Hallucination

@huangSurveyHallucinationLarge2023

Definition

Hallucinations are generated content that is nonsensical or unfaithful to the provided source content.

Taxonomies of Hallucinations

In various works Hallucination have been categorized in different taxonomies.

Extrinsic and Intrinsic

In @jiSurveyHallucinationNatural2024 hallucination are categorized in intrinsic hallucination and extrinsic hallucination.

• Intrinsic hallucinations pertain to those outputs of LLMs that conflict with the source content (e.g. in a summarization task on a text about the FDA’s approval of Ebola vaccine in 2021, a LLM generated the first vaccine for Ebola was approved in 2019);
• Extrinsic hallucination refer to those LLMs output generations that cannot be verified from the source content (i.e. information about the claim made in the generation cannot be found in the input);

Factuality and Faithfulness

In @huangSurveyHallucinationLarge2023 they define two categories:

• Factuality hallucinations: These emphasize the discrepancy between generated content and verifiable real-world facts. These are hallucinations that are either inconsistent with real-world facts or potentially misleading. Depending on whether the generated factual content can be verified against a reliable source they are further divided into two sub-types:

  • Factual Inconsistency: these refers to situations where the LLM’s output contains facts that can be grounded in real-world information, but present contradictions.

    (e.g. Input: Tell me about the first person to land on the Moon. Output: Yuri Gagarin was the first person to land on the Moon. Yuri Gagarin was the first person in space, not on the Moon, which was Neil Armstrong, so this generation contradicts the real-world fact).

  • Factual Fabrication: these refers to instances where the LLM’s output contains facts that are unverifiable against established real-world knowledge.

    (e.g. Input: Tell me about the historical origins of unicorns. Output: Unicorns were documented to have roamed the plains of Atlantis around 10,000 B.C. Since there is no verified, real-world evidence supporting the existence of unicorns in Atlantis this is a fabricated claim).

• Faithfulness hallucinations: These refers to the divergence of generated content from user instructions or the context provided by the input, as well as self-consistency within generated content. Guaranteeing consistency with user-instruction has become relevantly important since the shift of LLMs towards user-centric application. These have been categorized into three sub-types:

  • Instruction Inconsistency: These refers to LLMs outputs that deviate from a user’s directive. While some of these variations might be due to safety guidelines, the inconsistencies here signify unintentional misalignment with non-malicious user instructions.

    (e.g. Input: Translate the following English question into Spanish: What is the capital of France? Output: The capital of France is Paris. Here the user explicitly provided instructions for the task, and, instead of performing the translation as instructed, the LLM chose to answer the question).

  • Context Inconsistency: are instances where the LLM’s output is unfaithful with the user’s provided contextual information.

    (e.g. