LLM Hallucination Pattern

An LLM Hallucination Pattern is an AI error pattern that is a language model failure mode where large language models generate plausible-sounding information that is factually incorrect or unsupported by training data.

• AKA: LLM Confabulation, Model Hallucination, AI Hallucination.
• Context:
  • It can typically manifest as Factual Errors in generated text.
  • It can typically include Fabricated Citations to non-existent sources.
  • It can typically produce Inconsistent Statements across response segments.
  • It can typically generate Plausible Falsehoods with confident tones.
  • It can typically create Temporal Confusions about event sequences.
  • ...
  • It can often occur in Knowledge-Intensive Tasks requiring specific facts.
  • It can often emerge during Creative Generation with loose constraints.
  • It can often appear in Long-Form Responses with complex reasoning.
  • It can often arise from Training Data Gaps or distribution mismatches.
  • ...
  • It can range from being a Minor LLM Hallucination Pattern to being a Major LLM Hallucination Pattern, depending on its LLM hallucination impact severity.
  • It can range from being a Subtle LLM Hallucination Pattern to being an Obvious LLM Hallucination Pattern, depending on its LLM hallucination detection difficulty.
  • It can range from being a Domain-Specific LLM Hallucination Pattern to being a Cross-Domain LLM Hallucination Pattern, depending on its LLM hallucination scope breadth.
  • It can range from being a Rare LLM Hallucination Pattern to being a Frequent LLM Hallucination Pattern, depending on its LLM hallucination occurrence rate.
  • It can range from being a Correctable LLM Hallucination Pattern to being a Persistent LLM Hallucination Pattern, depending on its LLM hallucination mitigation resistance.
  • ...
  • It can be detected through Fact-Checking Systems and verification methods.
  • It can be mitigated by Retrieval-Augmented Generation and grounding techniques.
  • It can be measured using HaluEval Benchmarks and evaluation metrics.
  • It can be addressed through Prompt Engineering and output filtering.
  • ...
• Example(s):
  • Factual Hallucinations, such as:
    • Historical Event Hallucinations with incorrect dates or false details.
    • Scientific Fact Hallucinations containing wrong formulas or invalid theorems.
    • Biographical Hallucinations featuring fictitious achievements or false attributions.
  • Reference Hallucinations, such as:
    • Citation Hallucinations referencing non-existent papers.
    • URL Hallucinations generating invalid web addresses.
    • Quote Hallucinations attributing fabricated statements.
  • Logical Hallucinations, such as:
    • Mathematical Hallucinations with calculation errors.
    • Reasoning Hallucinations containing logical fallacys.
    • Consistency Hallucinations showing self-contradictions.
  • Entity Hallucinations inventing non-existent organizations or fictional persons.
  • ...
• Counter-Example(s):
  • Accurate LLM Generations, which provide factually correct information.
  • Uncertainty Expressions, which acknowledge knowledge limitations.
  • Creative Fictions, which are intentionally imaginative rather than unintentionally false.
• See: Large Language Model, AI Error Pattern, LLM Power User, Fact-Checking System, Retrieval-Augmented Generation, HaluEval Benchmark, Model Reliability, AI Safety, Prompt Engineering.

---