This report evaluates PDF-to-Markdown conversion using recent Vision-Language Models (VLMs) on challenging French documents. Document parsing is a critical step for Retrieval-Augmented Generation (RAG) pipelines, where transcription and layout errors propagate to downstream retrieval and grounding. Existing benchmarks often emphasize English or Chinese and can over-penalize benign formatting and linearization choices (e.g., line breaks, list segmentation, alternative table renderings) that are largely irrelevant for downstream use. We introduce a French-focused benchmark of difficult pages selected via model-disagreement sampling from a corpus of 60,000 documents, covering handwritten forms, complex layouts, dense tables, and graphics-rich pages. Evaluation is performed with unit-test-style checks that target concrete failure modes (text presence, reading order, and local table constraints) combined with category-specific normalization designed to discount presentation-only variance. Across 15 models, we observe substantially higher robustness for the strongest proprietary models on handwriting and forms, while several open-weights systems remain competitive on standard printed layouts.
This article deals with an optimization problem arising in industrial mail platforms. To increase the profitability of sorting machines, it is necessary to minimize the number of mail sorting sessions processed on these machines. This problem can be seen as a variant of the vector bin-packing problem, with a particular structure of the vector items. In this paper, we propose a 0-1 linear programming model to deal with small- and medium-size instances of this problem. It demonstrated good performance for certain instance categories. Computational results are reported as well. Copyright (C) 2024 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
In social and economic surveys, it can be difficult to directly reach units of the target population, and indirect sampling is often advocated to solve this issue. In indirect sampling, the sample is drawn from a frame population that is linked to the target population, and estimation of target population pa-rameters is typically achieved through the generalized weight share method (GWSM). This method provides a weight, for every unit of the target popu-lation, that depends on the one hand, on the sampling weights in the frame population and, on the other hand, on the link weights between the frame population and the target population. In the present study, we focus on the situation in which the units from the frame population are linked to one and only one unit from the target population (Many-to-One case). This situation is encountered at the French postal service where addresses are sampled instead of postman rounds. We aim at understanding of the impact of the link weights on the efficiency of the GWSM estimators. We derive variance expressions and optimality results for a large class of sampling designs. Moreover, we note that the Many-to-One case can lead to too many links to observe. We alleviate the problem by introducing an intermediate population and double indirect sampling. The question of the loss of precision in this situation is dis-cussed in detail through theoretical results and simulations. These findings help to explain the loss of precision of double GWSM estimators observed recently at the French postal service.
This paper addresses a mail sorting optimization problem encountered by a major French mail delivery provider La Poste. To simplify the operators’ handling during the sorting process, it is necessary to balance the mail flow between the outputs of each sorting machine. This problem can be considered as a variant of the well-known assembly line balancing problem (ALBP) with a particular structure of precedence constraints. At first, to handle efficiently small- and medium-size academic instances of this problem, we propose a mixed-integer linear programming (MILP) model aiming to minimize the difference between the most and least loaded outputs. Finally, to deal successfully with real industrial large-size instances, we develop an efficient heuristic, inspired by simulated annealing, which focuses on minimizing the nonlinear standard deviation between the output loads. Both approaches have demonstrated a very good overall performance for the respective instance categories. Computational results are reported as well.