António Loison

Sacha Muller, António Loison, Bilel Omrani et al.

Retrieval-Augmented Generation (RAG) has emerged as a common paradigm to use Large Language Models (LLMs) alongside private and up-to-date knowledge bases. In this work, we address the challenges of using LLM-as-a-Judge when evaluating grounded answers generated by RAG systems. To assess the calibration and discrimination capabilities of judge models, we identify 7 generator failure modes and introduce GroUSE (Grounded QA Unitary Scoring of Evaluators), a meta-evaluation benchmark of 144 unit tests. This benchmark reveals that existing automated RAG evaluation frameworks often overlook important failure modes, even when using GPT-4 as a judge. To improve on the current design of automated RAG evaluation frameworks, we propose a novel pipeline and find that while closed models perform well on GroUSE, state-of-the-art open-source judges do not generalize to our proposed criteria, despite strong correlation with GPT-4's judgement. Our findings suggest that correlation with GPT-4 is an incomplete proxy for the practical performance of judge models and should be supplemented with evaluations on unit tests for precise failure mode detection. We further show that finetuning Llama-3 on GPT-4's reasoning traces significantly boosts its evaluation capabilities, improving upon both correlation with GPT-4's evaluations and calibration on reference situations.

Manuel Faysse, Patrick Fernandes, Nuno M. Guerreiro et al. · meta-ai

We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware. To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources. To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81 % of the transparency criteria, far beyond the scores of even most open initiatives. This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.

António Loison, Quentin Macé, Antoine Edy et al.

Retrieval-Augmented Generation (RAG) pipelines must address challenges beyond simple single-document retrieval, such as interpreting visual elements (tables, charts, images), synthesizing information across documents, and providing accurate source grounding. Existing benchmarks fail to capture this complexity, often focusing on textual data, single-document comprehension, or evaluating retrieval and generation in isolation. We introduce ViDoRe v3, a comprehensive multimodal RAG benchmark featuring multi-type queries over visually rich document corpora. It covers 10 datasets across diverse professional domains, comprising ~26,000 document pages paired with 3,099 human-verified queries, each available in 6 languages. Through 12,000 hours of human annotation effort, we provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers. Our evaluation of state-of-the-art RAG pipelines reveals that visual retrievers outperform textual ones, late-interaction models and textual reranking substantially improve performance, and hybrid or purely visual contexts enhance answer generation quality. However, current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding. To encourage progress in addressing these challenges, the benchmark is released under a commercially permissive license at https://hf.co/vidore.

Théo Combey, António Loison, Maxime Faucher et al.

Neural network classifiers (NNCs) are known to be vulnerable to malicious adversarial perturbations of inputs including those modifying a small fraction of the input features named sparse or $L_0$ attacks. Effective and fast $L_0$ attacks, such as the widely used Jacobian-based Saliency Map Attack (JSMA) are practical to fool NNCs but also to improve their robustness. In this paper, we show that penalising saliency maps of JSMA by the output probabilities and the input features of the NNC allows to obtain more powerful attack algorithms that better take into account each input's characteristics. This leads us to introduce improved versions of JSMA, named Weighted JSMA (WJSMA) and Taylor JSMA (TJSMA), and demonstrate through a variety of white-box and black-box experiments on three different datasets (MNIST, CIFAR-10 and GTSRB), that they are both significantly faster and more efficient than the original targeted and non-targeted versions of JSMA. Experiments also demonstrate, in some cases, very competitive results of our attacks in comparison with the Carlini-Wagner (CW) $L_0$ attack, while remaining, like JSMA, significantly faster (WJSMA and TJSMA are more than 50 times faster than CW $L_0$ on CIFAR-10). Therefore, our new attacks provide good trade-offs between JSMA and CW for $L_0$ real-time adversarial testing on datasets such as the ones previously cited. Codes are publicly available through the link https://github.com/probabilistic-jsmas/probabilistic-jsmas.