[EMNLP 2024] The official GitHub repo for the survey paper "Knowledge Conflicts for LLMs: A Survey"

[ICLR'24 Spotlight] "Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts"

Code and data for "ConflictBank: A Benchmark for Evaluating the Influence of Knowledge Conflicts in LLM" (NeurIPS 2024 Track Datasets and Benchmarks)

Open-source memory layer for multi-agent LLM systems. Typed conflict resolution ensures agents share accurate, up-to-date knowledge stale facts are detected and resolved at ingestion, not left for the LLM to guess.

An improved contrastive decoding approach that can better resolve LLM knowledge conflicts between prior belief and contextual information

This project demonstrates a novel approach to automate and enhance key aspects of Requirements Engineering (RE) by integrating Retrieval-Augmented Generation (RAG)-Enhanced Large Language Models (LLMs) with dynamic Knowledge Graph (KG) construction, metaheuristic conflict resolution, and multi-criteria traceability optimization.

Dataset for the paper "Exploring Knowledge Conflicts for Faithful LLM Reasoning: Benchmark and Method"

Website of paper "tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge"

The code of article: "Debate-of-Thoughts: Resolving Knowledge Conflicts in LLMs Through Internal Deliberation"

Structured confabulation in LLMs under knowledge conflict. 168 experimental runs, 3 models, escape hatch that stops it every time.

This repository contains my explanations and presentations of recent research papers on: - Knowledge Injection in Large Language Models (LLMs) - Retrieval-Augmented Generation (RAG) - Handling knowledge conflicts - Knowledge Graphs (Neo4j, UMLS, ConceptNet)

LangGraph AI agent with 3 variants: baseline RAG, Multi-Agent Consensus Protocol (MCP) for conflicting sources, and DeepSeek LLM. Corporate knowledge base demo.

Swin-VIB is a knowledge-conflict control framework for retrieval-augmented LLMs that adaptively regulates external context injection with an information-bottleneck sliding-window mechanism to improve reliability and consistency in the wild.

DrugSense detects harmful drug interactions using a Neo4j knowledge graph and a local medical LLM. Enter prescribed medications, and it instantly checks for conflicts and generates clear, patient-safe warnings. It reduces medication errors and improves clinical decision-making.

MemoSynth-Lite is a lightweight AI-powered organizational memory system that stores, queries, and summarizes knowledge using semantic search (Qdrant), graph reasoning (Neo4j), and timeline tracking (DuckDB). It supports multilingual memory insights, conflict resolution, and works fully offline using local LLMs.