Conversational AI deployed at population scale has begun to produce documented psychological harm, emotional bonding, affective dependence and behavioural influence on vulnerable users, with legal filings, regulatory complaints and public safety disclosures involving owners or developers of Large Language Models (LLMs) between 2024 and 2026. This paper investigates whether analytic theology can supply a symbolic logical framework apt to encode moral safeguards inside LLM pipelines. The central thesis has two parts. First, the long human effort of normative rationalisation, from medieval scholastic disputation through von Wright’s deontic logic to contemporary analytic theology and neuro symbolic AI, has produced a repository of symbolic formalisation techniques that prefigures, in form, what bai call constitutional AI, in continuity with Adams. Second, justice, truth, dignity and respect are best treated as metaphysical objects with ontological structure rather than as practical conventions which is the property required for safeguards that resist degradation across long interactions. The contribution targets the engineering of moral computation: a formal language placed between metaphysical content and machine implementation.
Fernando Negrini· International Journal of Bio...· 0 citations
This dataset is an extension of the publication of the Database of Byzantine Book Epigrams (2023). The 2023 publication is a copy of the data used for the live application on dbbe.ugent.be. The present dataset aims to offer a simplified structure of the DBBE data, facilitating computational research and data analysis. The dataset provides structured access to the data behind dbbe.ugent.be, enabling intuitive exploration and reuse. It is based on production PostgreSQL and Elasticsearch instances and is organized around six core concepts. The data is updated weekly to reflect the latest research and additions. Occurrences:Individual epigrams reproduced as faithfully as possible from manuscripts, preserving orthography and punctuation. Types: Normalized or critically established texts linked to one or more Occurrences, enabling comparison across similar Occurrences and providing interpretive metadata such as translations, genres, and subjects. Verses: Verses of Occurrences, grouped in Verse Groups to record textual parallels across Occurrences. Manuscripts: Metadata on the context of manuscripts, including provenance and related documentation. Persons: Information on authors, scribes, editors, and other historical or modern contributors linked to Occurrences, Types, or Manuscripts. Bibliographies: Structured references connecting literature and publications to the relevant entities in the corpus. Exploring the Dataset The dataset can be explored and analyzed in multiple ways: Web-based viewing: Tools such as SQLite Viewer allow users to simply upload the SQLite file and inspect its contents directly in the browser, requiring no technical setup. Graphical interface and export: DB Browser for SQLite offers an intuitive interface for browsing the data and exporting tables to CSV for further analysis without need for scripting based interaction. AI-assisted semantic querying: The database file can be uploaded to a large language model for natural language queries. For example, one could ask: "Give me five occurrences from the eleventh century discussing harbours." To protect sensitive data, consider using self-hosted models or paid, secure AI services rather than free online tools. Custom development: The structured SQLite format provides a robust foundation for programmatic analysis, integration with other tools like datasette, or the development of custom applications tailored to specific research questions. Database schema For a full visual of the database schema, please visit yED live. 1. Occurrences This table stores individual Occurrences (= short epigrams or poems, literally how they have been found in a manuscript, including marks for gaps and missing text.) Columns include: id, created, modified, public_comment, incipit, text_stemmer, text_original, location_in_ms*, completion_date_floor, completion_date_ceiling, palaeographical_info, contextual_info, manuscript_id, title *Note that, in the current version, the location of occurrences within the manuscript is given as plain text (ex. p. 394-395 for pages or f. 18r-18v for folia). For manuscripts that have more than 1 way of numbering pages, the alternative location is marked as f. 14r -- (alt.) p. 27. Related tables: occurrence_person_role: Links Occurrences to Persons, indicating which Role a Person plays in the given Occurrence. Example: Scribe ( = historical person), transcriber (=modern person), contributor (=modern person)... occurrence_genre: genre attributed to this Occurrence (Can be more than 1) occurrence_metre: metre attributed to this Occurrence (Can be more than 1) occurrence_management: Internal information. For example: To do's in the processing of this Occurrence occurrence_acknowledgement: Plain text acknowledgement of people who helped in the publication of this Occurrence. occurrence_text_status: An Occurrence text can have statuses like partially/completely (un)known occurrence_related_occurrence and occurrence_relation_definition: An Occurrence can be related to other Occurrence if (a) some of their verses share Verse Groups or (b) they share a Type. The relationship type is defined in occurrence_relation_definition. This works in one direction: if occurrenceA --> related to --> occurrenceB is set, then occurrenceB --> related to --> occurrenceA is not set. occurrence_keyword: Keywords telling what the Occurrence is about 2. Verses This table contains verse-level information about an Occurrence. Columns include id, occurrence_id, manuscript_id, text, order_in_occurrence, verse_group_id. Verse Groups are groupings of similar verses across occurrences. 