Hallucinations Plague Large Language Models, But New Training Approaches Offer Hope

Large language models (LLMs) have significant limitations despite their recent popularity and hype, including hallucinations, lack of confidence estimates, and absence of citations. Overcoming these challenges is crucial for developing more reliable and trustworthy LLM-based applications.

Hallucinations: The core challenge: LLMs can generate content that appears convincing but is actually inaccurate or entirely false, known as hallucinations:

• Hallucinations are the most difficult issue to address, and their negative impact is only slightly mitigated by confidence estimates and citations.
• Contradictions in the training data contribute to the problem, as LLMs cannot self-inspect their training data for logical inconsistencies.

Bootstrapping consistent LLMs: A potential solution: One approach to mitigate hallucinations is to carefully curate the training data and use the model itself to select additional data:

• Start with a small, highly coherent, logical, and truthful dataset to train a base model.
• Use the base model to classify new text documents as consistent or inconsistent with the curated training corpus.
• Gradually extend the training data with consistent documents and train a larger, more consistent LLM.
• This approach has been explored by researchers at MIT, as described in their paper “Can Logic Help Save Large Language Models from Bias?” (https://arxiv.org/abs/2303.05670)

Confidence estimates and citations: Enhancing transparency: Incorporating confidence estimates and citations can help users assess the reliability of LLM-generated content:

• Confidence estimates assign a score to a prediction, indicating its likely factuality, but high confidence scores for incorrect answers can be problematic.
• OpenAI has released research on teaching models to express uncertainty in words (https://openai.com/index/teaching-models-to-express-their-uncertainty-in-words/).
• Citations provide sources for the generated text, which can be achieved using retrieval-augmented generation (RAG) techniques, as demonstrated by Perplexity.ai and WikiChat.

Expanding the idea: Multiple models with different world views: The bootstrapping approach could be further extended to create models with radically different world views:

• Curate different training corpora representing various sets of beliefs or world views.
• Train separate models on each corpus to create LLMs with distinct perspectives.

Looking ahead: More research is needed to explore the consistent data bootstrapping approach for LLMs and its potential to address the limitations of hallucinations, confidence estimates, and citations. Overcoming these challenges will be crucial for developing more reliable and trustworthy LLM-based applications in the future.