The field of Large Language Models (LLMs) is moving towards developing more reliable and trustworthy models by addressing the issue of hallucinations, which refers to the generation of confident but factually incorrect information. Recent research has focused on developing innovative methods for hallucination detection and mitigation, including metamorphic testing frameworks, attention probing techniques, and self-improving faithfulness-aware contrastive tuning. These approaches aim to improve the accuracy and reliability of LLMs, particularly in high-stakes domains such as law and enterprise applications. Noteworthy papers in this area include MetaRAG, which proposes a metamorphic testing framework for hallucination detection in Retrieval-Augmented Generation (RAG) systems, and SI-FACT, which presents a self-improving framework for mitigating knowledge conflict in LLMs. Overall, the field is advancing towards more robust and reliable LLMs that can be deployed in real-world applications with confidence.
Advances in Hallucination Detection and Mitigation for Large Language Models
Sources
MetaRAG: Metamorphic Testing for Hallucination Detection in RAG Systems
Inteligencia Artificial jur\'idica y el desaf\'io de la veracidad: an\'alisis de alucinaciones, optimizaci\'on de RAG y principios para una integraci\'on responsable
Cross-Layer Attention Probing for Fine-Grained Hallucination Detection
Investigating Symbolic Triggers of Hallucination in Gemma Models Across HaluEval and TruthfulQA
Unsupervised Hallucination Detection by Inspecting Reasoning Processes
SI-FACT: Mitigating Knowledge Conflict via Self-Improving Faithfulness-Aware Contrastive Tuning
SENTRA: Selected-Next-Token Transformer for LLM Text Detection
LLM Hallucination Detection: A Fast Fourier Transform Method Based on Hidden Layer Temporal Signals
Hallucination Detection with the Internal Layers of LLMs
JU-NLP at Touch\'e: Covert Advertisement in Conversational AI-Generation and Detection Strategies
DetectAnyLLM: Towards Generalizable and Robust Detection of Machine-Generated Text Across Domains and Models
Estimating Semantic Alphabet Size for LLM Uncertainty Quantification
AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt