Mohamed El Amine Seddik

Quentin Malartic, Nilabhra Roy Chowdhury, Ruxandra Cojocaru et al.

We introduce Falcon2-11B, a foundation model trained on over five trillion tokens, and its multimodal counterpart, Falcon2-11B-vlm, which is a vision-to-text model. We report our findings during the training of the Falcon2-11B which follows a multi-stage approach where the early stages are distinguished by their context length and a final stage where we use a curated, high-quality dataset. Additionally, we report the effect of doubling the batch size mid-training and how training loss spikes are affected by the learning rate. The downstream performance of the foundation model is evaluated on established benchmarks, including multilingual and code datasets. The foundation model shows strong generalization across all the tasks which makes it suitable for downstream finetuning use cases. For the vision language model, we report the performance on several benchmarks and show that our model achieves a higher average score compared to open-source models of similar size. The model weights and code of both Falcon2-11B and Falcon2-11B-vlm are made available under a permissive license.

Mohamed El Amine Seddik, José Henrique de Morais Goulart, Maxime Guillaud

This paper studies the deflation algorithm when applied to estimate a low-rank symmetric spike contained in a large tensor corrupted by additive Gaussian noise. Specifically, we provide a precise characterization of the large-dimensional performance of deflation in terms of the alignments of the vectors obtained by successive rank-1 approximation and of their estimated weights, assuming non-trivial (fixed) correlations among spike components. Our analysis allows an understanding of the deflation mechanism in the presence of noise and can be exploited for designing more efficient signal estimation methods.

Mohamed El Amine Seddik, Maxime Guillaud, Alexis Decurninge et al.

This work introduces an asymptotic study of Hotelling-type tensor deflation in the presence of noise, in the regime of large tensor dimensions. Specifically, we consider a low-rank asymmetric tensor model of the form $\sum_{i=1}^r β_i{\mathcal{A}}_i + {\mathcal{W}}$ where $β_i\geq 0$ and the ${\mathcal{A}}_i$'s are unit-norm rank-one tensors such that $\left| \langle {\mathcal{A}}_i, {\mathcal{A}}_j \rangle \right| \in [0, 1]$ for $i\neq j$ and ${\mathcal{W}}$ is an additive noise term. Assuming that the dominant components are successively estimated from the noisy observation and subsequently subtracted, we leverage recent advances in random tensor theory in the regime of asymptotically large tensor dimensions to analytically characterize the estimated singular values and the alignment of estimated and true singular vectors at each step of the deflation procedure. Furthermore, this result can be used to construct estimators of the signal-to-noise ratios $β_i$ and the alignments between the estimated and true rank-1 signal components.

Reda Alami, Ali Khalifa Almansoori, Ahmed Alzubaidi et al.

We demonstrate that preference optimization methods can effectively enhance LLM safety. Applying various alignment techniques to the Falcon 11B model using safety datasets, we achieve a significant boost in global safety score (from $57.64\%$ to $99.90\%$) as measured by LlamaGuard 3 8B, competing with state-of-the-art models. On toxicity benchmarks, average scores in adversarial settings dropped from over $0.6$ to less than $0.07$. However, this safety improvement comes at the cost of reduced general capabilities, particularly in math, suggesting a trade-off. We identify noise contrastive alignment (Safe-NCA) as an optimal method for balancing safety and performance. Our study ultimately shows that alignment techniques can be sufficient for building safe and robust models.

Mayug Maniparambil, Raiymbek Akshulakov, Yasser Abdelaziz Dahou Djilali et al.

Aligned text-image encoders such as CLIP have become the de facto model for vision-language tasks. Furthermore, modality-specific encoders achieve impressive performances in their respective domains. This raises a central question: does an alignment exist between uni-modal vision and language encoders since they fundamentally represent the same physical world? Analyzing the latent spaces structure of vision and language models on image-caption benchmarks using the Centered Kernel Alignment (CKA), we find that the representation spaces of unaligned and aligned encoders are semantically similar. In the absence of statistical similarity in aligned encoders like CLIP, we show that a possible matching of unaligned encoders exists without any training. We frame this as a seeded graph-matching problem exploiting the semantic similarity between graphs and propose two methods - a Fast Quadratic Assignment Problem optimization, and a novel localized CKA metric-based matching/retrieval. We demonstrate the effectiveness of this on several downstream tasks including cross-lingual, cross-domain caption matching and image classification. Code available at github.com/mayug/0-shot-llm-vision.

