法國農業信貸銀行

Claim-level Uncertainty Quantification (UQ) is a promising approach to mitigate the lack of reliability in Large Language Models (LLMs). We introduce MUCH, the first claim-level UQ benchmark designed for fair and reproducible evaluation of future methods under realistic conditions. It includes 4,873 samples across four European languages (English, French, Spanish, and German) and four instruction-tuned open-weight LLMs. Unlike prior claim-level benchmarks, we release 24 generation logits per token, facilitating the development of future white-box methods without re-generating data. Moreover, in contrast to previous benchmarks that rely on manual or LLM-based segmentation, we propose a new deterministic algorithm capable of segmenting claims using as little as 0.2

Localizing document tampering is extremely challenging, as manipulations are crafted to appear visually consistent and often leave only subtle traces that are nearly invisible to the human eye. In prior work, evaluation has been largely dominated by synthetic benchmarks that closely match the training distribution, and methods have shown steady progress under this setting. However, these gains often translate poorly to human-made forgeries and to cross-domain evaluation, where both the source documents and the tampering pipeline can change, leading to a distribution shift. In addition, since the introduction of the Frequency Perception Head for the discrete cosine transform (DCT) modality, it has become a standard choice, and subsequent work has largely focused on downstream modules and fusion strategies rather than revisiting the backbone itself. To help close this gap in cross-domain performance and improve the DCT backbone design, we propose DiffNet, a relatively simple yet effective RGB–DCT early-fusion architecture driven by two key design choices. First, to ensure that the decoder aggregates multi-scale inconsistency evidence rather than operating on raw, content-heavy activations, we apply a lightweight multi-level discrepancy transformation at the output of each backbone stage, replacing features with magnitude-only responses to learned zero-sum filters. Second, we design an efficient DCT-domain backbone that relies on a lightweight frequency-index-aware DCT–quantization joint embedding. Our approach achieves state-of-the-art performance on cross-domain and human-made document tampering localization, outperforming prior methods by around 30 7× higher throughput than the previous best model.

Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconstruction attack designed to extract sensitive fields from the training data of these models. We attack LayoutLM and BROS architectures, demonstrating that an adversary can perfectly reconstruct up to 4.1% of the fields of the documents used for fine-tuning, including some names, dates, and invoice amounts up to six-digit numbers. When our reconstruction attack is combined with a membership inference attack, our attack accuracy escalates to 22.5%. In addition, we introduce two new end-to-end metrics and evaluate our approach under various conditions: unimodal or bimodal data, LayoutLM or BROS backbones, four fine-tuning tasks, and two public datasets (FUNSD and SROIE). We also investigate the interplay between overfitting, predictive performance, and susceptibility to our attack. We conclude with a discussion on possible defenses against our attack and potential future research directions to construct robust document understanding models.

The decision for a financial institution to accept or deny a loan is based on the probability of a client paying back their debt in time. This probability is given by a model such as a logistic regression, and estimated based on, e.g., the clients’ characteristics, their credit history, the repayment performance. Historically, different models have been developed on different markets and/or credit products and/or addressed population. We show that this amounts to modelling default as a mixture model composed of a decision tree and logistic regression on its leaves (thereafter “logistic regression tree”). We seek to optimise this practice by considering the population to which a client belongs as a latent variable, which we will estimate. After exposing the context, the notations and the problem formalisation, we will conduct estimation using a Stochastic-Expectation-Maximisation (SEM) algorithm. We will finally show the performance on simulated data, and on real retail credit data from [COMPANY], as well as real open-source data.

During the last decade, the increase in computational capacity, the consolidation of new data processing methodologies and the availability of access to new information concerning both individuals and organizations, aided by the widespread internet usage, has increased the development and implementation of artificial intelligence (AI) within companies. The application of AI techniques in the banking sector attracts wide interest as the extraction of information from data is inherent to banks. As matter of fact, for many years now models play a crucial role in several banks processes and are strictly regulated when they drive capital measurement processes. Among banks’ risk models a special role is played by credit ones, as they manage the most relevant risk banks face and are often used in regulatory relevant processes. The new AI techniques, coupled with the usage of novel data, mostly unstructured ones related to borrowers’ behaviors, allow for an improvement of the accuracy of credit risk models, that so far relied on structured internal and external data. This paper takes inspiration from the Position Paper Aifirm 33/2022 and its English published translation (Locatelli, Pepe, Salis (eds), 2022. The paper is focused on literature review regarding the most common AI models in use in credit risk management, also adding a regulatory perspective due to the specific regime banking models are subject when they are used for regulatory purposes. Furthermore, the exploration of forthcoming challenges and future advancements considers a managerial perspective. It aims to uncover how credit risk managers can leverage the new AI toolbox and novel data to enhance the credit risk models’ predictive power, without overlooking the intrinsic problems associated with the interpretability of the results.