How to Mitigate Extrinsic Hallucinations in Large Language Models

By • min read

Introduction

Large language models (LLMs) can produce impressively coherent text, but they sometimes fabricate information, a phenomenon known as hallucination. When the generated content is not grounded in the provided context or real-world knowledge, it's called an extrinsic hallucination. This guide focuses on reducing such hallucinations by ensuring your LLM stays factual and honestly admits when it doesn't know an answer. By following these steps, you can make your model more reliable and trustworthy.

What You Need

• Access to a large language model (e.g., GPT-4, LLaMA, or similar) via API or local deployment
• Basic understanding of the model's pre-training dataset (size, domain, known biases)
• Ability to modify model prompts, fine-tune, or adjust inference parameters
• A set of factual evaluation examples to test output accuracy
• Optional: retrieval-augmented generation (RAG) system for grounding external knowledge

Step-by-Step Guide

Step 1: Distinguish Extrinsic from In-Context Hallucination

Before you can fix extrinsic hallucinations, you must identify them. In-context hallucination happens when the model contradicts the explicit context provided in the prompt (e.g., a source document). Extrinsic hallucination occurs when the output fabricates facts not supported by the model's training data or real-world knowledge. For example, if you ask about a historical event and the model creates a fictional date, that's extrinsic. To differentiate, check whether the model's claim can be verified externally. Use reliable sources or a knowledge base to cross-reference. This step lays the foundation for targeted mitigation.

Step 2: Ensure Factual Grounding in Pre-Training Data

Avoiding extrinsic hallucinations requires that the model's output be grounded in its pre-training data, which serves as a proxy for world knowledge. However, because pre-training datasets are enormous, manually checking every generation is impractical. Instead, you can implement strategies at inference time:

• Use retrieval-augmented generation (RAG): Attach a separate knowledge base to your LLM. The model first retrieves relevant facts, then generates output based solely on that retrieved information. This grounds responses in external, verifiable sources.
• Prompt engineering for factual constraints: Include instructions like “Only answer if you are certain the fact is widely accepted. Otherwise, say you don't know.” This forces the model to stay within its training data.
• Fine-tune on factual datasets: Train the model on examples where it must produce grounded answers and avoid fabricated details. Use datasets like Natural Questions or TriviaQA with strict accuracy requirements.

By making factual grounding a priority, you reduce the likelihood of the model inventing content unsupported by its training.

Step 3: Teach the Model to Acknowledge Uncertainty

Equally important is enabling the model to admit when it does not know an answer. Many hallucinations occur because models feel pressured to respond even when uncertain. To combat this:

• Incorporate refusal examples in training: Fine-tune on dialogues where the model gracefully declines to answer, e.g., “I'm sorry, I don't have enough information to answer that accurately.”
• Set confidence thresholds: During inference, use techniques like temperature scaling or logit analysis to detect low-confidence tokens. If confidence falls below a threshold, force the model to output a “don't know” response.
• Prompt explicitly for uncertainty: Add a system message like: “You are a helpful assistant that only answers when you are certain. If unsure, say 'I cannot confirm.'”

This step directly addresses the second requirement for avoiding extrinsic hallucinations: the model must be able to acknowledge ignorance without guessing.

Step 4: Evaluate and Iterate

Regularly test your model using a benchmark that measures both factual accuracy and uncertainty expression. Create a set of questions with known correct answers and some with ambiguous or false premises. Score responses both on correctness and appropriate refusals. Use this feedback to refine your prompts, fine-tuning data, or RAG setup. Over time, you'll reduce extrinsic hallucinations to acceptable levels.

Tips for Success

• Combine multiple strategies: RAG alone may not catch every hallucination; pair it with confidence checks and refusal training for best results.
• Beware of over-reliance on pre-training: Even the best datasets contain errors. Encourage the model to cross-check facts when possible.
• Monitor edge cases: Pay special attention to obscure topics or recently changed facts (e.g., current events) where the model's training data may be outdated.
• Use human review for critical outputs: In high-stakes applications, have a human verify any auto-generated content for hallucinations.
• Stay updated: As LLM research advances, new techniques for reducing hallucinations emerge. Keep learning and adapting your approach.