Jingwei Zuo, Maksim Velikanov, Ilyas Chahed et al.

In this report, we introduce Falcon-H1, a new series of large language models (LLMs) featuring hybrid architecture designs optimized for both high performance and efficiency across diverse use cases. Unlike earlier Falcon models built solely on Transformer or Mamba architectures, Falcon-H1 adopts a parallel hybrid approach that combines Transformer-based attention with State Space Models (SSMs), known for superior long-context memory and computational efficiency. We systematically revisited model design, data strategy, and training dynamics, challenging conventional practices in the field. Falcon-H1 is released in multiple configurations, including base and instruction-tuned variants at 0.5B, 1.5B, 1.5B-deep, 3B, 7B, and 34B parameters. Quantized instruction-tuned models are also available, totaling over 30 checkpoints on Hugging Face Hub. Falcon-H1 models demonstrate state-of-the-art performance and exceptional parameter and training efficiency. The flagship Falcon-H1-34B matches or outperforms models up to 70B scale, such as Qwen3-32B, Qwen2.5-72B, and Llama3.3-70B, while using fewer parameters and less data. Smaller models show similar trends: the Falcon-H1-1.5B-Deep rivals current leading 7B-10B models, and Falcon-H1-0.5B performs comparably to typical 7B models from 2024. These models excel across reasoning, mathematics, multilingual tasks, instruction following, and scientific knowledge. With support for up to 256K context tokens and 18 languages, Falcon-H1 is suitable for a wide range of applications. All models are released under a permissive open-source license, underscoring our commitment to accessible and impactful AI research.

Falcon LLM Team, Iheb Chaabane, Puneesh Khanna et al.

This work introduces Falcon-H1R, a 7B-parameter reasoning-optimized model that establishes the feasibility of achieving competitive reasoning performance with small language models (SLMs). Falcon-H1R stands out for its parameter efficiency, consistently matching or outperforming SOTA reasoning models that are $2\times$ to $7\times$ larger across a variety of reasoning-intensive benchmarks. These results underscore the importance of careful data curation and targeted training strategies (via both efficient SFT and RL scaling) in delivering significant performance gains without increasing model size. Furthermore, Falcon-H1R advances the 3D limits of reasoning efficiency by combining faster inference (through its hybrid-parallel architecture design), token efficiency, and higher accuracy. This unique blend makes Falcon-H1R-7B a practical backbone for scaling advanced reasoning systems, particularly in scenarios requiring extensive chain-of-thoughts generation and parallel test-time scaling. Leveraging the recently introduced DeepConf approach, Falcon-H1R achieves state-of-the-art test-time scaling efficiency, offering substantial improvements in both accuracy and computational cost. As a result, Falcon-H1R demonstrates that compact models, through targeted model training and architectural choices, can deliver robust and scalable reasoning performance.

Adam Younsi, Ahmed Attia, Abdalgader Abubaker et al.

Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4\% absolute on SAT MATH). Furthermore, we benchmark our PRM against existing open-source reward models, demonstrating superior alignment with reasoning quality and more consistent guidance for downstream generation. Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.

Mohamed El Amine Seddik, Changmin Wu, Johannes F. Lutzeyer et al.

The robustness of the much-used Graph Convolutional Networks (GCNs) to perturbations of their input is becoming a topic of increasing importance. In this paper, the random GCN is introduced for which a random matrix theory analysis is possible. This analysis suggests that if the graph is sufficiently perturbed, or in the extreme case random, then the GCN fails to benefit from the node features. It is furthermore observed that enhancing the message passing step in GCNs by adding the node feature kernel to the adjacency matrix of the graph structure solves this problem. An empirical study of a GCN utilised for node classification on six real datasets further confirms the theoretical findings and demonstrates that perturbations of the graph structure can result in GCNs performing significantly worse than Multi-Layer Perceptrons run on the node features alone. In practice, adding a node feature kernel to the message passing of perturbed graphs results in a significant improvement of the GCN's performance, thereby rendering it more robust to graph perturbations. Our code is publicly available at:https://github.com/ChangminWu/RobustGCN.

