HighlightCurated by Aramai Editorial

Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

Paper introducing Agentic Redux, a provably correct LLM agent architecture.

The authors present Agentic Redux, an LLM agent architecture that ensures semantically guaranteed correctness through typed lambda calculus. They also introduce Ontology-First Agent Design, a methodology for creating agent frameworks on problem domains using human-authored ontologies. The paper includes working code and two production-grade domain examples.

Based on: Provably Auditable and Safe LLM Agents from Human-Authored Ontologies · arXiv

HighlightCurated by Aramai Editorial

Taxonomy-aware deep learning for hierarchical marine species classification in underwater imagery

A taxonomy-aware deep learning framework for classifying marine species from underwater imagery.

The paper presents a taxonomy-aware deep learning framework that aligns with the hierarchical structure of biological classification. The system combines multiple techniques, including taxonomy-weighted loss and minimum-risk Bayesian inference, to improve classification accuracy. Evaluated on the FathomNet 2025 dataset, the system achieves competitive results in marine species classification.

Based on: Taxonomy-aware deep learning for hierarchical marine species classification in underwater imagery · arXiv

HighlightCurated by Aramai Editorial

The Eticas AI Risk Taxonomy: Open Infrastructure for Operationalizing AI Audits

A standardized evaluation framework for operationalizing AI audits.

This paper presents the Eticas AI Risk Taxonomy, a structured framework for evaluating and auditing AI systems. The taxonomy organizes risks into 76 subcategories across 10 categories and provides mappings to 18 external frameworks. It demonstrates a bridge from concept to graded finding by separating risks from their surface mechanisms.

Based on: The Eticas AI Risk Taxonomy: Open Infrastructure for Operationalizing AI Audits · arXiv

HighlightCurated by Aramai Editorial

From Prompt to Process: a Process Taxonomy and Comparative Assessment of Frameworks Supporting AI Software Development Agents

A study on operational frameworks for AI software development agents, proposing a six-dimension process taxonomy.

The paper presents a directed search of primary sources and selects six frameworks supporting AI software development agents. It proposes a six-dimension process taxonomy: specification, context, roles, execution, validation, and portability. The study applies the taxonomy to six frameworks and an out-of-sample case, highlighting convergence in process adoption and trade-offs between depth and portability.

Based on: From Prompt to Process: a Process Taxonomy and Comparative Assessment of Frameworks Supporting AI Software Development Agents · arXiv

HighlightCurated by Aramai Editorial

NeSyCat Torch: A Differentiable Tensor Implementation of Categorical Semantics for Neurosymbolic Learning

A framework for neurosymbolic learning that extends ULLER with a single inductive definition of truth.

The authors present NeSyCat Torch, a differentiable tensor implementation of categorical semantics for neurosymbolic learning. They extend the ULLER framework with a single inductive definition of truth, parametric in a strong monad and an aggregation structure on truth-values. The framework is implemented in probabilistic programming and tensor-based backends, allowing for efficient training in batches.

Based on: NeSyCat Torch: A Differentiable Tensor Implementation of Categorical Semantics for Neurosymbolic Learning · arXiv

HighlightCurated by Aramai Editorial

Travel-Oriented Reasoning Large Language Model via Domain-Specific Knowledge Graphs

A paper proposing a modular pipeline for building a travel-domain reasoning LLM grounded in an expert-designed knowledge graph.

The authors propose a modular pipeline to build a travel-domain reasoning large language model (LLM) using a domain-specific knowledge graph. The pipeline integrates a travel KG, bottom-up construction procedure, and supervised fine-tuning stage to embed domain knowledge into the LLM. The approach achieves high accuracy on a benchmark dataset.

Based on: Travel-Oriented Reasoning Large Language Model via Domain-Specific Knowledge Graphs · arXiv

HighlightCurated by Aramai Editorial

XGRAG: A Graph-Native Framework for Explaining KG-based Retrieval-Augmented Generation

A framework that generates causally grounded explanations for GraphRAG systems.

The paper introduces XGRAG, a graph-native framework for explaining knowledge graph-based retrieval-augmented generation. It employs graph-based perturbation strategies to quantify the contribution of individual graph components on the model answer. The authors conduct experiments comparing XGRAG against an existing explainability baseline and evaluate its robustness across various question types and LLMs.

Based on: XGRAG: A Graph-Native Framework for Explaining KG-based Retrieval-Augmented Generation · arXiv

HighlightCurated by Aramai Editorial

MemSyco-Bench: Benchmarking Sycophancy in Agent Memory

A benchmark for evaluating memory-induced sycophancy in agent systems.

MemSyco-Bench is a comprehensive benchmark that measures how retrieved memories influence downstream reasoning and decision-making in agent systems.,It covers five tasks to assess an agent's ability to reject, respect, resolve conflicts between memory and objective evidence, track updates, and use valid memory for personalization.,The benchmark aims to bridge the gap in existing memory benchmarks by evaluating the impact of retrieved memories on factual accuracy and objective reasoning.

Based on: MemSyco-Bench: Benchmarking Sycophancy in Agent Memory · arXiv

HighlightCurated by Aramai Editorial

MAECO-Lite: Modular Ontology for Dynamic Malware Analysis

A lightweight ontology designed to represent data and operationalize its processing for dynamic malware analysis.

The paper proposes MAECO-Lite, a modular ontology that separates enduring entities from runtime events. It is based on an ontological analysis of core MAEC constructs relevant to dynamic malware analysis. The ontology improves learning performance using description logic concept learning algorithms.

Based on: MAECO-Lite: Modular Ontology for Dynamic Malware Analysis · arXiv

HighlightCurated by Aramai Editorial

CORTEX: High-Quality Cross-Domain Organization of Web-Scale Corpora through Ontological Corpus Graph

A framework for elevating web-scale corpus construction to structured knowledge organization.

The paper presents Cortex, a three-layer heterogeneous structure (Ontological Corpus Graph) for organizing high-quality corpora. It refines content, evolves ontologies, and enables cross-domain alignment. Comprehensive experiments validate its effectiveness in quality refinement, domain organization, and data synthesis.

Based on: CORTEX: High-Quality Cross-Domain Organization of Web-Scale Corpora through Ontological Corpus Graph · arXiv

HighlightCurated by Aramai Editorial

Semantic constraint validation in knowledge representation for the semantic web: A survey, taxonomy and research challenges

A survey, taxonomy and research challenges on semantic constraint validation.

This paper presents a survey and taxonomy of semantic constraint validation techniques for knowledge representation on the Semantic Web. It identifies research challenges and gaps in current approaches. The authors aim to provide a comprehensive overview of existing methods and their limitations.

Based on: Semantic constraint validation in knowledge representation for the semantic web: A survey, taxonomy and research challenges · Data & Knowledge Engineering

HighlightCurated by Aramai Editorial

BIRDNet: Mining and Encoding Boolean Implication Knowledge Graphs as Interpretable Deep Neural Networks

A neural network architecture that encodes Boolean implication relationships as a layered graph.

The paper proposes BIRDNet, a neural network architecture that mines Boolean implication relationships from tabular data. The mined implications are encoded as a layered graph, where each hidden unit corresponds to one rule and binds only to its two features. This design results in a sparse and interpretable model that can recover known biological signatures.

Based on: BIRDNet: Mining and Encoding Boolean Implication Knowledge Graphs as Interpretable Deep Neural Networks · arXiv