Beyza Ermiş

Luiza Pozzobon, Beyza Ermis, Patrick Lewis et al. · meta-ai

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. Code and data are available at https://github.com/for-ai/black-box-api-challenges.

Meriem Boubdir, Edward Kim, Beyza Ermis et al. · nvidia

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 extensive evaluation of Elo behaviour, illustrating that individual Elo computations exhibit volatility and delving into the impact of varying the Elo rating system's hyperparameters. We show that these 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.

Ted Zadouri, Ahmet Üstün, Arash Ahmadian et al.

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. Our code used in all the experiments is publicly available here: https://github.com/for-ai/parameter-efficient-moe.

Meriem Boubdir, Edward Kim, Beyza Ermis et al. · nvidia

Human evaluation is increasingly critical for assessing large language models, capturing linguistic nuances, and reflecting user preferences more accurately than traditional automated metrics. However, the resource-intensive nature of this type of annotation process poses significant challenges. The key question driving our work: "is it feasible to minimize human-in-the-loop feedback by prioritizing data instances which most effectively distinguish between models?" We evaluate several metric-based methods and find that these metrics enhance the efficiency of human evaluations by minimizing the number of required annotations, thus saving time and cost, while ensuring a robust performance evaluation. We show that our method is effective across widely used model families, reducing instances of indecisive (or "tie") outcomes by up to 54% compared to a random sample when focusing on the top-20 percentile of prioritized instances. This potential reduction in required human effort positions our approach as a valuable strategy in future large language model evaluations.

Luiza Pozzobon, Beyza Ermis, Patrick Lewis et al.

Considerable effort has been dedicated to mitigating toxicity, but existing methods often require drastic modifications to model parameters or the use of computationally intensive auxiliary models. Furthermore, previous approaches have often neglected the crucial factor of language's evolving nature over time. In this work, we present a comprehensive perspective on toxicity mitigation that takes into account its changing nature. We introduce Goodtriever, a flexible methodology that matches the current state-of-the-art toxicity mitigation while achieving 43% relative latency reduction during inference and being more computationally efficient. By incorporating a retrieval-based approach at decoding time, Goodtriever enables toxicity-controlled text generation. Our research advocates for an increased focus on adaptable mitigation techniques, which better reflect the data drift models face when deployed in the wild. Code and data are available at https://github.com/for-ai/goodtriever.

Beyza Ermis, Y. Kenan Yılmaz, A. Taylan Cemgil et al. · deepmind

Probabilistic Latent Tensor Factorization (PLTF) is a recently proposed probabilistic framework for modelling multi-way data. Not only the common tensor factorization models but also any arbitrary tensor factorization structure can be realized by the PLTF framework. This paper presents full Bayesian inference via variational Bayes that facilitates more powerful modelling and allows more sophisticated inference on the PLTF framework. We illustrate our approach on model order selection and link prediction.

Beyza Ermis, Giovanni Zappella, Martin Wistuba et al.

In many real-world scenarios, data to train machine learning models becomes available over time. Unfortunately, these models struggle to continually learn new concepts without forgetting what has been learnt in the past. This phenomenon is known as catastrophic forgetting and it is difficult to prevent due to practical constraints. For instance, the amount of data that can be stored or the computational resources that can be used might be limited. Moreover, applications increasingly rely on large pre-trained neural networks, such as pre-trained Transformers, since the resources or data might not be available in sufficiently large quantities to practitioners to train the model from scratch. In this paper, we devise a method to incrementally train a model on a sequence of tasks using pre-trained Transformers and extending them with Adapters. Different than the existing approaches, our method is able to scale to a large number of tasks without significant overhead and allows sharing information across tasks. On both image and text classification tasks, we empirically demonstrate that our method maintains a good predictive performance without retraining the model or increasing the number of model parameters over time. The resulting model is also significantly faster at inference time compared to Adapter-based state-of-the-art methods.

Beyza Ermis, Giovanni Zappella, Martin Wistuba et al.

