Our Work

RIQE (Radiology Image Quality Evaluator) is a NIQE-style no-reference image quality model fitted on full-dose computed tomography instead of natural photographs, released with its validation and with …

RIQE (Radiology Image Quality Evaluator) is a NIQE-style no-reference image quality model fitted on full-dose computed tomography instead of natural photographs, released with its validation and with everything needed to rebuild it from public data.The artefact carries the intensity mapping that defines the model (a fixed Hounsfield window, part of the specification), the hyperparameters and the criterion that selected them, the identity of every slice used with its SHA-256, and the identity of the code that produced it. A verification script refits from the same public data and compares.On held-out patients RIQE ranks degraded images (simulated reduced dose, added noise, blur) worse than their source, but its preferences among filtered images do not follow the signal of small inserted lesions. It is suited to paired ranking and should not be the sole criterion for selecting, comparing or tuning denoisers. The score is a distance from the declared model, not a judgement of diagnostic quality, and is not comparable with NIQE values in the literature.Fitted on the CC BY 4.0 Chest and Liver subsets of the TCIA collection LDCT-and-Projection-data (DOI 10.7937/9npb-2637, version 7). Algorithm implemented from Mittal, Soundararajan and Bovik, IEEE SPL 20(3):209-212, 2013; no third-party code incorporated. Code: MIT licence; model, manifests and results: CC BY 4.0.

The mechanism screened in a computational electrocatalysis study is still usually written by hand, and assumed to consist only of proton-coupled electron transfers. We present HELLMM, a pipeline in wh…

The mechanism screened in a computational electrocatalysis study is still usually written by hand, and assumed to consist only of proton-coupled electron transfers. We present HELLMM, a pipeline in which a language model selects transformation rules, enumerates surface intermediates and filters them for chemical plausibility, while deterministic graph constraints and non-LLM atomistic and computational-hydrogen-electrode calculations evaluate what survives. Repeating it 23 times across three model checkpoints and four catalyst systems separates two sources of error. Conditional on a mechanism, the computed quantities reproduce to 0.5 mV; the run-to-run spread comes instead from the model selecting between competing mechanisms that are individually valid, partitioning exactly four and four on Pt(111) ORR and five and five on CO₂RR. The same energetics miss the literature Pt(111) ORR limiting potential by 0.5 V, an order of magnitude more than mechanism selection moves it and three more than the within-mechanism spread, so reproducibility and accuracy are separable problems. The audit also exposes failures that aggregate metrics hide: a descriptor blind to costly chemical steps, and prompt examples that bias enumeration.

Abstract—The Journal of Open Source Software (JOSS) is a diamond open-access journal with an open, collaborative, peer-review model for reusable open-source research software pack…

Abstract—The Journal of Open Source Software (JOSS) is a diamond open-access journal with an open, collaborative, peer-review model for reusable open-source research software packages and associated short papers. Over its ten-year history, the journal has had a consistent aim of improving research software packages and providing scholarly credit to the people behind them. However, research culture and technology have changed over this time, most dramatically in the last few years with the rise of the use of Generative AI for both text and then software. This has led to both positive and negatives changes for JOSS, which are described in this talk.

Este texto, de natureza ensaística, objetivou pensar a escola pública a partir de Paulo Freire. Para materializar seu objetivo, guiou-se em três apostas, as quais intentam a escola pública como espaço…

Este texto, de natureza ensaística, objetivou pensar a escola pública a partir de Paulo Freire. Para materializar seu objetivo, guiou-se em três apostas, as quais intentam a escola pública como espaço de formação crítica, ético-política e humanizadora das pessoas. São elas: i. A escola pública por via de outra política educacional; ii. A escola pública e outras relações com a prática social; e iii. A escola pública e a construção de outros currículos. Tais apostas não se configuram como soluções definitivas, mas como caminhos possíveis para o fortalecimento da escola enquanto espaço de resistência e emancipação humana. Reafirma-se, por fim, que pensar a escola pública a partir de Paulo Freire é também assumir um compromisso ético-político com a transformação do mundo.