3. Types This table contains prototypes of Occurrences. A lot of Occurrences have a high level of similarity. DBBE proposes prototypes for every group of similar Occurrences. Related tables: type_person_role: Links Types to Persons, indicating which Role a Person plays in the given Type. Example: Creator, Translitor, Editor, Contributor, ... type_genre: genre attributed to this Type. More than 1 Genre can be attributed. type_metre: metre attributed to this Type. More than 1 Metre can be attributed. type_management: Internal information. For example: To do's in the processing of this Type type_acknowledgement: Plain text acknowledgement of people who helped in the publication of this Type. type_text_status: Type text can be either completely known or partially unknown type_related_type: Groups of similar Types. The relationship is defined in type_relation_definition. This works in one direction: if typeA --> related to --> typeB is set, then typeB --> related to --> typeA is not set. type_tag: Explains the function of the Type (ex: introducing a subject, making a comment on the content,...). type_occurrence: Occurrences linked to this Type. Note that this is a many-to-many relationship: one occurrence can be linked to several types, one type can have several occurrence linked to it. type_editorial_status: editorial states for types. Currently only critical text / not a critical text. type_keyword: Keywords telling what the type is about 4. Manuscripts This table contains metadata about manuscripts. Related tables: manuscript_person_role: Any possible role a Person could play in the publication of this manuscript. Example: Patron ( = historical person), Illuminator (=historical person), contributor (=modern person)... manuscript_acknowledgement: Plain text acknowledgement of people who helped in the publication of this Manuscript. manuscript_content: Explains what the manuscript is about. Careful: content is a hierarchical table. For example, a manuscript can be about Biblica -> Novum Testamentum. In this table, the lowest leaf (Novum Testamentum) is stored. The parent_id column of the content table can be used to trace the full content. manuscript_identification: Links a manuscript to one or more IDs that were used in canonical works to refer to this manuscript (ex: Diktyon) manuscript_management: Internal information. For example: To do's in the processing of this manuscript manuscript_location: The location where the manuscript was written. Careful: location is a hierarchical table. If a manuscript was written in Brussels, it is linked to Brussels, but via the parent_id column of the location table, you could also see that Brussels is in Belgium. 5. Persons This table contains metadata about persons involved (authors, editors, patrons, etc.). Related tables: person_acknowledgement: Plain text acknowledgement of people who helped in the publication of the information on this (historical) person. person_identification: Used to link persons to canonical IDs set by different authorities. person_management: Internal information. For example: To do's in the processing of this person person_self_designation: Used for scribes: How a scribe describes himself person_office: Used for scribes: The official title of a person. 6. Bibliographies Bibliographies are modelled as concrete entity types, rather than a single table as in the original setup. article book book_chapter blog_post bib_varia: This table is usually avoided but contains entries for which no other bibliographical type exists. online_source phd Each bibliographic entity has: its own table as mentioned above a corresponding _person_role table (ex.: article_person_role: could contain authors, contributors, reviewers, ... for a given article) tables linking to the item the bibliography is about: manuscripts (ex: manuscript_article: contains articles about a given manuscript) occurrence (ex: occurrence_book: contains books about a given Occurrences) persons (ex: person_article: contains articles about - usually historical - persons) type (ex: type_article: contains articles about given Types) Additional structures: journal and journal_issue: Articles may be linked to journals and journal issues. Note that, for now, some of these bibliography tables were added for completeness sake: not every concept (Manuscript / Occurrence / Person / Type) has all types of bibliographies linked to it (online sources, PhDs, etc.). Lookup / Metadata Tables roles — defines roles for persons (ex. Author, Scribe, Contributor, ...) text_status — textual status of Occurrence or Type. (ex. Text completely known, text partially unknown,...) keywords — keywords for Occurrence and Type (ex. Holy Trinity, Seven Sages, Last Judgement, ...) tag — tag for Type: Explains the function of the Type (ex: introducing a subject, making a comment on the content,...). metre — metre classification (ex. Dodecasyllable, Elegiacs,...) genre — genre classification. (ex. Scribe-related epigram, Text-related epigram, Reader-related epigram) management — administrative metadata. (ex. Bibliography to check) acknowledgement — acknowledgement linked to occurrence, manuscripts, type, or persons. (ex. Inf