Mohamed El Amine Seddik

We study the spectral properties of sample covariance matrices constructed from pooled sequence representations, where token embeddings are drawn from a fixed two-class Gaussian mixture table and pooled via (fixed) attention weights. Working in the high-dimensional regime $d,V,N\to\infty$ with $d/V\toδ$ and $d/N\toγ$, we derive exact characterizations of the limiting eigenvalue distribution, outlier eigenvalues, and eigenvector alignment with the hidden signal. The bulk spectrum follows a non-Marchenko--Pastur law given by the free multiplicative convolution $κ(MP_δ\boxtimes MP_γ)$, reflecting the finite vocabulary structure. Signal recovery undergoes two successive BBP-type phase transitions characterized by the scalars: $δ,γ,α=w^{\top} R w$ and $κ=\|w\|^2$, where $w$ denotes the attention pooling weights and $R$ the positional correlation matrix. An aftermath of our analysis demonstrates that the optimal attention weights maximizing the signal-to-noise ratio $α/κ$ are given by the (normalized) top eigenvector of $R$, and we show (as a particular case of our analysis) that parameter-free causal self-attention with $τ/d$ score scaling yields deterministic harmonic weights that improve signal recovery over mean pooling whenever early tokens carry more signal. Extensive simulations confirm sharp agreement between theory and finite-dimensional experiments.

Mohamed El Amine Seddik, Suei-Wen Chen, Soufiane Hayou et al.

The phenomenon of model collapse, introduced in (Shumailov et al., 2023), refers to the deterioration in performance that occurs when new models are trained on synthetic data generated from previously trained models. This recursive training loop makes the tails of the original distribution disappear, thereby making future-generation models forget about the initial (real) distribution. With the aim of rigorously understanding model collapse in language models, we consider in this paper a statistical model that allows us to characterize the impact of various recursive training scenarios. Specifically, we demonstrate that model collapse cannot be avoided when training solely on synthetic data. However, when mixing both real and synthetic data, we provide an estimate of a maximal amount of synthetic data below which model collapse can eventually be avoided. Our theoretical conclusions are further supported by empirical validations.

Reda Alami, Abdalgader Abubaker, Mastane Achab et al.

This paper explores the effects of various forms of regularization in the context of language model alignment via self-play. While both reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) require to collect costly human-annotated pairwise preferences, the self-play fine-tuning (SPIN) approach replaces the rejected answers by data generated from the previous iterate. However, the SPIN method presents a performance instability issue in the learning phase, which can be mitigated by playing against a mixture of the two previous iterates. In the same vein, we propose in this work to address this issue from two perspectives: first, by incorporating an additional Kullback-Leibler (KL) regularization to stay at the proximity of the reference policy; second, by using the idea of fictitious play which smoothens the opponent policy across all previous iterations. In particular, we show that the KL-based regularizer boils down to replacing the previous policy by its geometric mixture with the base policy inside of the SPIN loss function. We finally discuss empirical results on MT-Bench as well as on the Hugging Face Open LLM Leaderboard.

Aymane El Firdoussi, Mohamed El Amine Seddik, Soufiane Hayou et al.

Synthetic data has gained attention for training large language models, but poor-quality data can harm performance (see, e.g., Shumailov et al. (2023); Seddik et al. (2024)). A potential solution is data pruning, which retains only high-quality data based on a score function (human or machine feedback). Previous work Feng et al. (2024) analyzed models trained on synthetic data as sample size increases. We extend this by using random matrix theory to derive the performance of a binary classifier trained on a mix of real and pruned synthetic data in a high dimensional setting. Our findings identify conditions where synthetic data could improve performance, focusing on the quality of the generative model and verification strategy. We also show a smooth phase transition in synthetic label noise, contrasting with prior sharp behavior in infinite sample limits. Experiments with toy models and large language models validate our theoretical results.

