[Paper Review] A Survey of Hallucination in Large Foundation Models

Vipula Rawte, Amit Sheth+1 more|arXiv (Cornell University)|Sep 12, 2023

Complex Systems and Time Series Analysis94 citations

TL;DR

A comprehensive survey of hallucination in large foundation models (LFMs) across text, image, video, and audio, detailing types, evaluation, detection, mitigation, datasets, and future directions.

ABSTRACT

Hallucination in a foundation model (FM) refers to the generation of content that strays from factual reality or includes fabricated information. This survey paper provides an extensive overview of recent efforts that aim to identify, elucidate, and tackle the problem of hallucination, with a particular focus on ``Large'' Foundation Models (LFMs). The paper classifies various types of hallucination phenomena that are specific to LFMs and establishes evaluation criteria for assessing the extent of hallucination. It also examines existing strategies for mitigating hallucination in LFMs and discusses potential directions for future research in this area. Essentially, the paper offers a comprehensive examination of the challenges and solutions related to hallucination in LFMs.

Motivation & Objective

• Categorize existing work on hallucination in LFMs.
• Examine LFMs across text, image, video, and audio modalities.
• Summarize detection methods, mitigation strategies, tasks, datasets, and evaluation metrics.
• Propose future research directions and open-source resources.

Proposed method

• Classify LFMs into four modalities: text, image, video, and audio.
• Review detection, mitigation, tasks, datasets, and evaluation metrics for hallucination.
• Summarize modality-specific works and datasets (e.g., HaluEval, Med-HALT, M-HalDetect).
• Highlight prompting, external knowledge, grounding, and data-augmentation approaches for mitigation.
• Offer future directions including automated evaluation and curated knowledge sources.

Figure 2: The evolution of “hallucination” papers for Large Foundation Models (LFMs) from March 2023 to September 2023.

Experimental results

Research questions

• RQ1 What are the main types of hallucination observed in LFMs across different modalities?
• RQ2 What detection and mitigation strategies are proposed for LFMs to reduce hallucinations?
• RQ3 What datasets and evaluation metrics are used to assess hallucination across text, image, video, and audio?
• RQ4 What future directions can advance reliable hallucination detection and mitigation in LFMs?
• RQ5 How do domain-specific and multilingual LFMs differ in their hallucination profiles and remediation?

Key findings

• The survey provides a taxonomy of hallucination types and a cross-modal overview (text, image, video, audio) of LFMs.
• It catalogs detection methods, mitigation techniques, datasets, and evaluation metrics across modalities.
• It highlights notable benchmarks and datasets such as HaluEval, Med-HALT, HALO, M-HalDetect, and Lp-MusicCaps.
• It discusses prompting-based, knowledge-grounding, and external-knowledge integration as mitigation directions.
• It emphasizes future research directions in automated evaluation, knowledge-curation, and ethical considerations.

Figure 3: An illustration of hallucination Luo et al. ( 2023 ) . Incorrect information is highlighted in Red .