Guokan Shang

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}.

Yanzhu Guo, Guokan Shang, Michalis Vazirgiannis et al.

This study investigates the consequences of training language models on synthetic data generated by their predecessors, an increasingly prevalent practice given the prominence of powerful generative models. Diverging from the usual emphasis on performance metrics, we focus on the impact of this training methodology on linguistic diversity, especially when conducted recursively over time. To assess this, we adapt and develop a set of novel metrics targeting lexical, syntactic, and semantic diversity, applying them in recursive finetuning experiments across various natural language generation tasks in English. Our findings reveal a consistent decrease in the diversity of the model outputs through successive iterations, especially remarkable for tasks demanding high levels of creativity. This trend underscores the potential risks of training language models on synthetic text, particularly concerning the preservation of linguistic richness. Our study highlights the need for careful consideration of the long-term effects of such training approaches on the linguistic capabilities of language models.

Virgile Rennard, Guokan Shang, Julie Hunter et al.

A system that could reliably identify and sum up the most important points of a conversation would be valuable in a wide variety of real-world contexts, from business meetings to medical consultations to customer service calls. Recent advances in deep learning, and especially the invention of encoder-decoder architectures, has significantly improved language generation systems, opening the door to improved forms of abstractive summarization, a form of summarization particularly well-suited for multi-party conversation. In this paper, we provide an overview of the challenges raised by the task of abstractive meeting summarization and of the data sets, models and evaluation metrics that have been used to tackle the problems.

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.

Yanzhu Guo, Guokan Shang, Virgile Rennard et al.

With the increasing amount of problematic peer reviews in top AI conferences, the community is urgently in need of automatic quality control measures. In this paper, we restrict our attention to substantiation -- one popular quality aspect indicating whether the claims in a review are sufficiently supported by evidence -- and provide a solution automatizing this evaluation process. To achieve this goal, we first formulate the problem as claim-evidence pair extraction in scientific peer reviews, and collect SubstanReview, the first annotated dataset for this task. SubstanReview consists of 550 reviews from NLP conferences annotated by domain experts. On the basis of this dataset, we train an argument mining system to automatically analyze the level of substantiation in peer reviews. We also perform data analysis on the SubstanReview dataset to obtain meaningful insights on peer reviewing quality in NLP conferences over recent years.

Moussa Kamal Eddine, Guokan Shang, Michalis Vazirgiannis

The rapid development of large pretrained language models has revolutionized not only the field of Natural Language Generation (NLG) but also its evaluation. Inspired by the recent work of BARTScore: a metric leveraging the BART language model to evaluate the quality of generated text from various aspects, we introduce DATScore. DATScore uses data augmentation techniques to improve the evaluation of machine translation. Our main finding is that introducing data augmented translations of the source and reference texts is greatly helpful in evaluating the quality of the generated translation. We also propose two novel score averaging and term weighting strategies to improve the original score computing process of BARTScore. Experimental results on WMT show that DATScore correlates better with human meta-evaluations than the other recent state-of-the-art metrics, especially for low-resource languages. Ablation studies demonstrate the value added by our new scoring strategies. Moreover, we report in our extended experiments the performance of DATScore on 3 NLG tasks other than translation.

Julie Hunter, Jérôme Louradour, Virgile Rennard et al.

We present the Claire French Dialogue Dataset (CFDD), a resource created by members of LINAGORA Labs in the context of the OpenLLM France initiative. CFDD is a corpus containing roughly 160 million words from transcripts and stage plays in French that we have assembled and publicly released in an effort to further the development of multilingual, open source language models. This paper describes the 24 individual corpora of which CFDD is composed and provides links and citations to their original sources. It also provides our proposed breakdown of the full CFDD dataset into eight categories of subcorpora and describes the process we followed to standardize the format of the final dataset. We conclude with a discussion of similar work and future directions.

Yang Zhang, Mersin Konomi, Christos Xypolopoulos et al.

Large Language Models (LLMs) are commonly trained on multilingual corpora that include Greek, yet reliable evaluation benchmarks for Greek-particularly those based on authentic, native-sourced content-remain limited. Existing datasets are often machine-translated from English, failing to capture Greek linguistic and cultural characteristics. We introduce GreekMMLU, a native-sourced benchmark for massive multitask language understanding in Greek, comprising 21,805 multiple-choice questions across 45 subject areas, organized under a newly defined subject taxonomy and annotated with educational difficulty levels spanning primary to professional examinations. All questions are sourced or authored in Greek from academic, professional, and governmental exams. We publicly release 16,857 samples and reserve 4,948 samples for a private leaderboard to enable robust and contamination-resistant evaluation. Evaluations of over 80 open- and closed-source LLMs reveal substantial performance gaps between frontier and open-weight models, as well as between Greek-adapted models and general multilingual ones. Finally, we provide a systematic analysis of factors influencing performance-including model scale, adaptation, and prompting-and derive insights for improving LLM capabilities in Greek.

