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.

Srishti Gureja, Lester James Validad Miranda, Shayekh Bin Islam, Rishabh Maheshwary, Drishti Sharma, Gusti Triandi Winata, Nathan Lambert, Sebastian Ruder, Sara Hooker, Marzieh Fadaee. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Synthetic data has driven recent state-of-the-art advancements, but reliance on a single oracle teacher model can lead to model collapse and bias propagation. These issues are particularly severe in multilingual settings, where no single model excels across all languages. In this study, we propose multilingual arbitration, which exploits performance variations among multiple models for each language. By strategically routing samples through a diverse set of models, each with unique strengths, we mitigate these challenges and enhance multilingual performance. Extensive experiments with state-of-the-art models demonstrate that our approach significantly surpasses single-teacher distillation, achieving up to 80% win rates over proprietary and open-weight models like Gemma 2, Llama 3.1, and Mistral v0.3, with the largest improvements in low-resource languages.

Shivalika Singh, Angelika Romanou, Clémentine Fourrier, David Ifeoluwa Adelani, Jian Gang Ngui, Daniel Vila-Suero, Peerat Limkonchotiwat, Kelly Marchisio, Wei Qi Leong, Yosephine Susanto, Raymond Ng, Shayne Longpre, Sebastian Ruder, Wei-Yin Ko, Antoine Bosselut, Alice Oh, Andre Martins, Leshem Choshen, Daphne Ippolito, Enzo Ferrante, Marzieh Fadaee, Beyza Ermis, Sara Hooker. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Recent advancements in large language models (LLMs) have shifted focus toward scaling inference-time compute-improving performance without retraining the model.A common approach is to sample multiple outputs in parallel, and select one of these as the final output.While existing work has focused on English and specific domains, we study how to robustly scale inference-time compute in a multilingual, multi-task setting: spanning open-ended generations, math and translation tasks, for open models at 8B and 111B scale, across seven languages.Our findings highlight the need for tailored sampling and selection strategies.We propose novel solutions tailored for this multi-faceted inference scenario, demonstrating notable gains across languages and tasks.Our methods achieve an average +6.8 jump in win-rates for 8B models on m-ArenaHard-v2.0prompts in non-English languages against proprietary models like Gemini.At larger scale, our 111B model shows a +9.0 improvement with just five samples compared to single-sample decoding.These results emphasize the importance of language-and taskaware approaches to democratize inferencetime improvements.

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.

The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (i.e., multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts. Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed.

David Ifeoluwa Adelani, Jessica Ojo, Israel Abebe Azime, Jian Yun Zhuang, Jesujoba Oluwadara Alabi, Xuanli He, Millicent Ochieng, Sara Hooker, Andiswa Bukula, En-Shiun Annie Lee, Chiamaka Ijeoma Chukwuneke, Happy Buzaaba, Blessing Kudzaishe Sibanda, Godson Koffi Kalipe, Jonathan Mukiibi, Salomon Kabongo Kabenamualu, Foutse Yuehgoh, Mmasibidi Setaka, Lolwethu Ndolela, Nkiruka Odu, Rooweither Mabuya, Salomey Osei, Shamsuddeen Hassan Muhammad, Sokhar Samb, Tadesse Kebede Guge, Tombekai Vangoni Sherman, Pontus Stenetorp. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

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.

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.

Aakanksha, Arash Ahmadian, Beyza Ermis, Seraphina Goldfarb-Tarrant, Julia Kreutzer, Marzieh Fadaee, Sara Hooker. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 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 widespread adoption of synthetic data raises new questions about how models generating the data can influence other large language models (LLMs).To start, our work exhaustively characterizes the impact of passive inheritance of model properties by systematically studying how the source of synthetic data shapes models' internal biases, calibration and preferences, and their generations' textual attributes, providing one of the most comprehensive studies to-date.We find that models are surprisingly sensitive towards certain attributes even when the synthetic data prompts appear "neutral", which invites the question "can we explicitly steer the generated data towards desired properties?"We demonstrate how such active inheritance can guide the generation profiles of models towards desirable non-differentiable attributes in both directions, e.g.increasing lexical diversity or reducing toxicity.Overall, our study broadens the understanding of the implicit biases inherited by LLMs and explores how we can leverage them to positive effect.

