Overview
Large language models (LLMs) are powerful, but they sometimes produce outputs that are unfaithful, fabricated, or nonsensical—a phenomenon broadly called hallucination. This guide narrows the focus to a specific subtype: extrinsic hallucination, where the model generates content that contradicts verifiable world knowledge or fails to admit when it lacks information. In contrast, in-context hallucination occurs when output contradicts provided context. Extrinsic hallucinations are harder to detect because they rely on the model's pre-training corpus—a proxy for world knowledge—which is too large to check per generation. The goal of this tutorial is to equip you with techniques to make LLMs more factual and honest about their limits.
Prerequisites
Before diving into mitigation strategies, ensure you have a basic understanding of:
- How LLMs are trained (pre-training, fine-tuning, vocabulary).
- Concepts like prompt engineering, context window, and token generation.
- Familiarity with Python for code examples (optional but helpful).
No advanced machine learning expertise is required, but a comfort with high-level architectural ideas will make the guide more accessible.
Step-by-Step Guide to Mitigating Extrinsic Hallucinations
Understanding the Two Types of Hallucination
First, distinguish between in‑context and extrinsic hallucination:
- In‑context hallucination: Output contradicts the source content provided in the prompt (e.g., a summary invents details).
- Extrinsic hallucination: Output contradicts established world knowledge not given in context (e.g., claiming a historical event happened on the wrong date).
Step 1 focuses on building factuality, while Step 3 addresses acknowledging uncertainty.
Step 1: Implement Retrieval-Augmented Generation (RAG)
RAG grounds generation in external, verified knowledge sources rather than relying solely on the model’s pre‑training data. This directly reduces extrinsic hallucination by providing a reliable context.
- Choose a knowledge base: Use a curated set of documents (e.g., Wikipedia dumps, company databases).
- Set up an embedding model: Convert queries and documents into vectors (e.g., using
sentence-transformers). - Implement a retriever: At inference time, retrieve the top‑k documents relevant to the prompt via cosine similarity.
- Feed retrieved content: Prepend or integrate the documents into the LLM’s context, instructing it to answer solely from that material.
Example code (pseudocode):
from transformers import pipeline
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
model = SentenceTransformer('all-MiniLM-L6-v2')
docs = [...] # list of text chunks
doc_embeddings = model.encode(docs)
def retrieve(query, k=3):
query_emb = model.encode([query])
scores = cosine_similarity(query_emb, doc_embeddings)[0]
top_indices = scores.argsort()[-k:][::-1]
return [docs[i] for i in top_indices]
def generate_with_rag(prompt):
retrieved_docs = retrieve(prompt)
context = "\n".join(retrieved_docs)
llm_input = f"Answer based on the provided text:\n{context}\n\nQuestion: {prompt}\nAnswer:"
generator = pipeline('text-generation', model='llama-7b')
return generator(llm_input, max_new_tokens=50)[0]['generated_text']This approach forces the model to stay grounded. Without it, the model might invent “facts” from its training distribution.
Step 2: Apply Confidence Thresholds and Uncertainty Signaling
Even with RAG, the model can hallucinate if it retrieves ambiguous documents. Teach the model to say “I don’t know” when uncertain.
- Logit inspection: Analyze output logits—if the probability of the top token is below a threshold (e.g., 0.3), decline to answer.
- Prompt engineering: Include instructions like “If you are not confident, respond with ‘I am unsure’.”
- Fine‑tuning: Train on examples where the correct answer is “I don’t know” when evidence is insufficient.
Example: Simple logit check
def generate_with_uncertainty(prompt, threshold=0.3):
outputs = model.generate(prompt, return_dict_in_generate=True, output_scores=True)
last_logits = outputs.scores[-1]
probs = torch.softmax(last_logits, dim=-1)
top_prob = probs.max().item()
if top_prob < threshold:
return "I am unsure."
else:
return tokenizer.decode(outputs.sequences[0])Step 3: Fine‑tune on Factual Data with Confidence Markers
Fine‑tuning can shape the model’s internal representations to be more factual and uncertainty‑aware.
- Curate a fact‑only dataset (e.g., from verified sources) and train the model to produce those answers.
- Include negative examples: Show prompts where the answer should be “I don’t know” because the fact is absent.
- Use reinforcement learning from human feedback (RLHF) where human annotators reward factual responses and penalize hallucinations.
This step is resource‑intensive but provides the deepest correction.
Common Mistakes
- Over‑relying on pre‑training data: Assuming the model “knows” everything from training. Extrinsic hallucination often emerges from rare or outdated information.
- Ignoring uncertainty: Pushing the model to always produce an answer, even when it lacks evidence. This encourages fabrication.
- Poor RAG retrieval: Using a low‑quality knowledge base or not filtering retrieved documents can lead to citing wrong sources.
- Setting confidence thresholds too low: A very low threshold (e.g., 0.1) may let hallucinations slip through; too high (e.g., 0.8) may make the model refuse valid answers.
- Neglecting post‑processing: Not verifying the output against the retrieved context. A simple “did the output contradict the context?” check can catch many errors.
Summary
Extrinsic hallucinations in LLMs stem from ungrounded outputs that contradict world knowledge. Mitigation involves a three‑pronged approach: using retrieval‑augmented generation to anchor responses, implementing confidence thresholds to decline unsure answers, and fine‑tuning with factual data and uncertainty markers. Avoid common pitfalls like overtrusting pre‑training or neglecting retrieval quality. By following this guide, you can make your LLM applications more reliable and honest.