Hallucination Taxonomy

An error where the generated output directly contradicts the provided source context or grounding documents. The model fabricates information that is logically inconsistent with the evidence it was given.

• Mechanism: The model overrides retrieved facts with parametric knowledge or statistical biases.
• Example: A source states 'Q2 revenue was $4M,' but the model outputs 'Q2 revenue reached $12M.'
• Mitigation: Faithfulness metrics and constrained decoding can detect and suppress these contradictions.

An error where the generated output contains factual claims that cannot be verified against any provided source or established world knowledge. The model invents entities, events, or citations wholesale.

• Mechanism: The model fills knowledge gaps with plausible-sounding but entirely fabricated tokens.
• Example: Inventing a non-existent research paper titled 'Neural Dynamics of Market Prediction' with a fake DOI.
• Mitigation: Cross-source verification and knowledge graph grounding can flag unsupported claims.

A nuanced error occurring when multiple retrieved sources disagree, and the model synthesizes a conflated or incorrect answer rather than acknowledging the ambiguity.

• Mechanism: The model fails to perform entity disambiguation or temporal grounding across conflicting evidence.
• Example: Two financial reports list different CEOs for the same year; the model picks the wrong one without noting the discrepancy.
• Mitigation: Attribution-aware chunking and source reliability scoring help the model weigh evidence appropriately.

A specific subtype where the model uses outdated or anachronistic information, applying facts from one time period to another without recognizing the temporal shift.

• Mechanism: Failure of temporal grounding; the model does not anchor claims to the correct date range.
• Example: Stating a person holds a position they left three years ago, based on older training data.
• Mitigation: Metadata filtering by recency and explicit date-range constraints in retrieval queries.

A hallucination caused by conflating two distinct entities that share a name or similar attributes, leading to a factually incorrect composite statement.

• Mechanism: The retrieval system or model fails to resolve a textual mention to a unique knowledge base identity.
• Example: Merging the biographies of two different scientists named 'Michael Collins' into a single, inaccurate profile.
• Mitigation: Entity linking pipelines and knowledge graph grounding enforce unique identity resolution.

A classification distinction where a response is loyal to the source but not faithful to reality, or vice versa. Loyalty measures consistency with the provided context; faithfulness measures alignment with objective truth.

• Loyal but Unfaithful: The source document itself contains an error, and the model accurately repeats it.
• Faithful but Disloyal: The model corrects a source error using its own parametric knowledge, violating grounding constraints.
• Mitigation: Groundedness checks must be paired with source reliability scores to handle this trade-off.

The primary axis of the taxonomy. Intrinsic hallucinations are logical contradictions within the model's own generated output. Extrinsic hallucinations are statements that cannot be verified against the provided source context or established world knowledge.

• Intrinsic Example: Stating 'Paris is in Germany' after previously stating 'Paris is the capital of France'.
• Extrinsic Example: Inventing a research paper title that does not exist in the retrieved corpus.
• Operational Impact: Intrinsic errors are detectable via self-consistency checks; extrinsic errors require external verification against a knowledge base.

A quantitative evaluation score measuring the degree to which a generated statement is logically entailed by and consistent with the provided source context. It is the primary automated gauge for extrinsic hallucination.

• Mechanism: Typically uses a Natural Language Inference model to classify each atomic claim as entailed, contradicted, or neutral relative to the source.
• Formula: Faithfulness = (Number of Entailed Claims) / (Total Number of Claims).
• Key Distinction: Faithfulness is independent of general world knowledge; it only cares about fidelity to the provided context, making it ideal for RAG evaluation.

A self-correcting technique where a language model generates an initial response, then systematically drafts and answers a series of independent verification questions to identify and correct its own factual errors.

• Process: (1) Generate baseline response, (2) Plan verification questions, (3) Execute verifications independently, (4) Generate final revised response.
• Taxonomy Link: CoVe is effective at catching both intrinsic contradictions and extrinsic factual errors by forcing the model to re-examine its own claims from a fresh context.
• Efficiency: Reduces hallucination rates without requiring external retrieval, though it is less reliable than grounding in a verified knowledge base.

A constrained text generation strategy that manipulates token probabilities during inference to favor words and phrases explicitly supported by a provided evidence document. It prevents hallucination at the generation level rather than detecting it post-hoc.

• Implementation: Adjusts logits to penalize tokens that would introduce entities or facts absent from the grounding document.
• Taxonomy Application: Directly targets extrinsic hallucination by making it mathematically improbable for the model to generate unsupported claims.
• Trade-off: Can reduce output fluency and creativity, making it suitable for high-stakes factual applications but less ideal for open-ended creative tasks.

An automated evaluation step that compares a generated summary or answer against its source material to identify contradictions, hallucinations, or unsupported inferences. It serves as the operational implementation of a hallucination taxonomy.

• Methods: Includes NLI-based entailment scoring, QA-based fact extraction, and LLM-as-a-judge pairwise comparison.
• Granularity: Can operate at the document level, sentence level, or atomic fact level, with finer granularity providing more precise taxonomic classification.
• Integration: Deployed as a guardrail in production RAG pipelines to filter or flag responses before they reach the end user.

A specific category within extrinsic hallucination where the model generates information that directly contradicts the provided source material. This is distinct from fabricating information absent from the source.

• Example: A source states 'Revenue grew 12%' but the model outputs 'Revenue grew 20%'.
• Root Cause: Often stems from the model's parametric knowledge overriding the retrieved context, a phenomenon known as context-memory interference.
• Mitigation: Addressed through strict grounded decoding and instruction hierarchy that prioritizes provided context over internal weights.