Amr Mohamed, Yang Zhang, Michalis Vazirgiannis et al.

Diffusion large language models (dLLMs) offer a promising alternative to autoregressive models, but their practical utility is severely hampered by slow, iterative sampling. We present SchED, a training-free, model-agnostic early-exit algorithm that aggregates full-span logit margins and halts decoding once a smooth, progress-dependent confidence threshold is met. We evaluated SchED on two dLLM families (Dream and LLaDA), in base and instruction-tuned variants across ten benchmarks spanning downstream tasks including multiple-choice question answering (MCQ), math, long-form QA/summarization, and translation. SchED delivers large, stable accelerations: on instruction-tuned models, it achieves $3.8$-$4.0\times$ speedups while retaining $99.8$-$100\%$ of the baseline score on average. On base models, SchED yields consistent speedup gains with $99.1$-$100\%$ performance retention, with up to $2.34\times$ under more aggressive settings. Using a conservative speed metric that heavily penalizes quality loss (QPS, $γ{=}4$), we show that SchED is robust and clearly outperforms prior confidence-based early-exit methods, which break down on long-form generation. An entropy analysis of the model's token predictions reveals that instruction tuning speeds up the decay of predictive entropy. By turning genuine confidence stabilization into computational savings, SchED makes dLLM decoding substantially more efficient.

Mingmeng Geng, Amr Mohamed, Guokan Shang et al.

The widespread use of large language models (LLMs) raises an important question: how do texts evolve when they are repeatedly processed by LLMs? In this paper, we define this iterative inference process as Markovian generation chains, where each step takes a specific prompt template and the previous output as input, without including any prior memory. In iterative rephrasing and round-trip translation experiments, the output either converges to a small recurrent set or continues to produce novel sentences over a finite horizon. Through sentence-level Markov chain modeling and analysis of simulated data, we show that iterative process can either increase or reduce sentence diversity depending on factors such as the temperature parameter and the initial input sentence. These results offer valuable insights into the dynamics of iterative LLM inference and their implications for multi-agent LLM systems.

Ahmad Chamma, Omar El Herraoui, Guokan Shang

We introduce MixtureKit, a modular open-source framework for constructing, training, and analyzing Mixture-of-Experts (MoE) models from arbitrary pre-trained or fine-tuned models. MixtureKit currently supports three complementary methods: (i) \emph{Traditional MoE}, which uses a single router per transformer block to select experts, (ii) \emph{BTX} (Branch-Train-Mix), which introduces separate routers for each specified sub-layer enabling fine-grained token routing, and (iii) \emph{BTS} (Branch-Train-Stitch), which keeps experts fully intact and introduces trainable stitch layers for controlled information exchange between hub and experts. MixtureKit automatically modifies the model configuration, patches decoder and causal LM classes, and saves a unified checkpoint ready for inference or fine-tuning. We further provide a visualization interface to inspect per-token routing decisions, expert weight distributions, and layer-wise contributions. Experiments with multilingual code-switched data (e.g. Arabic-Latin) show that a BTX-based model trained using MixtureKit can outperform baseline dense models on multiple benchmarks. We release MixtureKit as a practical foundation for research and development of MoE-based systems across diverse domains.

Yanzhu Guo, Guokan Shang, Chloé Clavel

The development and evaluation of Large Language Models (LLMs) has primarily focused on their task-solving capabilities, with recent models even surpassing human performance in some areas. However, this focus often neglects whether machine-generated language matches the human level of diversity, in terms of vocabulary choice, syntactic construction, and expression of meaning, raising questions about whether the fundamentals of language generation have been fully addressed. This paper emphasizes the importance of examining the preservation of human linguistic richness by language models, given the concerning surge in online content produced or aided by LLMs. We propose a comprehensive framework for evaluating LLMs from various linguistic diversity perspectives including lexical, syntactic, and semantic dimensions. Using this framework, we benchmark several state-of-the-art LLMs across all diversity dimensions, and conduct an in-depth case study for syntactic diversity. Finally, we analyze how different development and deployment choices impact the linguistic diversity of LLM outputs.

Amr Mohamed, Yang Zhang, Michalis Vazirgiannis et al.

Code-switching (CSW) is the act of alternating between two or more languages within a single discourse. This phenomenon is widespread in multilingual communities, and increasingly prevalent in online content, where users naturally mix languages in everyday communication. As a result, Large Language Models (LLMs), now central to content processing and generation, are frequently exposed to code-switched inputs. Given their widespread use, it is crucial to understand how LLMs process and reason about such mixed-language text. This paper presents a systematic evaluation of LLM comprehension under code-switching by generating CSW variants of established reasoning and comprehension benchmarks. While degradation is evident when foreign tokens disrupt English text$\unicode{x2013}$even under linguistic constraints$\unicode{x2013}$embedding English into other languages often improves comprehension. Though prompting yields mixed results, fine-tuning offers a more stable path to degradation mitigation.

