Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LightHouse: A Survey of AGI Hallucination

Feng Wang

TL;DR

LightHouse provides a structured survey of AGI hallucination, offering a taxonomy of three primary hallucination types and detailing how data distribution, information timeliness, and cross-modal ambiguity drive emergence. It covers mitigation across data, training, and inference, including RLHF, retrieval-augmented methods, and post-hoc approaches, and reviews evaluation benchmarks across language and multimodal modalities. The paper highlights the complex role of hallucinations as both a safety risk and a potential driver of creativity, and calls for data- and modality-specific research to advance robust AGI systems. This work informs researchers and practitioners about current gaps and practical directions for reducing hallucinations in next-generation AGI systems.

Abstract

With the development of artificial intelligence, large-scale models have become increasingly intelligent. However, numerous studies indicate that hallucinations within these large models are a bottleneck hindering the development of AI research. In the pursuit of achieving strong artificial intelligence, a significant volume of research effort is being invested in the AGI (Artificial General Intelligence) hallucination research. Previous explorations have been conducted in researching hallucinations within LLMs (Large Language Models). As for multimodal AGI, research on hallucinations is still in an early stage. To further the progress of research in the domain of hallucinatory phenomena, we present a bird's eye view of hallucinations in AGI, summarizing the current work on AGI hallucinations and proposing some directions for future research.

LightHouse: A Survey of AGI Hallucination

TL;DR

LightHouse provides a structured survey of AGI hallucination, offering a taxonomy of three primary hallucination types and detailing how data distribution, information timeliness, and cross-modal ambiguity drive emergence. It covers mitigation across data, training, and inference, including RLHF, retrieval-augmented methods, and post-hoc approaches, and reviews evaluation benchmarks across language and multimodal modalities. The paper highlights the complex role of hallucinations as both a safety risk and a potential driver of creativity, and calls for data- and modality-specific research to advance robust AGI systems. This work informs researchers and practitioners about current gaps and practical directions for reducing hallucinations in next-generation AGI systems.

Abstract

With the development of artificial intelligence, large-scale models have become increasingly intelligent. However, numerous studies indicate that hallucinations within these large models are a bottleneck hindering the development of AI research. In the pursuit of achieving strong artificial intelligence, a significant volume of research effort is being invested in the AGI (Artificial General Intelligence) hallucination research. Previous explorations have been conducted in researching hallucinations within LLMs (Large Language Models). As for multimodal AGI, research on hallucinations is still in an early stage. To further the progress of research in the domain of hallucinatory phenomena, we present a bird's eye view of hallucinations in AGI, summarizing the current work on AGI hallucinations and proposing some directions for future research.
Paper Structure (14 sections, 5 figures, 3 tables)

This paper contains 14 sections, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Definition for AGI Hallucination
  3. Conflict in Intrinsic Knowledge of Models
  4. Factual Conflict in Information Forgetting and Updating
  5. Conflict in Multimodal Fusion
  6. Emergence for AGI Hallucination
  7. Training Data Distribution
  8. Timeliness of Information
  9. Ambiguity in Different Modalities
  10. Mitigation for AGI Hallucination
  11. Evaluation for AGI Hallucination
  12. Discourse for AGI Hallucination
  13. Talking about Future
  14. Example Appendix

Figures (5)

  • Figure 1: Illustration of Hallucination in AGI. Which classify by three types: 1. Conflict in Intrinsic Knowledge of Models. 2. Factual Conflict in Information Forgetting and Updating. 3. Conflict in Multimodal Fusion. is input information, is forgotten information, is new information, and is error output.
  • Figure 2: The overview structure of this survey. we have explored the definition of hallucinations in the AGI, examined their causes, evaluated current works to mitigate them, assessed methods for hallucination analysis, engaged in dialogues concerning these phenomena, and contemplated future outlooks in the AGI.
  • Figure 3: Talking about Future.
  • Figure 4: Conflict in Intrinsic Knowledge examples.
  • Figure 5: Factual Conflict examples.