HighlightCurated by Aramai Editorial

mcp-proto-okn: Natural-language access to open scientific knowledge graphs through the Model Context Protocol

A Python-based Model Context Protocol server for natural-language access to open scientific knowledge graphs.

The MCP Server Proto-OKN enables AI assistants to discover, inspect, and query scientific knowledge graphs through natural language. It provides various functions such as graph routing, schema inspection, and SPARQL execution. The server is implemented in Python using the FastMCP framework and is available on GitHub.

Based on: mcp-proto-okn: Natural-language access to open scientific knowledge graphs through the Model Context Protocol · arXiv

HighlightCurated by Aramai Editorial

Decoupling Knowledge and Task Subspaces for Composable Parametric Retrieval Augmented Generation

Paper proposing a method to improve compositional robustness in parametric retrieval-augmented generation.

The authors introduce Orthogonal Subspace Decomposition (OSD) to separate reusable task behavior from document-specific knowledge adapters. This setup is used to examine the effect of orthogonalizing task and document updates on adapter composition in multi-document PRAG. Experiments show improved compositional robustness, especially when merging multiple document adapters.

Based on: Decoupling Knowledge and Task Subspaces for Composable Parametric Retrieval Augmented Generation · arXiv

HighlightCurated by Aramai Editorial

Quantifying the Affective Gap: A Zero-Shot Evaluation of LLMs on Fine-Grained Emotion Taxonomies

A zero-shot evaluation of three leading commercial large language models on fine-grained emotion taxonomies.

This paper presents a unified zero-shot evaluation of three LLMs (Claude, ChatGPT, and Gemini) on a 13-class emotion classification task. The results show that all models excel on certain emotions but consistently fail on others. The study highlights the limitations of current AI systems in fine-grained emotion classification.

Based on: Quantifying the Affective Gap: A Zero-Shot Evaluation of LLMs on Fine-Grained Emotion Taxonomies · arXiv

HighlightCurated by Aramai Editorial

Evidence-Supported Credit Risk Report Generation Using News-Centric Financial Knowledge Graphs

A framework for automatically constructing knowledge graphs from news events and company data to generate credit risk reports.

The authors present FinKG-News, a framework that extracts news events as anchors linked to companies in financial knowledge graphs. This approach is used to develop an in-context learning architecture for credit risk report generation across three core financial dimensions. The results show improved quality and reduced hallucinations compared to baselines.

Based on: Evidence-Supported Credit Risk Report Generation Using News-Centric Financial Knowledge Graphs · arXiv

HighlightCurated by Aramai Editorial

Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service

A study on automated detection and classification of potentially abusive clauses in Chilean Terms of Service using a retrieval-augmented generation framewo

The paper presents a framework for detecting and classifying potentially abusive clauses in Chilean Terms of Service. It combines efficient clause detection, hybrid dense-sparse retrieval, reranking, and prompt augmentation to support medium-sized open-weight language models. The study also contributes a refined legal annotation scheme and a practical design for AI-assisted consumer contract review.

Based on: Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service · arXiv

HighlightCurated by Aramai Editorial

A Training-Free Mixture-of-Agents Framework for Multi-Document Summarization using LLMs and Knowledge Graphs

Proposes a training-free framework for multi-document summarization leveraging large language models and knowledge graphs.

The paper presents a training-free mixture-of-agents framework for multi-document summarization, decomposing the task into specialized agent tasks. The approach leverages large language models and knowledge graphs to capture complex inter-document relationships without requiring labeled data or fine-tuning. Experiments demonstrate state-of-the-art or competitive performance across four datasets in English and Vietnamese.

Based on: A Training-Free Mixture-of-Agents Framework for Multi-Document Summarization using LLMs and Knowledge Graphs · arXiv

HighlightCurated by Aramai Editorial

222P A hybrid text-knowledge graph retrieval-augmented generation system for clinical decision support in soft tissue sarcoma

A study on a hybrid text-knowledge graph retrieval-augmented generation system for clinical decision support.

The study presents a system that combines text and knowledge graphs to aid in clinical decision-making. It focuses on soft tissue sarcoma, analyzing proteomic data from patient samples. The analysis highlights potential pathways for further investigation and identifies potential drug candidates.

Based on: 222P A hybrid text-knowledge graph retrieval-augmented generation system for clinical decision support in soft tissue sarcoma · ESMO rare cancers.

HighlightCurated by Aramai Editorial

Automating Geometry-Intensive Compliance Checking in BIM: Graph-Based Semantic Reasoning Framework

A graph-based semantic reasoning framework for automating compliance checks in Building Information Modeling.

The paper proposes a Spatial-Geometric Reasoning System (SGR-BIM) to automate geometry-intensive compliance checking in BIM. SGR-BIM constructs a cross-modal knowledge graph aligning user intent, regulatory semantics, and BIM geometry. The framework achieves high accuracy on expert-verified queries from fire safety codes.

Based on: Automating Geometry-Intensive Compliance Checking in BIM: Graph-Based Semantic Reasoning Framework · arXiv

HighlightCurated by Aramai Editorial

A Neuro-Symbolic Approach to Strategy Synthesis for Strategic Logics

Paper introducing a neuro-symbolic framework for strategy synthesis in Multi-Agent Systems.

The authors propose a generate-and-certify architecture that integrates large language models into the model-checking pipeline. This approach uses LLM guidance to navigate large combinatorial strategy spaces while preserving formal soundness. The paper instantiates this framework for bounded strategic reasoning in NatATL and introduces a new dataset.

Based on: A Neuro-Symbolic Approach to Strategy Synthesis for Strategic Logics · arXiv

HighlightCurated by Aramai Editorial

Ontology Memory-Augmented ASR Correction for Long Text-Speech Interleaved Conversations

Proposes an ontology memory-augmented framework for long text-speech interleaved conversations.

The paper proposes a framework that organizes preceding interaction history into a dynamically updatable ontology memory to improve automatic speech recognition correction. The framework stores entities, terminology, and semantic relations as retrievable nodes for context-grounded correction. Experiments show improved results over direct correction in most paired backbone-setting combinations.

Based on: Ontology Memory-Augmented ASR Correction for Long Text-Speech Interleaved Conversations · arXiv

HighlightCurated by Aramai Editorial

SeedER: Seed-and-Expand Retrieval from Knowledge Graphs

A retrieval framework for knowledge graphs that leverages structure through iterative expansion.

The authors introduce SeedER, a retrieval framework for knowledge graphs that explicitly uses graph structure to improve efficiency and effectiveness.,SeedER first seeds a compact set of core nodes using lightweight dense and entity-based retrieval, then selectively expands this set via a learned policy.,This design enables efficient discovery of query-relevant nodes while controlling expansion cost.

Based on: SeedER: Seed-and-Expand Retrieval from Knowledge Graphs · arXiv

HighlightCurated by Aramai Editorial

Talk is (Not) Cheap: A Taxonomy and Benchmark Coverage Audit for LLM Attacks

A framework for auditing the coverage of LLM attack benchmarks.

The authors introduce a reusable framework for auditing the collective coverage of LLM attack benchmarks. They construct a 4x6 matrix grounded in STRIDE, covering 932 arXiv security studies from 2023-2026. The study reveals that existing benchmarks cover at most 25% of the threat surface and highlights evaluation gaps.

Based on: Talk is (Not) Cheap: A Taxonomy and Benchmark Coverage Audit for LLM Attacks · arXiv