Hamza Cherkaoui, Malik Tiomoko, Mohamed El Amine Seddik et al.

Bootstrap methods have long been a cornerstone of ensemble learning in machine learning. This paper presents a theoretical analysis of bootstrap techniques applied to the Least Square Support Vector Machine (LSSVM) ensemble in the context of large and growing sample sizes and feature dimensionalities. Leveraging tools from Random Matrix Theory, we investigate the performance of this classifier that aggregates decision functions from multiple weak classifiers, each trained on different subsets of the data. We provide insights into the use of bootstrap methods in high-dimensional settings, enhancing our understanding of their impact. Based on these findings, we propose strategies to select the number of subsets and the regularization parameter that maximize the performance of the LSSVM. Empirical experiments on synthetic and real-world datasets validate our theoretical results.

Aymane El Firdoussi, Mohamed El Amine Seddik

This paper provides theoretical insights into high-dimensional binary classification with class-conditional noisy labels. Specifically, we study the behavior of a linear classifier with a label noisiness aware loss function, when both the dimension of data $p$ and the sample size $n$ are large and comparable. Relying on random matrix theory by supposing a Gaussian mixture data model, the performance of the linear classifier when $p,n\to \infty$ is shown to converge towards a limit, involving scalar statistics of the data. Importantly, our findings show that the low-dimensional intuitions to handle label noise do not hold in high-dimension, in the sense that the optimal classifier in low-dimension dramatically fails in high-dimension. Based on our derivations, we design an optimized method that is shown to be provably more efficient in handling noisy labels in high dimensions. Our theoretical conclusions are further confirmed by experiments on real datasets, where we show that our optimized approach outperforms the considered baselines.

Hugo Lebeau, Mohamed El Amine Seddik, José Henrique de Morais Goulart

We study the estimation of a planted signal hidden in a recently introduced nested matrix-tensor model, which is an extension of the classical spiked rank-one tensor model, motivated by multi-view clustering. Prior work has theoretically examined the performance of a tensor-based approach, which relies on finding a best rank-one approximation, a problem known to be computationally hard. A tractable alternative approach consists in computing instead the best rank-one (matrix) approximation of an unfolding of the observed tensor data, but its performance was hitherto unknown. We quantify here the performance gap between these two approaches, in particular by deriving the precise algorithmic threshold of the unfolding approach and demonstrating that it exhibits a BBP-type transition behavior. This work is therefore in line with recent contributions which deepen our understanding of why tensor-based methods surpass matrix-based methods in handling structured tensor data.

Aymane El Firdoussi, El Mahdi Chayti, Mohamed El Amine Seddik et al.

Fine-tuning has proven to be highly effective in adapting pre-trained models to perform better on new desired tasks with minimal data samples. Among the most widely used approaches are reparameterization methods, which update a target module by augmenting its frozen weight matrix with an additional trainable weight matrix. The most prominent example is Low Rank Adaption (LoRA), which gained significant attention in recent years. In this paper, we introduce a new class of reparameterization methods for transfer learning, designed to enhance the generalization ability of fine-tuned models. We establish the effectiveness of our approach in a high-dimensional binary classification setting using tools from Random Matrix Theory, and further validate our theoretical findings through more realistic experiments, such as fine-tuning LLMs.

Omar El Mansouri, Mohamed El Amine Seddik, Salem Lahlou

Reinforcement learning from human feedback (RLHF) or verifiable rewards (RLVR), the standard paradigm for aligning LLMs or building recent SOTA reasoning models, is highly sensitive to noise from inconsistent or erroneous rewards. Yet, the interaction between such noise and widely used group-based policy optimization methods remains underexplored. We introduce a noise-robust Group Relative Policy Optimization (GRPO) and Done Right GRPO (Dr.GRPO) framework that explicitly models reward corruption as Bernoulli noise. Our method applies noise correction after estimating reward flip probabilities to debias the learning signal, yielding provably unbiased gradient estimates. Theoretical analysis shows that group-based methods inherently mitigate individual-level noise, and our correction strategy amplifies this robustness. Empirically, we observe consistent improvements across math and code tasks when applying our noise correction to standard reward model usage, with particular gains of up to 6.7 percentage points in accuracy on math tasks and 1.5 on code tasks under realistic reward model conditions. This work bridges label-noise correction from supervised learning with modern RLHF, offering both theoretical insights and a practical algorithm for noisy real-world deployment.