Kristoffel Demoen, Sofia Belioti, Gilbert Bentein et al.· Ghent University Academic Bi...· 0 citations
# Separating Secrets from Placeholders: A Hybrid CNN-CodeBERT Framework for Three-Class Credential Leakage Detection ## Overview This repository contains the code and dataset for our 3-class credential leakage detection framework, which distinguishes between: - **Class 0:** No Leak- **Class 1:** Genuine Leak- **Class 2:** Placeholder/Weak Leak --- ## Recommended Experiment Order 1. **`main_method/`** — Train and evaluate the proposed hybrid model (4 seeds). This is the primary experiment.2. **`ablation/`** — Run ablation study to evaluate each architectural component (seed 42 only). Requires the same train/val/test splits as the main method.3. **`lolo/`** — Run Leave-One-Language-Out cross-validation using the full dataset. Run once per language (10 runs total).4. **`baselines/keysentinel/`** — Run KEYSENTINEL baseline. Use `--config fair` for the results reported in the paper.5. **`baselines/passfinder/`** — Run PassFinder baseline. All experiments are independent and can be run in any order. The recommended order above follows the paper's research questions (RQ1 → RQ5). --- ## Repository Structure ```├── data/ # Dataset (see data/README.md)│ ├── Sanitized_CCLD_dataset.csv│ └── splits/│ ├── train.csv│ ├── val.csv│ └── test.csv│├── main_method/ # Proposed hybrid model (RQ1, RQ2)├── ablation/ # Ablation study (RQ3)├── lolo/ # Leave-One-Language-Out evaluation (RQ4)├── baselines/│ ├── keysentinel/ # KEYSENTINEL baseline│ └── passfinder/ # PassFinder baseline``` --- ## Experiments | Folder | Description | Data Used ||--------|-------------|-----------|| `main_method/` | Proposed CharCNN + CodeBERT + Adapter model, 4 seeds | train/val/test splits || `ablation/` | 6 architectural variants, seed 42 | train/val/test splits || `lolo/` | Leave-One-Language-Out cross-validation | Full dataset || `baselines/keysentinel/` | KEYSENTINEL adapted for snippet-level 3-class task | train/val/test splits || `baselines/passfinder/` | PassFinder adapted for snippet-level 3-class task | train/val/test splits | Each folder contains its own `README.md` with setup and run instructions. --- ```bibtex@article{baby2026separating, title={Separating Secrets from Placeholders: A Hybrid CNN-CodeBERT Framework for Three-Class Credential Leakage Detection}, author={Baby, Maksuda Bilkis and Shah, Khushika and Liang, Naiyue and Zhang, Lei}, journal={arXiv preprint arXiv:2605.31520}, year={2026}}``` ## Acknowledgements Parts of the writing and implementation in this prototype were developed with assistance from Large Language Models (LLMs). These tools were used solely as helpers for tasks such as brainstorming, drafting code snippets, and refining text. All core ideas, research design, experimental decisions, and contributions in this work originate from the author(s). All code and experiments were executed, inspected, and validated by the author(s). Any errors or omissions remain the responsibility of the author(s).
Mst Maksuda Bilkis Baby, Khushika Shah, 梁乃跃 et al.· Zenodo (CERN European Organi...· 0 citations
The file attached is the replication package for the experiments presented in the paper “Generalizing Large Language Models’ Hyperparameters and Prompting for Code Generation: Removing the Blindspot”. This package contains the prompts, generated code for case study, and evaluation results.
Generative AI offers new opportunities for individualized and adaptive learning, e.g., through large language model (LLM)-based feedback systems. While LLMs can produce factually correct feedback for relatively straightforward conceptual tasks, delivering high-quality feedback for tasks that require advanced domain expertise—such as physics problem solving—remains a substantial challenge. This study presents the design and implementation of an LLM-based feedback system for physics problem solving grounded in evidence-centered design and reports a first evaluation within the German Physics Olympiad. Participants rated the usefulness and correctness of the generated feedback for each implemented problem. The collected ratings indicate that the feedback was generally perceived as useful and highly correct. However, an in-depth analysis revealed that the feedback contained errors in 20% of cases—errors that often went unnoticed by the students. We discuss the risks associated with uncritical reliance on LLM-based feedback and outline potential directions for generating more adaptive and reliable LLM-based feedback in the future.
