A recent survey conducted by Feng Wang from Soochow University has shed light on the phenomenon of AGI hallucination, categorizing its types, causes, and current mitigation approaches. As Artificial General Intelligence (AGI) continues to advance, researchers are increasingly focused on addressing the challenges posed by hallucinations in AI models.
The survey identifies three main types of AGI hallucinations: conflict in intrinsic knowledge of models, factual conflict in information forgetting and updating, and conflict in multimodal fusion. These hallucinations can manifest in various modalities, including language, vision, video, audio, and 3D or agent-based systems.
Factors contributing to the emergence of AGI hallucinations include training data distribution, timeliness of information, and ambiguity in different modalities. The survey emphasizes the importance of high-quality data and effective training techniques in mitigating these hallucinations.
Current mitigation strategies are discussed in three stages: data preparation, model training, and model inference and post-processing. Techniques such as RLHF (Reinforcement Learning from Human Feedback) and knowledge-based approaches are highlighted as effective methods for reducing hallucinations.
The survey also delves into evaluation methodologies for AGI hallucinations, including rule-based, large model-based, and human-based approaches. It notes that while some hallucinations can stimulate a model’s creativity, finding the right balance between hallucination and creative output remains a challenge.
Looking ahead, the authors stress the need for robust datasets in areas like audio, 3D modeling, and agent-based systems. They also highlight the importance of enhancing knowledge updating in models while retaining foundational information.
As AGI technology continues to evolve, understanding and mitigating hallucinations will be crucial for developing reliable and safe AI systems. This comprehensive survey provides valuable insights and sets the stage for future research in this critical area.