In many real-world scenarios, data to train machine learning models become available over time. However, neural network models struggle to continually learn new concepts without forgetting what has been learnt in the past. This phenomenon is known as catastrophic forgetting and it is often difficult to prevent due to practical constraints, such as the amount of data that can be stored or the limited computation sources that can be used. Moreover, training large neural networks, such as Transformers, from scratch is very costly and requires a vast amount of training data, which might not be available in the application domain of interest. A recent trend indicates that dynamic architectures based on an expansion of the parameters can reduce catastrophic forgetting efficiently in continual learning, but this needs complex tuning to balance the growing number of parameters and barely share any information across tasks. As a result, they struggle to scale to a large number of tasks without significant overhead. In this paper, we validate in the computer vision domain a recent solution called Adaptive Distillation of Adapters (ADA), which is developed to perform continual learning using pre-trained Transformers and Adapters on text classification tasks. We empirically demonstrate on different classification tasks that this method maintains a good predictive performance without retraining the model or increasing the number of model parameters over the time. Besides it is significantly faster at inference time compared to the state-of-the-art methods.

Ondrej Bohdal, Lukas Balles, Martin Wistuba et al.

Hyperparameter optimization (HPO) and neural architecture search (NAS) are methods of choice to obtain the best-in-class machine learning models, but in practice they can be costly to run. When models are trained on large datasets, tuning them with HPO or NAS rapidly becomes prohibitively expensive for practitioners, even when efficient multi-fidelity methods are employed. We propose an approach to tackle the challenge of tuning machine learning models trained on large datasets with limited computational resources. Our approach, named PASHA, extends ASHA and is able to dynamically allocate maximum resources for the tuning procedure depending on the need. The experimental comparison shows that PASHA identifies well-performing hyperparameter configurations and architectures while consuming significantly fewer computational resources than ASHA.

Luiza Pozzobon, Patrick Lewis, Sara Hooker et al.

To date, toxicity mitigation in language models has almost entirely been focused on single-language settings. As language models embrace multilingual capabilities, it's crucial our safety measures keep pace. Recognizing this research gap, our approach expands the scope of conventional toxicity mitigation to address the complexities presented by multiple languages. In the absence of sufficient annotated datasets across languages, we employ translated data to evaluate and enhance our mitigation techniques. We also compare finetuning mitigation approaches against retrieval-augmented techniques under both static and continual toxicity mitigation scenarios. This allows us to examine the effects of translation quality and the cross-lingual transfer on toxicity mitigation. We also explore how model size and data quantity affect the success of these mitigation efforts. Covering nine languages, our study represents a broad array of linguistic families and levels of resource availability, ranging from high to mid-resource languages. Through comprehensive experiments, we provide insights into the complexities of multilingual toxicity mitigation, offering valuable insights and paving the way for future research in this increasingly important field. Code and data are available at https://github.com/for-ai/goodtriever.

Shivalika Singh, Angelika Romanou, Clémentine Fourrier et al.

Cultural biases in multilingual datasets pose significant challenges for their effectiveness as global benchmarks. These biases stem not only from differences in language but also from the cultural knowledge required to interpret questions, reducing the practical utility of translated datasets like MMLU. Furthermore, translation often introduces artefacts that can distort the meaning or clarity of questions in the target language. A common practice in multilingual evaluation is to rely on machine-translated evaluation sets, but simply translating a dataset is insufficient to address these challenges. In this work, we trace the impact of both of these issues on multilingual evaluations and ensuing model performances. Our large-scale evaluation of state-of-the-art open and proprietary models illustrates that progress on MMLU depends heavily on learning Western-centric concepts, with 28% of all questions requiring culturally sensitive knowledge. Moreover, for questions requiring geographic knowledge, an astounding 84.9% focus on either North American or European regions. Rankings of model evaluations change depending on whether they are evaluated on the full portion or the subset of questions annotated as culturally sensitive, showing the distortion to model rankings when blindly relying on translated MMLU. We release Global MMLU, an improved MMLU with evaluation coverage across 42 languages -- with improved overall quality by engaging with compensated professional and community annotators to verify translation quality while also rigorously evaluating cultural biases present in the original dataset. This comprehensive Global MMLU set also includes designated subsets labeled as culturally sensitive and culturally agnostic to allow for more holistic, complete evaluation.