Replication Package The package contains the datasets, source code, validation results, inter-rater agreement results, and generated outputs used throughout the study "Architectural Degradation: How t…

Replication Package The package contains the datasets, source code, validation results, inter-rater agreement results, and generated outputs used throughout the study "Architectural Degradation: How to Measure and to Remediate". It helps to reproduce the literature screening, full-text extraction, LLM-based analysis, validation, consolidation of the final categories, agreement analysis, and the analyses reported in the paper. Note: All hyperlinks referring to files in this replication package are intended to be used after downloading the package. Local file links may not work when browsing the package online. License All generated data is provided under DATA_LICENSE Creative Commons 4.0 Attribution License. All scripts are provided under SCRIPT_LICENSE MIT License. Package Structure Approaches for Architectural degradation/ │ ├── README.md ├── INSTALL.md ├── requirements.txt ├── DATA_LICENSE ├── SCRIPT_LICENSE ├── Final Full read and analysis (architecture degradation measure and remediation).xlsx ├── Validation architecture degradation measure and remediation.xlsx │ ├── Code and outputs/ │ ├── Title Abstract Round Code and output/ │ ├── T-A Validation code and Outputs/ │ ├── Full Read Code/ │ ├── Full Read Round Validation Outputs/ │ ├── Final round code and outputs/ │ └── Final Round Validation code and outputs/ │ ├── Inter Agreement(Gwet AC & Cohen K)/ └── Code for figures and tables/ Data Files 1. Final Full read and analysis (architecture degradation measure and remediation).xlsx This workbook contains the study process, screening results, extracted full-text data, final coded datasets, and analysis results used in the study. Workbook Contents Sheet Name Description Contents Provides an overview of the workbook contents. Process Documents the study workflow and the procedures followed for screening, extraction, coding, and analysis. White (titleabstract) Contains the title and abstract screening results for the white literature (peer-reviewed studies). Grey(titleabstract) Contains the title and abstract screening results for the grey literature. Full read (Accepted) Contains the extracted information from the studies accepted for full-text analysis. Tools Contains the final consolidated tools identified for measuring architectural degradation. Approaches to Measure Contains the final consolidated approaches identified for measuring architectural degradation. Metrics Contains the final consolidated metrics identified for measuring architectural degradation. Approches to remediate Contains the final consolidated approaches identified for remediating architectural degradation. 2. Validation architecture degradation measure and remediation.xlsx This workbook contains the validation datasets and summary statistics used to evaluate the reliability and consistency of the screening, extraction, and final coding stages. Workbook Contents Sheet Name Description Contents Provides an overview of the validation workbook. Stats Contains summary statistics and validation results used in the study. Title abstract (validation) Contains the validation data for the title and abstract screening stage. Approaches to measure (final round validation) Contains the final-round validation results for approaches to measure architectural degradation. tools (Final round validation) Contains the final-round validation results for tools. metrics (final round validation) Contains the final-round validation results for metrics. Approaches to remediate (final round validation) Contains the final-round validation results for remediation approaches. Approaches to measure (full read) Contains the full-read validation results for approaches to measure architectural degradation. Tools (full read) Contains the full-read validation results for tools. Metrics (full read) Contains the full-read validation results for metrics. Remediation Approches (full read) Contains the full-read validation results for remediation approaches. Code and Outputs The Code and outputs directory contains the source code, intermediate data, validation scripts, and generated outputs used throughout the study. Folder Description Title Abstract Round Code and output Contains the code and outputs used for title and abstract screening of the white and grey literature. T-A Validation code and Outputs Contains the code and outputs used to validate the title and abstract screening results. Full Read Code Contains the code and outputs used for full-text extraction and identification of measurement approaches, tools, metrics, and remediation approaches. Full Read Round Validation Outputs Contains the outputs produced by the independent LLM validators for the full-read extraction stage. Final round code and outputs Contains the multi-round consolidation code, intermediate outputs, and final outputs for approaches to measure, tools, metrics, and remediation approaches. Final Round Validation code and outputs Contains the code and outputs used to validate the final consolidated results. Inter Agreement (Gwet AC & Cohen K) Contains the code and outputs used to calculate Cohen's kappa and Gwet's AC1 agreement statistics. Code for figures and tables Contains the code, input data, and generated outputs used to create the figures and tables reported in the paper, including taxonomy summaries, LaTeX tables, TikZ files, figures, and Excel outputs. Replication of the Results This section describes the procedure for reproducing the title and abstract screening, full-text analysis, LLM-based extraction and validation, final-round consolidation, agreement analysis, and final results reported in the study. The workflow uses Marco-o1 as the primary model and Mistral, Qwen, and Llama as independent validation models. The computationally intensive LLM experiments were originally executed on a supercomputer. Setting Up the Environment Follow the instructions in INSTALL.md to configure the required software environment. The required Python dependencies are listed in requirements.txt. After configuring the environment, execute the replication stages below in the specified order. 