In-Context Hallucination

A type of LLM hallucination where model output contradicts or is inconsistent with information explicitly provided in the input context, representing a failure in reading comprehension and faithfulness to source material.

Definition

In-context hallucination occurs when language models generate content that:

• Contradicts explicitly provided context information
• Shows inconsistency with source content in the input
• Demonstrates failure in reading comprehension
• Produces outputs that are unfaithful to the given materials

This is distinct from extrinsic-hallucination, where outputs are fabricated and not grounded in world knowledge or training data.

Core Characteristics

According to lilian-weng's taxonomy:

• Context dependence: The evaluation baseline is the provided context, not external knowledge
• Faithfulness requirement: Model outputs should be consistent with source content
• Reading comprehension failure: Represents a fundamental inability to accurately process given information

Evaluation Approach

In-context hallucination is typically easier to evaluate than extrinsic-hallucination because:

• The source material is explicitly provided
• Verification can be done against known context
• No external knowledge retrieval required
• Clear ground truth for comparison

Relationship to Other Concepts

• Complementary to extrinsic-hallucination in the broader hallucination taxonomy
• Critical component of llm-reliability assessment
• Related to factual-accuracy but focused on context consistency
• Important for knowledge-grounding in retrieval-augmented systems

Practical Implications

Understanding in-context hallucination is crucial for:

• Evaluating reading comprehension capabilities
• Developing faithful summarization systems
• Building reliable question-answering applications
• Implementing context-aware generation controls

Detection Methods

Common approaches include:

• Automated consistency checking against provided context
• Human evaluation of faithfulness to source material
• Semantic similarity measurement between output and input
• Contradiction detection using natural language inference

See also

• extrinsic-hallucination
• hallucination
• llm-reliability
• knowledge-grounding
• factual-accuracy