John Dang, Shivalika Singh, Daniel D'souza et al.

We introduce the Aya Expanse model family, a new generation of 8B and 32B parameter multilingual language models, aiming to address the critical challenge of developing highly performant multilingual models that match or surpass the capabilities of monolingual models. By leveraging several years of research at Cohere For AI and Cohere, including advancements in data arbitrage, multilingual preference training, and model merging, Aya Expanse sets a new state-of-the-art in multilingual performance. Our evaluations on the Arena-Hard-Auto dataset, translated into 23 languages, demonstrate that Aya Expanse 8B and 32B outperform leading open-weight models in their respective parameter classes, including Gemma 2, Qwen 2.5, and Llama 3.1, achieving up to a 76.6% win-rate. Notably, Aya Expanse 32B outperforms Llama 3.1 70B, a model with twice as many parameters, achieving a 54.0% win-rate. In this short technical report, we present extended evaluation results for the Aya Expanse model family and release their open-weights, together with a new multilingual evaluation dataset m-ArenaHard.

Çağatay Yıldız, Nishaanth Kanna Ravichandran, Nitin Sharma et al.

Continual learning (CL) in large language models (LLMs) is an evolving domain that focuses on developing efficient and sustainable training strategies to adapt models to emerging knowledge and achieve robustness in dynamic environments. Our primary emphasis is on continual domain-adaptive pretraining, a process designed to equip LLMs with the ability to integrate new information from various domains while retaining previously learned knowledge. Since existing works concentrate mostly on continual fine-tuning for a limited selection of downstream tasks or training domains, we introduce a new benchmark designed to measure the adaptability of LLMs to changing pretraining data landscapes. We further examine the impact of model size on learning efficacy and forgetting, as well as how the progression and similarity of emerging domains affect the knowledge transfer within these models. Our findings uncover several key insights: (i) continual pretraining consistently improves <1.5B models studied in this work and is also superior to domain adaptation, (ii) larger models always achieve better perplexity than smaller ones when continually pretrained on the same corpus, (iii) smaller models are particularly sensitive to continual pretraining, showing the most significant rates of both learning and forgetting, (iv) continual pretraining boosts downstream task performance of GPT-2 family, (v) continual pretraining enables LLMs to specialize better when the sequence of domains shows semantic similarity while randomizing training domains leads to better transfer and final performance otherwise. We posit that our research establishes a new benchmark for CL in LLMs, providing a more realistic evaluation of knowledge retention and transfer across diverse domains.

Saurabh Dash, Yiyang Nan, John Dang et al.

Building multimodal language models is fundamentally challenging: it requires aligning vision and language modalities, curating high-quality instruction data, and avoiding the degradation of existing text-only capabilities once vision is introduced. These difficulties are further magnified in the multilingual setting, where the need for multimodal data in different languages exacerbates existing data scarcity, machine translation often distorts meaning, and catastrophic forgetting is more pronounced. To address the aforementioned challenges, we introduce novel techniques spanning both data and modeling. First, we develop a synthetic annotation framework that curates high-quality, diverse multilingual multimodal instruction data, enabling Aya Vision models to produce natural, human-preferred responses to multimodal inputs across many languages. Complementing this, we propose a cross-modal model merging technique that mitigates catastrophic forgetting, effectively preserving text-only capabilities while simultaneously enhancing multimodal generative performance. Aya-Vision-8B achieves best-in-class performance compared to strong multimodal models such as Qwen-2.5-VL-7B, Pixtral-12B, and even much larger Llama-3.2-90B-Vision. We further scale this approach with Aya-Vision-32B, which outperforms models more than twice its size, such as Molmo-72B and LLaMA-3.2-90B-Vision. Our work advances multilingual progress on the multi-modal frontier, and provides insights into techniques that effectively bend the need for compute while delivering extremely high performance.