Holger Maus, Fabian Kieser, Stefan Petersen et al.· The Physics Educator· 0 citations
With the wide and cross-domain adoption of Large Language Models, it becomes crucial to assess the extent to which the statistical correlations in training data hide subtle and potentially troubling biases. Gender bias in LLMs has been widely investigated from the perspectives of work, hobbies, and emotions typically associated with a specific gender. In this study, we introduce a novel perspective. We investigate whether LLMs can predict an individual’s gender based solely on online shopping histories and whether these predictions are influenced by gender biases and stereotypes. Using a dataset of historical online purchases from users in the United States, we evaluate the ability of | LLMs to predict the buyer’s gender, and we analyse their reasoning and product–gender co-occurrences. Results indicate that while models can infer gender with moderate accuracy, their decisions are often rooted in stereotypical associations between product categories and gender that differ from the actual behaviors of the individuals present in the dataset. Furthermore, explicit instructions to avoid bias reduce the certainty of model predictions, but do not eliminate stereotypical patterns. Our findings highlight the persistent nature of gender biases in LLMs and emphasize the need for robust bias-mitigation strategies as LLMs are currently amplifying the stereotypes analyzed.
Massimiliano Luca, Ciro Beneduce, Bruno Lepri et al.· ACM Transactions on Intellig...· 0 citations
Foundation models derive their value from broad general capability across domains, yet deployment usually rewards specialization - creating a fundamental question for general intelligence: when performance is pushed upward in one region of capability space, is competence elsewhere conserved, redistributed, or destroyed? Synthesizing evidence from continual learning, transfer learning, multi-task optimization, parameter-efficient adaptation, model merging, vision-language adaptation, alignment, and 2023–2026 large-language-model studies, we find that the evidence rejects a universal specialization tax: domain-adaptive pretraining can produce positive transfer, whereas sequential fine-tuning, narrow supervised adaptation, and conflicting objectives can cause catastrophic forgetting, feature distortion, degraded zero-shot transfer, weakened instruction following, or loss of safety behavior depending on task relatedness, update locality, data mixture, optimization geometry, and model capacity. We propose the Generality-Specialization Frontier (GSF), a deployment-oriented framework that treats specialization gain and transfer retention as a Pareto problem, introducing distance-stratified transfer evaluation, invariant retention tests, worst-case regression reporting, and a normalized transfer-elasticity measure. We further propose G-S Bench, an evaluation protocol comparing full fine-tuning, replay, parameter-efficient updates, modular routing, weight interpolation, and non-parametric alternatives under matched target gains, concluding that becoming exceptional at one domain reduces transfer elsewhere only when specialization overwrites shared representations faster than the system preserves broadly useful structure - meaning the tradeoff is an architectural and optimization choice rather than an inevitable law of intelligence.
Sahir Maharaj· Zenodo (CERN European Organi...· 0 citations
Eleven free-tier conversational AI assistants (ChatGPT, Google Gemini, Microsoft Copilot, Meta AI, Perplexity, Grok, Mistral Le Chat, DeepAI, DuckDuckGo AI, DeepSeek, Kimi) were each given the same 24-item cancer prompt library, in a fresh session, verbatim, with no follow-up turns. Responses were collected 21-24 July 2026 from a device in Bangladesh. That gives 11 x 24 = 264 responses. Every response was scored on six 0-5 domains by two reviewers with the platform masked; scores more than one point apart were adjudicated to consensus, and the two safety domains (D2, D6) were resolved to the more conservative value: D1 clinical correctness D2 patient safety D3 care-seeking appropriateness D4 LMIC feasibility D5 lay clarity D6 absence of harm composite = sum of D1..D6, range 0-30 Prompt library structure (4 domains, 24 items): S1-S6 cancer symptom interpretation SC1-SC6 cancer screening guidance T1-T6 treatment decision support L1-L6 LMIC-specific resource-constrained scenarios Within S, SC and T the odd-numbered item is the "standard" version and the even-numbered item is the "LMIC-paired" version of the same clinical topic, so there are 9 matched standard/LMIC pairs. L1-L6 are LMIC-specific and have no standard counterpart. Net: 9 standard + 9 LMIC-paired + 6 LMIC-specific = 24.
Md Shohidullah Samrat, Mst. Jarfin Begum Jyoti, Abid Hasan Rafi· Zenodo (CERN European Organi...· 0 citations
INFRA-COMPASS is an innovative software tool utilizing Large Language Models (LLMs) to systematically compile and update an inventory of state and local codes and ordinances related to energy infrastructure. Given the multitude of jurisdictions involved, automating the inventory process through AI is crucial for monitoring and understanding the implementation of these ordinances. The inventory aids in improving and expediting the permitting process in several ways. It enhances transparency for stakeholders, including energy developers, government officials, and communities, by detailing existing siting and permitting requirements. It also facilitates comparisons of regulatory variations across jurisdictions and over time, linking these to permit timelines, technical assistance efforts, and deployment rates to identify best practices. INFRA-COMPASS also provides insights for jurisdictions revising their rules and helps developers conduct comparative analyses to better assess energy development opportunities. Furthermore, INFRA-COMPASS improves the accuracy of energy modeling tools by ensuring they accurately reflect current policies.
