, English, and German. Our semantic MLE model outperforms other unsupervised models for Turkish language. Our proposed models could be also used for any other low-resource language with concatenative morphology.

Abstract Identifying factors that make certain languages harder to model than others is essential to reach language equality in future Natural Language Processing technologies. Free-order case-marking languages, such as Russian, Latin, or Tamil, have proved more challenging than fixed-order languages for the tasks of syntactic parsing and subject-verb agreement prediction. In this work, we investigate whether this class of languages is also more difficult to translate by state-of-the-art Neural Machine Translation (NMT) models. Using a variety of synthetic languages and a newly introduced translation challenge set, we find that word order flexibility in the source language only leads to a very small loss of NMT quality, even though the core verb arguments become impossible to disambiguate in sentences without semantic cues. The latter issue is indeed solved by the addition of case marking. However, in medium- and low-resource settings, the overall NMT quality of fixed-order languages remains unmatched.

Recent advances in multilingual dependency parsing have brought the idea of a truly universal parser closer to reality.However, crosslanguage interference and restrained model capacity remain major obstacles.To address this, we propose a novel multilingual task adaptation approach based on contextual parameter generation and adapter modules.This approach enables to learn adapters via language embeddings while sharing model parameters across languages.It also allows for an easy but effective integration of existing linguistic typology features into the parsing network.The resulting parser, UDapter, outperforms strong monolingual and multilingual baselines on the majority of both high-resource and lowresource (zero-shot) languages, showing the success of the proposed adaptation approach.Our in-depth analyses show that soft parameter sharing via typological features is key to this success.1