Aakanksha, Arash Ahmadian, Seraphina Goldfarb-Tarrant et al.

Large Language Models (LLMs) have been adopted and deployed worldwide for a broad variety of applications. However, ensuring their safe use remains a significant challenge. Preference training and safety measures often overfit to harms prevalent in Western-centric datasets, and safety protocols frequently fail to extend to multilingual settings. In this work, we explore model merging in a diverse multi-task setting, combining safety and general-purpose tasks within a multilingual context. Each language introduces unique and varied learning challenges across tasks. We find that objective-based merging is more effective than mixing data, with improvements of up to 8% and 10% in general performance and safety respectively. We also find that language-based merging is highly effective -- by merging monolingually fine-tuned models, we achieve a 4% increase in general performance and 7% reduction in harm across all languages on top of the data mixtures method using the same available data. Overall, our comprehensive study of merging approaches provides a useful framework for building strong and safe multilingual models.

Israfel Salazar, Manuel Fernández Burda, Shayekh Bin Islam et al. · mila

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

Aidan Peppin, Julia Kreutzer, Alice Schoenauer Sebag et al.

Despite advances in large language model capabilities in recent years, a large gap remains in their capabilities and safety performance for many languages beyond a relatively small handful of globally dominant languages. This paper provides researchers, policymakers and governance experts with an overview of key challenges to bridging the "language gap" in AI and minimizing safety risks across languages. We provide an analysis of why the language gap in AI exists and grows, and how it creates disparities in global AI safety. We identify barriers to address these challenges, and recommend how those working in policy and governance can help address safety concerns associated with the language gap by supporting multilingual dataset creation, transparency, and research.

Aakanksha, Arash Ahmadian, Beyza Ermis et al.

A key concern with the concept of "alignment" is the implicit question of "alignment to what?". AI systems are increasingly used across the world, yet safety alignment is often focused on homogeneous monolingual settings. Additionally, preference training and safety measures often overfit to harms common in Western-centric datasets. Here, we explore the viability of different alignment approaches when balancing dual objectives: addressing and optimizing for a non-homogeneous set of languages and cultural preferences while minimizing both global and local harms. We collect the first set of human annotated red-teaming prompts in different languages distinguishing between global and local harm, which serve as a laboratory for understanding the reliability of alignment techniques when faced with preference distributions that are non-stationary across geographies and languages. While this setting is seldom covered by the literature to date, which primarily centers on English harm mitigation, it captures real-world interactions with AI systems around the world. We establish a new precedent for state-of-the-art alignment techniques across 6 languages with minimal degradation in general performance. Our work provides important insights into cross-lingual transfer and novel optimization approaches to safeguard AI systems designed to serve global populations.

Beyza Ermis, Patrick Ernst, Yannik Stein et al.

Personalization is a crucial aspect of many online experiences. In particular, content ranking is often a key component in delivering sophisticated personalization results. Commonly, supervised learning-to-rank methods are applied, which suffer from bias introduced during data collection by production systems in charge of producing the ranking. To compensate for this problem, we leverage contextual multi-armed bandits. We propose novel extensions of two well-known algorithms viz. LinUCB and Linear Thompson Sampling to the ranking use-case. To account for the biases in a production environment, we employ the position-based click model. Finally, we show the validity of the proposed algorithms by conducting extensive offline experiments on synthetic datasets as well as customer facing online A/B experiments.

Claire Vernade, Alexandra Carpentier, Tor Lattimore et al.

