This repository is a curated list of papers focused on detecting and mitigating hallucinations in Large Language Models (LLMs). It serves researchers and practitioners aiming to improve the factual accuracy and trustworthiness of LLM outputs across various domains, including question answering, summarization, and vision-language tasks.

How It Works

The collection highlights diverse approaches to hallucination detection and mitigation. Methods range from analyzing semantic similarities and embedding spaces to leveraging internal model states, external knowledge bases, and even fine-grained AI feedback. Some papers focus on preemptive detection before generation, while others propose post-generation correction or uncertainty quantification techniques.

Quick Start & Requirements

• This is a curated list of research papers, not a runnable software package.
• Links to papers, datasets, and code repositories are provided within the README.

Highlighted Details

• Covers a broad spectrum of hallucination types, including factuality and faithfulness.
• Includes benchmarks and datasets specifically designed for evaluating hallucination detection and mitigation.
• Features methods applicable to both text-only and multimodal (vision-language) LLMs.
• Discusses various evaluation metrics, from traditional statistical measures to model-based and human-centric assessments.

Maintenance & Community

• The repository is maintained by EdinburghNLP.
• It cites a comprehensive list of academic papers, indicating active curation.
• Links to related surveys and shared tasks are provided.

Licensing & Compatibility

• The repository itself is licensed under the MIT License.
• Individual papers and linked code repositories will have their own licenses.

Limitations & Caveats

• This resource is a collection of research papers and does not provide a unified tool or framework for hallucination detection.
• The effectiveness and applicability of the discussed methods vary depending on the specific LLM and task.