Pavlo Pinchuk, Guilherme Castelão, Anthony Lopez et al.· Zenodo (CERN European Organi...· 0 citations
This research focuses on enterprise profiling in scenarios where large volumes of diverse texts—such as registration records, annual reports, news articles, and bidding notices—are continuously generated. Instead of relying solely on a single data representation or classification model, we developed a comprehensive natural language processing (NLP) pipeline for extracting key information and identifying industries. The pipeline consists of several steps. First, we use a BERT-BiLSTM-CRF model to identify core e/nterprise entities. Then, we combine TF-IDF with BERT embeddings to create a hybrid feature scheme that captures both lexical cues and contextual semantics. To address the challenge of imbalanced industry labels, we apply SMOTE in the dense semantic space and pair it with Focal Loss to enhance learning for minority classes. Additionally, we introduce a Stacking strategy to integrate outputs from different models, making predictions more stable. Tests on a self-compiled dataset covering ten national economic sectors and about 50,000 enterprises show that our method achieves a macro-F1 score of 95.4%. It outperforms traditional machine learning baselines and single deep learning models, offering more reliable recognition for minority classes. These results suggest that our framework is well-suited for applications such as supply chain partner discovery, industrial mapping, and targeted investment promotion.
Xin-Yi Xu· Applied and Computational En...· 0 citations
Eleven free-tier conversational AI assistants (ChatGPT, Google Gemini, Microsoft Copilot, Meta AI, Perplexity, Grok, Mistral Le Chat, DeepAI, DuckDuckGo AI, DeepSeek, Kimi) were each given the same 24-item cancer prompt library, in a fresh session, verbatim, with no follow-up turns. Responses were collected 21-24 July 2026 from a device in Bangladesh. That gives 11 x 24 = 264 responses. Every response was scored on six 0-5 domains by two reviewers with the platform masked; scores more than one point apart were adjudicated to consensus, and the two safety domains (D2, D6) were resolved to the more conservative value: D1 clinical correctness D2 patient safety D3 care-seeking appropriateness D4 LMIC feasibility D5 lay clarity D6 absence of harm composite = sum of D1..D6, range 0-30 Prompt library structure (4 domains, 24 items): S1-S6 cancer symptom interpretation SC1-SC6 cancer screening guidance T1-T6 treatment decision support L1-L6 LMIC-specific resource-constrained scenarios Within S, SC and T the odd-numbered item is the "standard" version and the even-numbered item is the "LMIC-paired" version of the same clinical topic, so there are 9 matched standard/LMIC pairs. L1-L6 are LMIC-specific and have no standard counterpart. Net: 9 standard + 9 LMIC-paired + 6 LMIC-specific = 24.
Md Shohidullah Samrat, Mst. Jarfin Begum Jyoti, Abid Hasan Rafi· Zenodo (CERN European Organi...· 0 citations
Abstract The explosive growth of large-scale video archives from surveillance networks, online platforms, and personal devices has made efficient and semantically rich video retrieval a critical challenge. Existing approaches based on deep multimodal embeddings have significantly improved retrieval accuracy. However, they often lack scalability, modularity, and system-level integration with indexing and metadata management. In this work, we present a modular and scalable pipeline for semantic video indexing and retrieval, tailored to person-centric search. The proposed architecture decouples a web-based front-end from a back-end organized into two pipelines. The indexing pipeline performs video chunking, person detection and tracking, crop selection, metadata enrichment, and semantic vectorization. The retrieval pipeline supports visual, textual, and hybrid queries, including face-based matching. The system leverages YOLO11 and BoT-SORT for real-time person detection and tracking, SigLIP2 for multilingual vision-language embeddings, and InsightFace for face recognition, storing all representations in a vector database with rich, traceable metadata. We further fine-tune the SigLIP2-SO400M-Patch14-384 checkpoint on a curated mixture of person-centric image-text datasets and evaluate the resulting model on the RSTPReid benchmark. Experimental results show that our approach achieves state-of-the-art performance on the RSTPReid benchmark under the considered setting. In particular, it achieves competitive Recall@k performance with respect to recent text-based person search methods and significantly improves mean Average Precision, reaching 0.68 against a best competing value of 0.54.
Rashid Hmaidan, Sebastiano Milardo, Irene Donato et al.· Multimedia Tools and Applica...· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