1. Title and Abstract Screening Use the scripts and outputs provided in: Code and outputs/Title Abstract Round Code and output/ The scripts process the retrieved white and grey literature and generate the title and abstract screening decisions. The corresponding results are provided in the White (titleabstract) and Grey(titleabstract) sheets of: Final Full read and analysis (architecture degradation measure and remediation).xlsx Studies accepted during screening proceed to validation and full-text analysis. 2. Title and Abstract Validation Use the validation code and outputs provided in: Code and outputs/T-A Validation code and Outputs/ The title and abstract screening results are independently validated using Mistral, Qwen, and Llama. Human validation is performed on the selected validation sample according to the validation procedure described in the paper. The resulting validation data are provided in the Title abstract (validation) sheet of: Validation architecture degradation measure and remediation.xlsx The corresponding agreement analysis is available under: Inter Agreement(Gwet AC & Cohen K)/Title Abstract(GWET AC and Cohen K)/ 3. Full-Text Analysis with Marco-o1 Use the code and outputs provided in: Code and outputs/Full Read Code/ Marco-o1 is used as the primary model to analyze the accepted studies and extract information concerning approaches to measure architectural degradation, tools, metrics, and approaches to remediate architectural degradation. For the original supercomputer execution, the general command structure is: sbatch <SBATCH_SCRIPT> "MODEL_NAME" <PORT> <OUTPUT_FILE> The extracted full-read data are consolidated in the Full read (Accepted) sheet of: Final Full read and analysis (architecture degradation measure and remediation).xlsx Note: Where full-text publications are not distributed with the replication package, researchers must obtain the corresponding papers independently before reproducing PDF-based extraction. 4. Full-Read Validation After the primary-model extraction, the full-read results are independently validated using Mistral, Qwen, and Llama. The generated validation outputs are provided in: Code and outputs/Full Read Round Validation Outputs/ The resulting validation data are consolidated in the corresponding full-read sheets of: Validation architecture degradation measure and remediation.xlsx These include validation for: Approaches to measure architectural degradation Tools Metrics Remediation approaches The corresponding Cohen's kappa and Gwet's AC1 analyses are provided in the relevant folders under: Inter Agreement(Gwet AC & Cohen K)/ 5. Final-Round Analysis After completing the full-read extraction and validation, the extracted information is consolidated through additional analysis rounds. The code, intermediate outputs, and final outputs are provided in: Code and outputs/Final round code and outputs/ The directory contains separate analysis pipelines for: Approaches to Measure Tools Metrics Remediation The scripts progressively consolidate the extracted items into the final categories used in the study. The resulting final datasets are provided in the Tools, Approaches to Measure, Metrics, and Approches to remediate sheets of: Final Full read and analysis (architecture degradation measure and remediation).xlsx 6. Final-Round Validation The final consolidated results are independently validated using Mistral, Qwen, and Llama, together with human validation according to the procedure described in the paper. The validation code and outputs are provided in: Code and outputs/Final Round Validation code and outputs/ Separate validation outputs are provided for approaches to measure, tools, metrics, and remediation approaches. The resulting validation data are consolidated in the corresponding final round validation sheets of: Validation architecture degradation measure and remediation.xlsx These validated results constitute the final coded data used for the study analysis. 7. Inter-Rater Agreement Analysis The scripts and outputs used to calculate agreement between human and LLM validation decisions are provided in: Inter Agreement(Gwet AC & Cohen K)/ The agreement analysis includes Cohen's kappa and Gwet's AC1 for the title/abstract stage, full-read validation, and final-round validation where applicable. The directory contains the input validation datasets together with the generated pairwise agreement and agreement-summary outputs. 8. Data Analysis, Figures, and Tables The final analysis uses the consolidated datasets from: Final Full read and analysis (architecture degradation measure and remediation).xlsx The corresponding analysis results are provided in the AnalysisResults sheet. The code, input data, and generated outputs used to create the figures and tables reported in the paper are provided in: Code for figures and tables/ Reproducing the Figures and Tables The complete figure- and table-generation workflow is contained in: Code for figures and tables/ From the root of the replication package, first install the required dependencies: python3 -m pip install -r requirements.txt Then enter the analysis directory and run the complete pipeline: cd "Code for figures and tables" python scripts/pipeline.py The pipeline runs the taxonomy construction, figure generation, path analysis, LaTeX table generation, TikZ generation, and Excel-output generation in sequence. The generated outputs are stored under: Code for figures and tables/results/current/ This directory contains the generated figures, tables, TikZ files, and Excel outputs used for the final analysis. The supplied consolidated data allow this stage to be reproduced independently without rerunning the LLM-based extraction and validation stages. Local Replication Without a Supercomputer Researchers without access to a SLURM-based supercomputer can execute a smaller-scale version of the LLM pipeline locally using an OpenAI-compatible vLLM endpoint, provided that sufficient computational resources are available. Detailed environment requirements and configuration instructions are provided in INSTALL.md. For example, Marco-o1 can be served using: python -m vllm.entrypoints.openai.api_server \ --model AIDC-AI/Marco-o1 \ --port 8000 \ --tensor-parallel-size 1

