HighlightCurated by Aramai Editorial

A homotopy-type-theoretic generalization of neurosymbolic inference

Paper proposing a framework for neurosymbolic systems using homotopy type theory.

The authors develop a framework for neurosymbolic systems based on homotopy type theory, which preserves information about symmetries and distinct proofs. They prove a conservativity theorem and show that the framework exposes symmetry behind reasoning shortcuts. The proposed method is demonstrated to be better calibrated than ensemble methods on MNIST benchmarks.

Based on: A homotopy-type-theoretic generalization of neurosymbolic inference · arXiv

HighlightCurated by Aramai Editorial

DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation

A framework for learnable evidence control in multi-hop retrieval-augmented generation.

The paper introduces DynaKRAG, a unified framework that formulates multi-hop evidence acquisition as state-conditioned control over atomic evidence operations. It uses a learned controller to select the next operation and updates the evidence state accordingly. The authors evaluate DynaKRAG on several benchmarks and demonstrate its effectiveness in coordinating retrieval, diagnosis, and gap-directed acquisition.

Based on: DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation · arXiv

HighlightCurated by Aramai Editorial

A Multi-Agent and synergistic Knowledge Graph retrieval-augmented generation framework for intelligent maintenance

This paper proposes a framework for intelligent maintenance using multi-agent and synergistic knowledge graph retrieval-augmented generation.

The authors present a framework that combines multi-agent systems with knowledge graph retrieval-augmented generation to improve intelligent maintenance. The framework is designed to enhance the efficiency and effectiveness of maintenance tasks by leveraging the strengths of both approaches. This work contributes to the development of more advanced maintenance systems that can adapt to complex environments.

Based on: A Multi-Agent and synergistic Knowledge Graph retrieval-augmented generation framework for intelligent maintenance · Journal of Manufacturing Systems

HighlightCurated by Aramai Editorial

Learning Red Agent Policy from Observations for Neurosymbolic Autonomous Cyber Agents

A paper proposing a policy learning technique for partially observable reinforcement learning agents in autonomous cyber-defense.

The authors propose an imitation learning-based policy learning technique to predict red agent actions in autonomous cyber-environments. The method is integrated with a neurosymbolic cyber-defense agent and achieves high prediction accuracy across diverse scenarios. This approach addresses the challenge of partially observable systems, where defender actions are not directly observable.

Based on: Learning Red Agent Policy from Observations for Neurosymbolic Autonomous Cyber Agents · arXiv

HighlightCurated by Aramai Editorial

Proof of Unlearning for Semantic Knowledge Bases in Large Language Models-Enabled Semantic Communication

A framework for efficiently and verifiably updating large language model-enabled semantic knowledge bases.

The authors propose a proof-of-unlearning framework for updating large language models (LLMs) used in semantic knowledge bases. The framework tracks the evolution of unlearning by measuring drifts in the LoRA adapter subspace. Experimental results demonstrate its effectiveness. This work addresses the challenge of removing outdated, malicious, or privacy-sensitive content from LLMs without retraining.

Based on: Proof of Unlearning for Semantic Knowledge Bases in Large Language Models-Enabled Semantic Communication · IEEE Communications Magazine

HighlightCurated by Aramai Editorial

Agent-Native Immune System: Architecture, Taxonomy, and Engineering

A biologically inspired defense architecture for autonomous agents.

The paper introduces the Agent-Native Immune System (ANIS), a framework that addresses vulnerabilities in AI agents. It presents a six-layer Immune Tower, a taxonomy of Agent Viruses and Vaccines, and Continual Immune Learning. The ANIS serves as a dynamic defense mechanism during runtime.

Based on: Agent-Native Immune System: Architecture, Taxonomy, and Engineering · arXiv

HighlightCurated by Aramai Editorial

TruthfulRAG: Resolving Factual-level Conflicts in Retrieval-Augmented Generation with Knowledge Graphs

A framework for resolving factual conflicts between LLMs' internal knowledge and external information using knowledge graphs.

This paper proposes TruthfulRAG, a framework that leverages knowledge graphs to resolve factual conflicts in RAG systems. It constructs KGs from retrieved content, identifies relevant knowledge through query-based graph retrieval, and employs entropy-based filtering mechanisms to mitigate inconsistencies. The authors claim that TruthfulRAG outperforms existing methods in resolving knowledge conflicts and improving the robustness of RAG systems.

Based on: TruthfulRAG: Resolving Factual-level Conflicts in Retrieval-Augmented Generation with Knowledge Graphs · Proceedings of the AAAI Conference on Artificial Intelligence

HighlightCurated by Aramai Editorial

MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection

A post-hoc causal memory auditing framework for memory-augmented LLM agents.

The paper proposes MemAudit, a framework that combines counterfactual influence scores and structural anomaly detection to identify malicious memories in LLM agents. It evaluates the effectiveness of MemAudit against memory injection attacks and demonstrates significant reductions in attack success rates. The framework aims to address post-hoc auditing of poisoned agent memory.

Based on: MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection · arXiv

HighlightCurated by Aramai Editorial

Text Knows What, Tables Know When: Clinical Timeline Reconstruction via Retrieval-Augmented Multimodal Alignment

A framework for reconstructing clinical timelines from text and structured EHR data.

The paper introduces a retrieval-augmented multimodal alignment framework to improve the temporal precision of absolute clinical timelines extracted from text. It formulates timeline reconstruction as a graph-based multistep process, using both unstructured narratives and structured electronic health record (EHR) data. The approach is evaluated on the i2m4 benchmark, showing improved accuracy and concordance compared to unimodal text-only reconstruction.

Based on: Text Knows What, Tables Know When: Clinical Timeline Reconstruction via Retrieval-Augmented Multimodal Alignment · arXiv

HighlightCurated by Aramai Editorial

ChronoMedKG: A Temporally-Grounded Biomedical Knowledge Graph and Benchmark for Clinical Reasoning

A temporal biomedical knowledge graph that contains evidence-linked triples covering 13,431 diseases.

ChronoMedKG is a temporally-grounded biomedical knowledge graph constructed through a multi-agent pipeline. It contains 460,497 evidence-linked triples and adds temporal grounding for 6,250 diseases absent from other resources. The authors also introduce ChronoTQA, a benchmark of questions across eight task types.

Based on: ChronoMedKG: A Temporally-Grounded Biomedical Knowledge Graph and Benchmark for Clinical Reasoning · arXiv

HighlightCurated by Aramai Editorial

Forensic Trajectory Signatures for Agent Memory Poisoning Detection

A research paper on detecting memory poisoning attacks in large language models.

The authors discover a behavioral invariant in LLM agents under persistent memory poisoning, which can be used to detect attacks. A simple rule and a Random Forest classifier are proposed to exploit this invariant, achieving high accuracy. The signature is overdetermined and generalizes to frontier models without retraining.

Based on: Forensic Trajectory Signatures for Agent Memory Poisoning Detection · arXiv

HighlightCurated by Aramai Editorial

Automating Cause-Effect Specification with Knowledge Graphs and Large Language Models

A paper proposing a framework for automating cause-and-effect logic generation using knowledge graphs and large language models.

The authors present a semantic-AI framework that combines a knowledge graph with a constrained large language model to automate the generation of cause-and-effect logic. The framework builds on an established modular alignment ontology and demonstrates its application on a modular process plant. This approach aims to reduce manual effort in creating engineering specifications.

Based on: Automating Cause-Effect Specification with Knowledge Graphs and Large Language Models · arXiv