Daily Digest — Sep 24

UniDataAgent: An Ontology-Grounded Agent for Enterprise Question-to-Report Automation

• UniDataAgent achieves 95.0% strict accuracy on real enterprise business questions using ontology grounding, significantly outperforming document RAG at 72.5%. • The system separates semantic acquisition from online execution by utilizing an Ontology Acquisition and Validation stage alongside a Question-to-Report Execution stage across 27 enterprise tables.

Large Knowledge Model: From Papers to a Scientific Reasoning Landscape

• Researchers introduce the Large Knowledge Model (LKM), a scientific infrastructure that transforms academic literature into a computationally accessible reasoning resource featuring source-grounded reasoning graphs.

The Path Matters: Evaluating Small Language Models Beyond Answer Accuracy in KGQA

• Small language models differ substantially in answer accuracy and path fidelity across Kinship and MQuAKE-ST knowledge graph question answering benchmarks, where the two metrics sometimes favor different models.

OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models

• OCCAM is a new framework for open-set causal concept explanation and ontology induction in vision models, designed to interpret decisions of deep image classifiers in black-box settings.

CreativityBench: Evaluating Agent Creative Reasoning via Affordance-Based Tool Repurposing

• CreativityBench is introduced as a benchmark for evaluating the creative reasoning of LLMs through affordance-based tool repurposing, featuring a knowledge base of 4K entities and 150K+ annotations, and 14K grounded tasks.

Rule-based autocorrection of Piping and Instrumentation Diagrams (P&IDs) on graphs

• A new rule-based method utilizes a graph representation of P&IDs, with 33 developed rules, to automate error detection and correction. • The pyDEXPI Python package generates P&ID graphs from DEXPI-standard P&IDs, facilitating the application of these rules. • A case study demonstrates the reliability and effectiveness of this rule-based autocorrection method for revising P&IDs.

C$^2$-Cite: Contextual-Aware Citation Generation for Attributed Large Language Models

• The C^2-Cite framework addresses limitations in attributed LLMs by enhancing contextual awareness of citation markers, improving the integration of retrieved knowledge. • A contextual citation alignment mechanism encodes document contexts into citation symbol representations and aligns marker numbers using a citation router function.