Hadi Abdine

Abdelaziz Bounhar, Rania Hossam Elmohamady Elbadry, Hadi Abdine et al.

Steering Large Language Models (LLMs) through activation interventions has emerged as a lightweight alternative to fine-tuning for alignment and personalization. Recent work on Bi-directional Preference Optimization (BiPO) shows that dense steering vectors can be learned directly from preference data in a Direct Preference Optimization (DPO) fashion, enabling control over truthfulness, hallucinations, and safety behaviors. However, dense steering vectors often entangle multiple latent factors due to neuron multi-semanticity, limiting their effectiveness and stability in fine-grained settings such as cultural alignment, where closely related values and behaviors (e.g., among Middle Eastern cultures) must be distinguished. In this paper, we propose Yet another Policy Optimization (YaPO), a \textit{reference-free} method that learns \textit{sparse steering vectors} in the latent space of a Sparse Autoencoder (SAE). By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Empirically, we show that YaPO converges faster, achieves stronger performance, and exhibits improved training stability compared to dense steering baselines. Beyond cultural alignment, YaPO generalizes to a range of alignment-related behaviors, including hallucination, wealth-seeking, jailbreak, and power-seeking. Importantly, YaPO preserves general knowledge, with no measurable degradation on MMLU. Overall, our results show that YaPO provides a general recipe for efficient, stable, and fine-grained alignment of LLMs, with broad applications to controllability and domain adaptation. The associated code and data are publicly available\footnote{https://github.com/MBZUAI-Paris/YaPO}.

Abdelaziz Bounhar, Hadi Abdine, Evan Dufraisse et al.

Large language models (LLMs) trained for step-by-step reasoning often become excessively verbose, raising inference cost. Standard Reinforcement Learning with Verifiable Rewards (RLVR) pipelines filter out ``easy'' problems for training efficiency, leaving the model to train primarily on harder problems that require longer reasoning chains. This skews the output length distribution upward, resulting in a \textbf{model that conflates ``thinking longer'' with ``thinking better''}. In this work, we show that retaining and modestly up-weighting moderately easy problems acts as an implicit length regularizer. Exposing the model to solvable short-chain tasks constrains its output distribution and prevents runaway verbosity. The result is \textbf{\emph{emergent brevity for free}}: the model learns to solve harder problems without inflating the output length, \textbf{ despite the absence of any explicit length penalization}. RLVR experiments using this approach on \textit{Qwen3-4B-Thinking-2507} (with a 16k token limit) achieve baseline pass@1 AIME25 accuracy while generating solutions that are, on average, nearly twice as short. The code is available at \href{https://github.com/MBZUAI-Paris/Frugal-AI}{GitHub}, with datasets and models on \href{https://huggingface.co/collections/MBZUAI-Paris/k2-think-mini-68dcfa8b114686a4bd3dc2bc}{Hugging Face}.

Hadi Abdine, Michail Chatzianastasis, Costas Bouyioukos et al.

In recent years, significant progress has been made in the field of protein function prediction with the development of various machine-learning approaches. However, most existing methods formulate the task as a multi-classification problem, i.e. assigning predefined labels to proteins. In this work, we propose a novel approach, Prot2Text, which predicts a protein's function in a free text style, moving beyond the conventional binary or categorical classifications. By combining Graph Neural Networks(GNNs) and Large Language Models(LLMs), in an encoder-decoder framework, our model effectively integrates diverse data types including protein sequence, structure, and textual annotation and description. This multimodal approach allows for a holistic representation of proteins' functions, enabling the generation of detailed and accurate functional descriptions. To evaluate our model, we extracted a multimodal protein dataset from SwissProt, and demonstrate empirically the effectiveness of Prot2Text. These results highlight the transformative impact of multimodal models, specifically the fusion of GNNs and LLMs, empowering researchers with powerful tools for more accurate function prediction of existing as well as first-to-see proteins.

Guokan Shang, Hadi Abdine, Yousef Khoubrane et al.