Stochastic linear bandits are a natural and well-studied model for structured exploration/exploitation problems and are widely used in applications such as online marketing and recommendation. One of the main challenges faced by practitioners hoping to apply existing algorithms is that usually the feedback is randomly delayed and delays are only partially observable. For example, while a purchase is usually observable some time after the display, the decision of not buying is never explicitly sent to the system. In other words, the learner only observes delayed positive events. We formalize this problem as a novel stochastic delayed linear bandit and propose ${\tt OTFLinUCB}$ and ${\tt OTFLinTS}$, two computationally efficient algorithms able to integrate new information as it becomes available and to deal with the permanently censored feedback. We prove optimal $\tilde O(\smash{d\sqrt{T}})$ bounds on the regret of the first algorithm and study the dependency on delay-dependent parameters. Our model, assumptions and results are validated by experiments on simulated and real data.

Beyza Ermis, Ali Taylan Cemgil

Deep neural networks with their large number of parameters are highly flexible learning systems. The high flexibility in such networks brings with some serious problems such as overfitting, and regularization is used to address this problem. A currently popular and effective regularization technique for controlling the overfitting is dropout. Often, large data collections required for neural networks contain sensitive information such as the medical histories of patients, and the privacy of the training data should be protected. In this paper, we modify the recently proposed variational dropout technique which provided an elegant Bayesian interpretation to dropout, and show that the intrinsic noise in the variational dropout can be exploited to obtain a degree of differential privacy. The iterative nature of training neural networks presents a challenge for privacy-preserving estimation since multiple iterations increase the amount of noise added. We overcome this by using a relaxed notion of differential privacy, called concentrated differential privacy, which provides tighter estimates on the overall privacy loss. We demonstrate the accuracy of our privacy-preserving variational dropout algorithm on benchmark datasets.

Beyza Ermis, Ali Taylan Cemgil

Large data collections required for the training of neural networks often contain sensitive information such as the medical histories of patients, and the privacy of the training data must be preserved. In this paper, we introduce a dropout technique that provides an elegant Bayesian interpretation to dropout, and show that the intrinsic noise added, with the primary goal of regularization, can be exploited to obtain a degree of differential privacy. The iterative nature of training neural networks presents a challenge for privacy-preserving estimation since multiple iterations increase the amount of noise added. We overcome this by using a relaxed notion of differential privacy, called concentrated differential privacy, which provides tighter estimates on the overall privacy loss. We demonstrate the accuracy of our privacy-preserving dropout algorithm on benchmark datasets.

Cedric Archambeau, Beyza Ermis

We introduce incremental variational inference and apply it to latent Dirichlet allocation (LDA). Incremental variational inference is inspired by incremental EM and provides an alternative to stochastic variational inference. Incremental LDA can process massive document collections, does not require to set a learning rate, converges faster to a local optimum of the variational bound and enjoys the attractive property of monotonically increasing it. We study the performance of incremental LDA on large benchmark data sets. We further introduce a stochastic approximation of incremental variational inference which extends to the asynchronous distributed setting. The resulting distributed algorithm achieves comparable performance as single host incremental variational inference, but with a significant speed-up.

Beyza Ermis, A. Taylan Cemgil

Probabilistic approaches for tensor factorization aim to extract meaningful structure from incomplete data by postulating low rank constraints. Recently, variational Bayesian (VB) inference techniques have successfully been applied to large scale models. This paper presents full Bayesian inference via VB on both single and coupled tensor factorization models. Our method can be run even for very large models and is easily implemented. It exhibits better prediction performance than existing approaches based on maximum likelihood on several real-world datasets for missing link prediction problem.

Beyza Ermiş, Evrim Acar, A. Taylan Cemgil

This study deals with the missing link prediction problem: the problem of predicting the existence of missing connections between entities of interest. We address link prediction using coupled analysis of relational datasets represented as heterogeneous data, i.e., datasets in the form of matrices and higher-order tensors. We propose to use an approach based on probabilistic interpretation of tensor factorisation models, i.e., Generalised Coupled Tensor Factorisation, which can simultaneously fit a large class of tensor models to higher-order tensors/matrices with com- mon latent factors using different loss functions. Numerical experiments demonstrate that joint analysis of data from multiple sources via coupled factorisation improves the link prediction performance and the selection of right loss function and tensor model is crucial for accurately predicting missing links.