Téléchargements / Downloads01_Deligne_EN.pdf02_Deligne_EN.tex03_Deligne_Source.zip04_Deligne_FR.pdf05_Deligne_FR.tex06_DELIGNE_PROVENANCE_AUDIT_D016_REV26_20260930.zip07_README_FR_EN.mdPierre Deligne …

Téléchargements / Downloads01_Deligne_EN.pdf02_Deligne_EN.tex03_Deligne_Source.zip04_Deligne_FR.pdf05_Deligne_FR.tex06_DELIGNE_PROVENANCE_AUDIT_D016_REV26_20260930.zip07_README_FR_EN.mdPierre Deligne — 39 travaux / 39 papers Français Édition cumulative du 30 septembre 2026 : D001–D023, D025–D031, D033–D036, D038–D040, D042–D043. D016 est intégré à la révision 26. D024, D032, D037 et D041 ne sont pas inclus dans cette édition ; cette exclusion ne signifie pas absence de fichiers retournés. 1. Lecteur anglais — 1173 pages 2. LaTeX anglais intégral 3. Sources et dépendances complètes 4. Lecteur français — 1188 pages 5. LaTeX français intégral 6. Provenance consolidée Les fichiers LaTeX cumulatifs contiennent tous les corps de texte, et non de simples appels à des sources non publiées. L'archive de sources contient les figures, styles, sources individuelles et instructions nécessaires à la reproduction. Les témoins et l'appareil éditorial restent distincts des lecteurs. Comparaison de D016 et intégration initiale : OpenAI Codex — GPT-6 Sol, effort Ultra. Nouvelle vérification ciblée et réparations du conditionnement, du lanceur de compilation et des titres courants français de D017 : OpenAI Codex — GPT-6.1 Sol, effort Ultra. Les attributions antérieures restent documentées dans les éditions et la provenance. Aucune traduction, correction ni relecture humaine n'est revendiquée. La nouvelle vérification porte sur D016, pages physiques 82–97 et modifications de la révision 25 à 26, son intégration cumulative, le conditionnement des sources et la reproductibilité des PDF. Elle n'est pas une nouvelle lecture intégrale des 39 travaux. Aucune nouvelle erreur mathématique ni erreur dans le corps de traduction n'a été trouvée dans ce périmètre. Les réparations concrètes sont documentées dans la provenance. Le corps et la géométrie des 1188 pages françaises sont inchangés par la réparation des 52 titres courants de D017 ; les dix avertissements de débordement hérités ont été vérifiés sans rognage observé. D032 demeure un chantier local distinct, non publié ici. Dépôt primaire des lecteurs et sources. Conservation DOI et provenance. Le PDF reste l'aperçu de lecture ; le LaTeX et les sources éditables sont des livrables primaires. Les éditions antérieures restent accessibles. English Cumulative edition dated 30 September 2026: D001–D023, D025–D031, D033–D036, D038–D040, D042–D043. D016 revision 26 is integrated. D024, D032, D037 and D041 are not included in this edition; their exclusion does not mean that returned files are absent. 1. English reader — 1173 pages 2. Complete English LaTeX 3. Complete sources and dependencies 4. French reader — 1188 pages 5. Complete French LaTeX 6. Consolidated provenance The cumulative LaTeX files contain all paper bodies, not thin masters pointing to unpublished inputs. The source archive includes figures, styles, individual sources and reproduction instructions. Witnesses and editorial apparatus remain separate from readers. D016 comparison and initial integration: OpenAI Codex — GPT-6 Sol, Ultra effort. Fresh scoped verification and repairs to packaging, the build launcher and D017's French running headers: OpenAI Codex — GPT-6.1 Sol, Ultra effort. Earlier attribution remains documented in the editions and provenance. No human translation, correction or review is claimed. The fresh review covers D016 physical pages 82–97 and revision 25-to-26 changes, cumulative integration, source packaging and PDF reproducibility. It is not a new full reading of all 39 papers. No new mathematical or translation-body errors were found in that scope. Concrete repairs are documented in the provenance. The header repair leaves body text and geometry unchanged on all 1188 French pages; ten inherited overfull warnings were checked without observed clipping. D032 remains separate local work, not published here. Primary reader/source repository. DOI preservation and provenance. The PDF remains the reading preview; LaTeX and editable sources are primary deliverables. Previous editions remain accessible.Sources / sources : GitHub — c33d13d9b59cf75ea6772d813d48d17723e83d75

