Hallucinations in large language models (LLM) and their causes

The emergence of large language models (LLM) such as Llama, PaLM and GPT-4 has revolutionized natural language processing (NLP), significantly advancing text understanding and generation. However, despite their notable capabilities, LLMs are prone to producing hallucinations, content that is objectively incorrect or inconsistent with user input. This phenomenon substantially challenges its reliability in real-world applications, requiring a comprehensive understanding of its principles, causes, and mitigation strategies.

Definition and types of hallucinations

Hallucinations in LLMs are generally classified into two main types: factuality hallucinations and fidelity hallucinations.

1. Objective hallucination: This type involves discrepancies between the generated content and verifiable real-world facts. It is further divided into:

• Factual inconsistency: It occurs when the result contains factual information that contradicts known facts. For example, an LLM might incorrectly claim that Charles Lindbergh was the first to walk on the moon instead of Neil Armstrong.
• Factual manufacturing: It involves creating completely unverifiable facts, such as making up historical details about unicorns.

1. Fidelity hallucination: This type refers to the divergence of the generated content from the user's instructions or the provided context. Includes:

• Instruction inconsistency: When the result does not follow the user's directive, such as answering a question instead of translating it according to the instructions.
• Context inconsistency: It occurs when the generated content contradicts the contextual information provided, such as misrepresenting the origin of the Nile River.
• Logical inconsistency: It involves internal contradictions within the generated content, often observed in reasoning tasks.

Causes of hallucinations in LLM

The root causes of hallucinations in LLMs span the entire spectrum of development, from data acquisition to training and inference. These causes can be broadly classified into three parts:

1. Data-related causes:

• Faulty data sources: Misinformation and biases in pre-training data can lead to hallucinations. For example, heuristic data collection methods can inadvertently introduce incorrect information, leading to imitative falsehoods.
• Limits of knowledge: LLMs may lack up-to-date specialized factual or domain knowledge, resulting in factual fabrications. For example, they might provide outdated information about recent events or need more experience in specific medical fields.
• Lower data utilization: LLMs can produce hallucinations due to spurious correlations and failures in knowledge retrieval, even with extensive knowledge. For example, they might incorrectly claim that Toronto is the capital of Canada due to the frequent simultaneous appearance of “Toronto” and “Canada” in the training data.

2. Causes related to training:

• Architectural defects: The unidirectional nature of transformer-based architectures can hinder the ability to capture intricate contextual dependencies, increasing the risk of hallucinations.
• Exposure bias: Discrepancies between training (where models are based on ground-truth tokens) and inference (where models are based on their results) can lead to cascading errors.
• Alignment problems: Misalignment between the capabilities of the model and the demands of the alignment data can lead to hallucinations. Additionally, belief misalignment, where models produce results that diverge from their internal beliefs to align with human feedback, can also cause hallucinations.

3. Causes related to inference:

• Decoding strategies: The randomness inherent in stochastic sampling strategies may increase the likelihood of hallucinations. Higher sampling temperatures result in more uniform token probability distributions, leading to the selection of less probable tokens.
• Imperfect decoding representations: Insufficient attention to context and the softmax bottleneck can limit the model's ability to predict the next token, causing hallucinations.

Mitigation strategies

Several strategies have been developed to address hallucinations, improve data quality, improve training processes, and refine decoding methods. Key approaches include:

1. Improving data quality: Ensure the accuracy and integrity of training data to minimize the introduction of misinformation and bias.
2. Training improvements: Develop better architectural designs and training strategies, such as bidirectional context models and techniques to mitigate exposure bias.
3. Advanced decoding techniques: Employ more sophisticated decoding methods that balance randomness and precision to reduce the occurrence of hallucinations.

Conclusion

Hallucinations in LLMs present significant challenges to their practical implementation and reliability. Understanding the various types and underlying causes of hallucinations is crucial to developing effective mitigation strategies. By improving data quality, improving training methodologies, and refining decoding techniques, the NLP community can work to create more accurate and reliable LLMs for real-world applications.

Sources

• https://arxiv.org/pdf/2311.05232