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.

In the machine learning ecosystem, hardware selection is often regarded as a mere utility, overshadowed by the spotlight on algorithms and data. This is especially relevant in contexts like ML-as-a-service platforms, where users often lack control over the hardware used for model deployment. This paper investigates the influence of hardware on the delicate balance between model performance and fairness. We demonstrate that hardware choices can exacerbate existing disparities, attributing these discrepancies to variations in gradient flows and loss surfaces across different demographic groups. Through both theoretical and empirical analysis, the paper not only identifies the underlying factors but also proposes an effective strategy for mitigating hardware-induced performance imbalances.

In Natural Language Processing (NLP), the Elo rating system, originally designed for ranking players in dynamic games such as chess, is increasingly being used to evaluate Large Language Models (LLMs) through "A vs B" paired comparisons. However, while popular, the system's suitability for assessing entities with constant skill levels, such as LLMs, remains relatively unexplored. We study two fundamental axioms that evaluation methods should adhere to: reliability and transitivity. We conduct an extensive evaluation of Elo behavior across simulated and real-world scenarios, demonstrating that individual Elo computations can exhibit significant volatility. We show that both axioms are not always satisfied, raising questions about the reliability of current comparative evaluations of LLMs. If the current use of Elo scores is intended to substitute the costly head-to-head comparison of LLMs, it is crucial to ensure the ranking is as robust as possible. Guided by the axioms, our findings offer concrete guidelines for enhancing the reliability of LLM evaluation methods, suggesting a need for reassessment of existing comparative approaches.

General-purpose artificial intelligence (AI) systems are built on massive swathes of public web data, assembled into corpora such as C4, RefinedWeb, and Dolma. To our knowledge, we conduct the first, large-scale, longitudinal audit of the consent protocols for the web domains underlying AI training corpora. Our audit of 14,000 web domains provides an expansive view of crawlable web data and how codified data use preferences are changing over time. We observe a proliferation of AI-specific clauses to limit use, acute differences in restrictions on AI developers, as well as general inconsistencies between websites' expressed intentions in their Terms of Service and their robots.txt. We diagnose these as symptoms of ineffective web protocols, not designed to cope with the widespread re-purposing of the internet for AI. Our longitudinal analyses show that in a single year (2023-2024) there has been a rapid crescendo of data restrictions from web sources, rendering ~5\%+ of all tokens in C4, or 28%+ of the most actively maintained, critical sources in C4, fully restricted from use. For Terms of Service crawling restrictions, a full 45% of C4 is now restricted. If respected or enforced, these restrictions are rapidly biasing the diversity, freshness, and scaling laws for general-purpose AI systems. We hope to illustrate the emerging crises in data consent, for both developers and creators. The foreclosure of much of the open web will impact not only commercial AI, but also non-commercial AI and academic research.

Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds.Similarly, black-box commercially available APIs for detecting toxicity, such as the Perspective API, are not static, but frequently retrained to address any unattended weaknesses and biases.We evaluate the implications of these changes on the reproducibility of findings that compare the relative merits of models and methods that aim to curb toxicity.Our findings suggest that research that relied on inherited automatic toxicity scores to compare models and techniques may have resulted in inaccurate findings.Rescoring all models from HELM, a widely respected living benchmark, for toxicity with the recent version of the API led to a different ranking of widely used foundation models.We suggest caution in applying apples-to-apples comparisons between studies and lay recommendations for a more structured approach to evaluating toxicity over time. 1

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.

Pushing the boundaries of machine learning often requires exploring different hardware and software combinations. However, this ability to experiment with different systems can be at odds with the drive for efficiency, which has produced increasingly specialized AI hardware and incentivized consolidation around a narrow set of ML frameworks. Exploratory research can be further restricted if software and hardware are co-evolving, making it even harder to stray away from a given tooling stack. While this friction increasingly impacts the rate of innovation in machine learning, to our knowledge the lack of portability in tooling has not been quantified. In this work we ask: How portable are popular ML software frameworks? We conduct a large scale study of the portability of mainstream ML frameworks across different hardware types. Our findings paint an uncomfortable picture -- frameworks can lose more than 40% of their key functions when ported to other hardware. Worse, even when functions are portable, the slowdown in their performance can be extreme. Collectively, our results reveal how costly straying from a narrow set of hardware-software combinations can be - and thus how specialization incurs an exploration cost that can impede innovation in machine learning research.