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.

Virgile Rennard, Guokan Shang, Damien Grari et al.

Recent advances in deep learning, and especially the invention of encoder-decoder architectures, has significantly improved the performance of abstractive summarization systems. The majority of research has focused on written documents, however, neglecting the problem of multi-party dialogue summarization. In this paper, we present a dataset of French political debates for the purpose of enhancing resources for multi-lingual dialogue summarization. Our dataset consists of manually transcribed and annotated political debates, covering a range of topics and perspectives. We highlight the importance of high quality transcription and annotations for training accurate and effective dialogue summarization models, and emphasize the need for multilingual resources to support dialogue summarization in non-English languages. We also provide baseline experiments using state-of-the-art methods, and encourage further research in this area to advance the field of dialogue summarization. Our dataset will be made publicly available for use by the research community.

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.

Amr Mohamed, Mingmeng Geng, Michalis Vazirgiannis et al.

As large language models are increasingly responsible for online content, concerns arise about the impact of repeatedly processing their own outputs. Inspired by the "broken telephone" effect in chained human communication, this study investigates whether LLMs similarly distort information through iterative generation. Through translation-based experiments, we find that distortion accumulates over time, influenced by language choice and chain complexity. While degradation is inevitable, it can be mitigated through strategic prompting techniques. These findings contribute to discussions on the long-term effects of AI-mediated information propagation, raising important questions about the reliability of LLM-generated content in iterative workflows.

Virgile Rennard, Guokan Shang, Michalis Vazirgiannis et al.

We introduce an extractive summarization system for meetings that leverages discourse structure to better identify salient information from complex multi-party discussions. Using discourse graphs to represent semantic relations between the contents of utterances in a meeting, we train a GNN-based node classification model to select the most important utterances, which are then combined to create an extractive summary. Experimental results on AMI and ICSI demonstrate that our approach surpasses existing text-based and graph-based extractive summarization systems, as measured by both classification and summarization metrics. Additionally, we conduct ablation studies on discourse structure and relation type to provide insights for future NLP applications leveraging discourse analysis theory.

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.

Moussa Kamal Eddine, Guokan Shang, Antoine J. -P. Tixier et al.

Fast and reliable evaluation metrics are key to R&D progress. While traditional natural language generation metrics are fast, they are not very reliable. Conversely, new metrics based on large pretrained language models are much more reliable, but require significant computational resources. In this paper, we propose FrugalScore, an approach to learn a fixed, low cost version of any expensive NLG metric, while retaining most of its original performance. Experiments with BERTScore and MoverScore on summarization and translation show that FrugalScore is on par with the original metrics (and sometimes better), while having several orders of magnitude less parameters and running several times faster. On average over all learned metrics, tasks, and variants, FrugalScore retains 96.8% of the performance, runs 24 times faster, and has 35 times less parameters than the original metrics. We make our trained metrics publicly available, to benefit the entire NLP community and in particular researchers and practitioners with limited resources.

Guokan Shang, Antoine Jean-Pierre Tixier, Michalis Vazirgiannis et al.

Recent work in Dialogue Act (DA) classification approaches the task as a sequence labeling problem, using neural network models coupled with a Conditional Random Field (CRF) as the last layer. CRF models the conditional probability of the target DA label sequence given the input utterance sequence. However, the task involves another important input sequence, that of speakers, which is ignored by previous work. To address this limitation, this paper proposes a simple modification of the CRF layer that takes speaker-change into account. Experiments on the SwDA corpus show that our modified CRF layer outperforms the original one, with very wide margins for some DA labels. Further, visualizations demonstrate that our CRF layer can learn meaningful, sophisticated transition patterns between DA label pairs conditioned on speaker-change in an end-to-end way. Code is publicly available.

Guokan Shang, Antoine Jean-Pierre Tixier, Michalis Vazirgiannis et al.

Abstractive community detection is an important spoken language understanding task, whose goal is to group utterances in a conversation according to whether they can be jointly summarized by a common abstractive sentence. This paper provides a novel approach to this task. We first introduce a neural contextual utterance encoder featuring three types of self-attention mechanisms. We then train it using the siamese and triplet energy-based meta-architectures. Experiments on the AMI corpus show that our system outperforms multiple energy-based and non-energy based baselines from the state-of-the-art. Code and data are publicly available.

Guokan Shang, Wensi Ding, Zekun Zhang et al.

We introduce a novel graph-based framework for abstractive meeting speech summarization that is fully unsupervised and does not rely on any annotations. Our work combines the strengths of multiple recent approaches while addressing their weaknesses. Moreover, we leverage recent advances in word embeddings and graph degeneracy applied to NLP to take exterior semantic knowledge into account, and to design custom diversity and informativeness measures. Experiments on the AMI and ICSI corpus show that our system improves on the state-of-the-art. Code and data are publicly available, and our system can be interactively tested.