We introduce Atlas-Chat, the first-ever collection of LLMs specifically developed for dialectal Arabic. Focusing on Moroccan Arabic, also known as Darija, we construct our instruction dataset by consolidating existing Darija language resources, creating novel datasets both manually and synthetically, and translating English instructions with stringent quality control. Atlas-Chat-2B, 9B, and 27B models, fine-tuned on the dataset, exhibit superior ability in following Darija instructions and performing standard NLP tasks. Notably, our models outperform both state-of-the-art and Arabic-specialized LLMs like LLaMa, Jais, and AceGPT, e.g., our 9B model gains a 13% performance boost over a larger 13B model on DarijaMMLU, in our newly introduced evaluation suite for Darija covering both discriminative and generative tasks. Furthermore, we perform an experimental analysis of various fine-tuning strategies and base model choices to determine optimal configurations. All our resources are publicly accessible, and we believe our work offers comprehensive design methodologies of instruction-tuning for low-resource languages, which are often neglected in favor of data-rich languages by contemporary LLMs.

Iakovos Evdaimon, Hadi Abdine, Christos Xypolopoulos et al.

The era of transfer learning has revolutionized the fields of Computer Vision and Natural Language Processing, bringing powerful pretrained models with exceptional performance across a variety of tasks. Specifically, Natural Language Processing tasks have been dominated by transformer-based language models. In Natural Language Inference and Natural Language Generation tasks, the BERT model and its variants, as well as the GPT model and its successors, demonstrated exemplary performance. However, the majority of these models are pretrained and assessed primarily for the English language or on a multilingual corpus. In this paper, we introduce GreekBART, the first Seq2Seq model based on BART-base architecture and pretrained on a large-scale Greek corpus. We evaluate and compare GreekBART against BART-random, Greek-BERT, and XLM-R on a variety of discriminative tasks. In addition, we examine its performance on two NLG tasks from GreekSUM, a newly introduced summarization dataset for the Greek language. The model, the code, and the new summarization dataset will be publicly available.

Hadi Abdine, Yanzhu Guo, Virgile Rennard et al.

With the significant increase in users on social media platforms, a new means of political campaigning has appeared. Twitter and Facebook are now notable campaigning tools during elections. Indeed, the candidates and their parties now take to the internet to interact and spread their ideas. In this paper, we aim to identify political communities formed on Twitter during the 2022 French presidential election and analyze each respective community. We create a large-scale Twitter dataset containing 1.2 million users and 62.6 million tweets that mention keywords relevant to the election. We perform community detection on a retweet graph of users and propose an in-depth analysis of the stance of each community. Finally, we attempt to detect offensive tweets and automatic bots, comparing across communities in order to gain insight into each candidate's supporter demographics and online campaign strategy.

Hadi Abdine, Moussa Kamal Eddine, Michalis Vazirgiannis et al.

Word sense induction (WSI) is a difficult problem in natural language processing that involves the unsupervised automatic detection of a word's senses (i.e. meanings). Recent work achieves significant results on the WSI task by pre-training a language model that can exclusively disambiguate word senses, whereas others employ previously pre-trained language models in conjunction with additional strategies to induce senses. In this paper, we propose a novel unsupervised method based on hierarchical clustering and invariant information clustering (IIC). The IIC is used to train a small model to optimize the mutual information between two vector representations of a target word occurring in a pair of synthetic paraphrases. This model is later used in inference mode to extract a higher quality vector representation to be used in the hierarchical clustering. We evaluate our method on two WSI tasks and in two distinct clustering configurations (fixed and dynamic number of clusters). We empirically demonstrate that, in certain cases, our approach outperforms prior WSI state-of-the-art methods, while in others, it achieves a competitive performance.

Iakovos Evdaimon, Giannis Nikolentzos, Christos Xypolopoulos et al.

Graph generation has emerged as a crucial task in machine learning, with significant challenges in generating graphs that accurately reflect specific properties. Existing methods often fall short in efficiently addressing this need as they struggle with the high-dimensional complexity and varied nature of graph properties. In this paper, we introduce the Neural Graph Generator (NGG), a novel approach which utilizes conditioned latent diffusion models for graph generation. NGG demonstrates a remarkable capacity to model complex graph patterns, offering control over the graph generation process. NGG employs a variational graph autoencoder for graph compression and a diffusion process in the latent vector space, guided by vectors summarizing graph statistics. We demonstrate NGG's versatility across various graph generation tasks, showing its capability to capture desired graph properties and generalize to unseen graphs. We also compare our generator to the graph generation capabilities of different LLMs. This work signifies a shift in graph generation methodologies, offering a more practical and efficient solution for generating diverse graphs with specific characteristics.

Yang Zhang, Amr Mohamed, Hadi Abdine et al.

