EdinburghNLP/awesome-hallucination-detection

What it solves

This project addresses the tendency of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) to "hallucinate"—generating information that is factually incorrect, contradictory to provided context, or unsupported by visual data. It serves as a centralized repository for research papers, datasets, and methods designed to detect, measure, and mitigate these errors.

How it works

The repository organizes state-of-the-art research into various approaches for managing hallucinations. These include:

• Detection: Methods that analyze internal model signals (like attention maps, entropy, or hidden states) to flag potential hallucinations without needing external labels.
• Mitigation: Techniques such as "steering" model activations, suppressing specific layers that cause unfaithful responses, or using reinforcement learning to prioritize factual accuracy.
• Evaluation: Benchmarks and metrics designed to test how well models handle factual claims, scientific critiques, and multimodal data, while highlighting the limitations of older evaluation methods like ROUGE.

Who it’s for

This resource is for AI researchers, developers, and engineers working on building more reliable, fact-aware LLM applications. It is particularly useful for those looking to implement RAG (Retrieval-Augmented Generation) systems or improve the factual consistency of models in production environments.

Highlights

• Curated Research: A collection of 145+ academic papers covering hallucination detection and mitigation.
• Diverse Methodologies: Includes training-free inference techniques, RL-based fine-tuning, and topological analysis of attention graphs.
• Multimodal Support: Features specific research on reducing hallucinations in vision-language models by steering visual information.
• Benchmark Access: Provides links to datasets and leaderboards for testing model factuality and critique skills.
• Practical Tools: Highlights open-source code and libraries that can be integrated into existing workflows.