Initial release of XEmoGPT, an explainable multimodal emotion recognition framework with cue-level perception and reasoning. This repository accompanies the paper "XEmoGPT: An Explainable Multimo…

Initial release of XEmoGPT, an explainable multimodal emotion recognition framework with cue-level perception and reasoning. This repository accompanies the paper "XEmoGPT: An Explainable Multimodal Emotion Recognition Framework with Cue-Level Perception and Reasoning"

L’étude vise à analyser trois problématiques reliées à la dynamique et à la morphologie des vallées : l’érosion, la sédimenta…

L’étude vise à analyser trois problématiques reliées à la dynamique et à la morphologie des vallées : l’érosion, la sédimentation et leurs rôles économique et biologique. On part du postulat que la zone d’accumulation permet une mobilisation des sédiments propices à l’économie rurale et au fonctionnement optimal des écosystèmes naturels. La méthodologie privilégie l’enquête morphologique et la cartographie de l’occupation du sol. Les résultats montrent d’importants dépôts solides réalisés en grande partie en début de saison humide : en mars-avril les colluvions grossières dans les zones d’accumulation atteignent 5460,1 kg/h par hectare en milieux indurés et 7317,6 kg/ha en milieux non indurés, soit environ trois fois les valeurs cumulées des dépôts au cours du reste de l’année. De même, l’érosion des berges apparaît très importante : le bilan sédimentaire est de -4,4 mm en milieu birrimien et -5,5 mm sur granite. Pendant ce temps, la capacité des marigots se révèle relativement faible : les débits solides sont mesurés à 1284 kg/s en moyenne dans les vallées dominées par les interfluves granitiques et à 688 kg/s dans les vallées en aval des versants sur roches birrimiennes. Il en résulte une puissante topographie de nappe et une structuration du système de vallées en types de milieu. Les secteurs caractérisés par des dépôts grossiers d’origine colluviale et alluviale, présentent une occupation dominée par des formations forestières. À l’inverse, les secteurs d’accumulation, constitués de matériaux plus fins et offrant de meilleures possibilités agronomiques, sont majoritairement affectés aux activités agricoles.

This presentation was held at SolarPACES 2026 (15-18 September 2026) by Marco Colimbi (Politecnico di Milano)