Benjamin Dadoun, Soufiane Hayou, Hanan Salam et al.

Deep neural networks are known to suffer from exploding or vanishing gradients as depth increases, a phenomenon closely tied to the spectral behavior of the input-output Jacobian. Prior work has identified critical initialization schemes that ensure Jacobian stability, but these analyses are typically restricted to fully connected networks with i.i.d. weights. In this work, we go significantly beyond these limitations: we establish a general stability theorem for deep neural networks that accommodates sparsity (such as that introduced by pruning) and non-i.i.d., weakly correlated weights (e.g. induced by training). Our results rely on recent advances in random matrix theory, and provide rigorous guarantees for spectral stability in a much broader class of network models. This extends the theoretical foundation for initialization schemes in modern neural networks with structured and dependent randomness.

Mohamed El Amine Seddik, Mastane Achab, Henrique Goulart et al.

In this paper, we propose a nested matrix-tensor model which extends the spiked rank-one tensor model of order three. This model is particularly motivated by a multi-view clustering problem in which multiple noisy observations of each data point are acquired, with potentially non-uniform variances along the views. In this case, data can be naturally represented by an order-three tensor where the views are stacked. Given such a tensor, we consider the estimation of the hidden clusters via performing a best rank-one tensor approximation. In order to study the theoretical performance of this approach, we characterize the behavior of this best rank-one approximation in terms of the alignments of the obtained component vectors with the hidden model parameter vectors, in the large-dimensional regime. In particular, we show that our theoretical results allow us to anticipate the exact accuracy of the proposed clustering approach. Furthermore, numerical experiments indicate that leveraging our tensor-based approach yields better accuracy compared to a naive unfolding-based algorithm which ignores the underlying low-rank tensor structure. Our analysis unveils unexpected and non-trivial phase transition phenomena depending on the model parameters, ``interpolating'' between the typical behavior observed for the spiked matrix and tensor models.

Mohamed El Amine Seddik, Maxime Guillaud, Romain Couillet

Relying on random matrix theory (RMT), this paper studies asymmetric order-$d$ spiked tensor models with Gaussian noise. Using the variational definition of the singular vectors and values of (Lim, 2005), we show that the analysis of the considered model boils down to the analysis of an equivalent spiked symmetric \textit{block-wise} random matrix, that is constructed from \textit{contractions} of the studied tensor with the singular vectors associated to its best rank-1 approximation. Our approach allows the exact characterization of the almost sure asymptotic singular value and alignments of the corresponding singular vectors with the true spike components, when $\frac{n_i}{\sum_{j=1}^d n_j}\to c_i\in (0, 1)$ with $n_i$'s the tensor dimensions. In contrast to other works that rely mostly on tools from statistical physics to study random tensors, our results rely solely on classical RMT tools such as Stein's lemma. Finally, classical RMT results concerning spiked random matrices are recovered as a particular case.

Abdallah Benzine, Mohamed El Amine Seddik, Julien Desmarais

Most recent person re-identification approaches are based on the use of deep convolutional neural networks (CNNs). These networks, although effective in multiple tasks such as classification or object detection, tend to focus on the most discriminative part of an object rather than retrieving all its relevant features. This behavior penalizes the performance of a CNN for the re-identification task, since it should identify diverse and fine grained features. It is then essential to make the network learn a wide variety of finer characteristics in order to make the re-identification process of people effective and robust to finer changes. In this article, we introduce Deep Miner, a method that allows CNNs to "mine" richer and more diverse features about people for their re-identification. Deep Miner is specifically composed of three types of branches: a Global branch (G-branch), a Local branch (L-branch) and an Input-Erased branch (IE-branch). G-branch corresponds to the initial backbone which predicts global characteristics, while L-branch retrieves part level resolution features. The IE-branch for its part, receives partially suppressed feature maps as input thereby allowing the network to "mine" new features (those ignored by G-branch) as output. For this special purpose, a dedicated suppression procedure for identifying and removing features within a given CNN is introduced. This suppression procedure has the major benefit of being simple, while it produces a model that significantly outperforms state-of-the-art (SOTA) re-identification methods. Specifically, we conduct experiments on four standard person re-identification benchmarks and witness an absolute performance gain up to 6.5% mAP compared to SOTA.