Despite widespread use of LLMs as conversational agents, evaluations of performance fail to capture a crucial aspect of communication: interpreting language in context---incorporating its pragmatics. Humans interpret language using beliefs and prior knowledge about the world. For example, we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No". To investigate whether LLMs have the ability to make this type of inference, known as an implicature, we design a simple task and evaluate four categories of widely used state-of-the-art models. We find that, despite only evaluating on utterances that require a binary inference (yes or no), models in three of these categories perform close to random. However, LLMs instruction-tuned at the example-level perform significantly better. These results suggest that certain fine-tuning strategies are far better at inducing pragmatic understanding in models. We present our findings as the starting point for further research into evaluating how LLMs interpret language in context and to drive the development of more pragmatic and useful models of human discourse.

Knowledge distillation is a popular technique that has been shown to produce remarkable gains in average accuracy. However, recent work has shown that these gains are not uniform across subgroups in the data, and can often come at the cost of accuracy on rare subgroups and classes. Robust optimization is a common remedy to improve worst-class accuracy in standard learning settings, but in distillation it is unknown whether it is best to apply robust objectives when training the teacher, the student, or both. This work studies the interplay between robust objectives for the teacher and student. Empirically, we show that that jointly modifying the teacher and student objectives can lead to better worst-class student performance and even Pareto improvement in the trade-off between worst-class and overall performance. Theoretically, we show that the per-class calibration of teacher scores is key when training a robust student. Both the theory and experiments support the surprising finding that applying a robust teacher training objective does not always yield a more robust student.

Abstract Recent work in natural language processing (NLP) has yielded appealing results from scaling model parameters and training data; however, using only scale to improve performance means that resource consumption also grows. Such resources include data, time, storage, or energy, all of which are naturally limited and unevenly distributed. This motivates research into efficient methods that require fewer resources to achieve similar results. This survey synthesizes and relates current methods and findings in efficient NLP. We aim to provide both guidance for conducting NLP under limited resources, and point towards promising research directions for developing more efficient methods.

In machine learning, a question of great interest is understanding what examples are challenging for a model to classify. Identifying atypical examples ensures the safe de-ployment of models, isolates samples that require further human inspection and provides interpretability into model behavior. In this work, we propose Variance of Gradients (VoG) as a valuable and efficient metric to rank data by difficulty and to surface a tractable subset of the most chal-lenging examples for human-in-the-loop auditing. We show that data points with high VoG scores are far more difficult for the model to learn and over-index on corrupted or mem-orized examples. Further, restricting the evaluation to the test set instances with the lowest VoG improves the model's generalization performance. Finally, we show that VoG is a valuable and efficient ranking for out-of-distribution detection.

Multilingual models are often particularly dependent on scaling to generalize to a growing number of languages.Compression techniques are widely relied upon to reconcile the growth in model size with real world resource constraints, but compression can have a disparate effect on model performance for lowresource languages.It is thus crucial to understand the trade-offs between scale, multilingualism, and compression.In this work, we propose an experimental framework to characterize the impact of sparsifying multilingual pre-trained language models during finetuning.Applying this framework to mBERT named entity recognition models across 40 languages, we find that compression confers several intriguing and previously unknown generalization properties.In contrast to prior findings, we find that compression may improve model robustness over dense models.We additionally observe that under certain sparsification regimes compression may aid, rather than disproportionately impact the performance of low-resource languages.

We propose an empirical measure of the approximate accuracy of feature importance estimates in deep neural networks. Our results across several large-scale image classification datasets show that many popular interpretability methods produce estimates of feature importance that are not better than a random designation of feature importance. Only certain ensemble based approaches---VarGrad and SmoothGrad-Squared---outperform such a random assignment of importance. The manner of ensembling remains critical, we show that some approaches do no better then the underlying method but carry a far higher computational burden.