Curriculum learning has shown promise in improving training efficiency and generalization in various machine learning domains, yet its potential in pretraining language models remains underexplored, prompting our work as the first systematic investigation in this area. We experimented with different settings, including vanilla curriculum learning, pacing-based sampling, and interleaved curricula-guided by six difficulty metrics spanning linguistic and information-theoretic perspectives. We train models under these settings and evaluate their performance on eight diverse benchmarks. Our experiments reveal that curriculum learning consistently improves convergence in early and mid-training phases, and can yield lasting gains when used as a warmup strategy with up to $3.5\%$ improvement. Notably, we identify compression ratio, lexical diversity, and readability as effective difficulty signals across settings. Our findings highlight the importance of data ordering in large-scale pretraining and provide actionable insights for scalable, data-efficient model development under realistic training scenarios.

Christos Xypolopoulos, Guokan Shang, Xiao Fei et al.

Large language models have evolved to process multiple modalities beyond text, such as images and audio, which motivates us to explore how to effectively leverage them for graph reasoning tasks. The key question, therefore, is how to transform graphs into linear sequences of tokens, a process we term "graph linearization", so that LLMs can handle graphs naturally. We consider that graphs should be linearized meaningfully to reflect certain properties of natural language text, such as local dependency and global alignment, in order to ease contemporary LLMs, trained on trillions of textual tokens, better understand graphs. To achieve this, we developed several graph linearization methods based on graph centrality and degeneracy. These methods are further enhanced using node relabeling techniques. The experimental results demonstrate the effectiveness of our methods compared to the random linearization baseline. Our work introduces novel graph representations suitable for LLMs, contributing to the potential integration of graph machine learning with the trend of multimodal processing using a unified transformer model.

Guokan Shang, Hadi Abdine, Ahmad Chamma et al.

We introduce Nile-Chat-4B, 3x4B-A6B, and 12B, a collection of LLMs for Egyptian dialect, uniquely designed to understand and generate texts written in both Arabic and Latin scripts. Specifically, with Nile-Chat-3x4B-A6B, we introduce a novel language adaptation approach by leveraging the Branch-Train-MiX strategy to merge script-specialized experts, into a single MoE model. Our Nile-Chat models significantly outperform leading multilingual and Arabic LLMs, such as LLaMa, Jais, and ALLaM, on our newly introduced Egyptian evaluation benchmarks, which span both understanding and generative tasks. Notably, our 12B model yields a 14.4% performance gain over Qwen2.5-14B-Instruct on Latin-script benchmarks. All our resources are publicly available. We believe this work presents a comprehensive methodology for adapting LLMs to dual-script languages, addressing an often overlooked aspect in modern LLM development.

Hadi Abdine, Yanzhu Guo, Moussa Kamal Eddine et al.

DaSciM (Data Science and Mining) part of LIX at Ecole Polytechnique, established in 2013 and since then producing research results in the area of large scale data analysis via methods of machine and deep learning. The group has been specifically active in the area of NLP and text mining with interesting results at methodological and resources level. Here follow our different contributions of interest to the AFIA community.

Stella Douka, Hadi Abdine, Michalis Vazirgiannis et al.

Language models have proven to be very useful when adapted to specific domains. Nonetheless, little research has been done on the adaptation of domain-specific BERT models in the French language. In this paper, we focus on creating a language model adapted to French legal text with the goal of helping law professionals. We conclude that some specific tasks do not benefit from generic language models pre-trained on large amounts of data. We explore the use of smaller architectures in domain-specific sub-languages and their benefits for French legal text. We prove that domain-specific pre-trained models can perform better than their equivalent generalised ones in the legal domain. Finally, we release JuriBERT, a new set of BERT models adapted to the French legal domain.

Hadi Abdine, Christos Xypolopoulos, Moussa Kamal Eddine et al.

Distributed word representations are popularly used in many tasks in natural language processing. Adding that pretrained word vectors on huge text corpus achieved high performance in many different NLP tasks. This paper introduces multiple high-quality word vectors for the French language where two of them are trained on massive crawled French data during this study and the others are trained on an already existing French corpus. We also evaluate the quality of our proposed word vectors and the existing French word vectors on the French word analogy task. In addition, we do the evaluation on multiple real NLP tasks that shows the important performance enhancement of the pre-trained word vectors compared to the existing and random ones. Finally, we created a demo web application to test and visualize the obtained word embeddings. The produced French word embeddings are available to the public, along with the finetuning code on the NLU tasks and the demo code.