Mohamed El Amine Seddik, Cosme Louart, Mohamed Tamaazousti et al.

This paper shows that deep learning (DL) representations of data produced by generative adversarial nets (GANs) are random vectors which fall within the class of so-called \textit{concentrated} random vectors. Further exploiting the fact that Gram matrices, of the type $G = X^T X$ with $X=[x_1,\ldots,x_n]\in \mathbb{R}^{p\times n}$ and $x_i$ independent concentrated random vectors from a mixture model, behave asymptotically (as $n,p\to \infty$) as if the $x_i$ were drawn from a Gaussian mixture, suggests that DL representations of GAN-data can be fully described by their first two statistical moments for a wide range of standard classifiers. Our theoretical findings are validated by generating images with the BigGAN model and across different popular deep representation networks.

John Lin, Mohamed El Amine Seddik, Mohamed Tamaazousti et al.

We propose a novel learning approach, in the form of a fully-convolutional neural network (CNN), which automatically and consistently removes specular highlights from a single image by generating its diffuse component. To train the generative network, we define an adversarial loss on a discriminative network as in the GAN framework and combined it with a content loss. In contrast to existing GAN approaches, we implemented the discriminator to be a multi-class classifier instead of a binary one, to find more constraining features. This helps the network pinpoint the diffuse manifold by providing two more gradient terms. We also rendered a synthetic dataset designed to help the network generalize well. We show that our model performs well across various synthetic and real images and outperforms the state-of-the-art in consistency.

Mohamed El Amine Seddik, Mohamed Tamaazousti, John Lin

A common issue of deep neural networks-based methods for the problem of Single Image Super-Resolution (SISR), is the recovery of finer texture details when super-resolving at large upscaling factors. This issue is particularly related to the choice of the objective loss function. In particular, recent works proposed the use of a VGG loss which consists in minimizing the error between the generated high resolution images and ground-truth in the feature space of a Convolutional Neural Network (VGG19), pre-trained on the very "large" ImageNet dataset. When considering the problem of super-resolving images with a distribution "far" from the ImageNet images distribution (\textit{e.g.,} satellite images), their proposed \textit{fixed} VGG loss is no longer relevant. In this paper, we present a general framework named \textit{Generative Collaborative Networks} (GCN), where the idea consists in optimizing the \textit{generator} (the mapping of interest) in the feature space of a \textit{features extractor} network. The two networks (generator and extractor) are \textit{collaborative} in the sense that the latter "helps" the former, by constructing discriminative and relevant features (not necessarily \textit{fixed} and possibly learned \textit{mutually} with the generator). We evaluate the GCN framework in the context of SISR, and we show that it results in a method that is adapted to super-resolution domains that are "far" from the ImageNet domain.

Youssef Tamaazousti, Hervé Le Borgne, Céline Hudelot et al.

A representation is supposed universal if it encodes any element of the visual world (e.g., objects, scenes) in any configuration (e.g., scale, context). While not expecting pure universal representations, the goal in the literature is to improve the universality level, starting from a representation with a certain level. To do so, the state-of-the-art consists in learning CNN-based representations on a diversified training problem (e.g., ImageNet modified by adding annotated data). While it effectively increases universality, such approach still requires a large amount of efforts to satisfy the needs in annotated data. In this work, we propose two methods to improve universality, but pay special attention to limit the need of annotated data. We also propose a unified framework of the methods based on the diversifying of the training problem. Finally, to better match Atkinson's cognitive study about universal human representations, we proposed to rely on the transfer-learning scheme as well as a new metric to evaluate universality. This latter, aims us to demonstrates the interest of our methods on 10 target-problems, relating to the classification task and a variety of visual domains.