Daily Digest — Sep 25

Ontology-Mediated Neurosymbolic Constraint Acquisition from Multiple Stakeholders

• An OWL configuration ontology mediates between neural constraint sources and downstream consumers to capture and formalize requirements from multiple stakeholders. • The framework employs LLM assistants to elicit soft stakeholder preferences while utilizing description logic to identify unsatisfiability against hard hardware specifications.

Clinical Knowledge Graphs for Chest X-Ray Device Reasoning

• An uncertainty-aware clinical knowledge graph successfully represents device instances, tip estimates, and placement assessments for chest radiographs while preserving provenance and temporal links. • The evaluation utilizes predictions from the complete RANZCR CLiP test archive containing 30,083 studies across 3,255 patients to materialize 914,632 evidence nodes and 884,549 typed relationships.

CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding

• CodeGraph integrates 158 million nodes and 1 billion typed edges across 14 programming languages by applying a code-specialized Large Language Model to the 167 million files of the Stack-Edu corpus. • The pipeline grounds extracted semantic entities in Wikidata using a three-stage linking procedure consisting of a deterministic SPARQL stage, a Deep Research Agent, and a hierarchy-rollup stage.

ARGUS: Role-Aware Event Knowledge Graphs for U.S. Employment-Discrimination Complaints

• ARGUS introduces a source-grounded pipeline that integrates a 5W1H-inspired schema, legal-domain models, and LLM-based structured generation to construct document-level Event Knowledge Graphs from U.S. employment-discrimination complaints on CourtListener.

HypeMed: Enhancing Medication Recommendations with Hypergraph-Based Patient Relationships

• HypeMed introduces a hypergraph-based framework to unify intra-visit coherence modeling and inter-visit augmentation for medication recommendations. • The MedRep module pre-trains clinical visit representations as hyperedges using knowledge-aware contrastive learning, creating a stable embedding space.

Tarab: A Multi-Dialect Corpus of Arabic Lyrics and Poetry

• The Tarab Corpus contains 2.56 million verses and over 13.5 million tokens, representing the largest open Arabic corpus of creative text spanning classical and contemporary works. • The corpus is balanced between songs and poems, covering Classical Arabic, Modern Standard Arabic, and six regional varieties, with structured metadata for comparative analysis.

Prompt Architecture Determines Reasoning Quality: A Variable Isolation Study on the Car Wash Problem

• The STAR (Situation-Task-Action-Result) reasoning framework improves LLM accuracy on the car wash problem from 0% to 85% by enforcing structured goal articulation. • Adding user profile context via vector database retrieval increases accuracy by 10 percentage points, and RAG context adds another 5 percentage points, reaching 100% accuracy.

Prompt Engineering for Scale Development in Generative Psychometrics

• Adaptive prompting outperforms other strategies in generative psychometrics by reducing semantic redundancy and improving structural validity of LLM-generated personality assessment items. • The AI-GENIE framework shows that adaptive prompting significantly increases the quality and size of item pools, with benefits scaling with LLM capability.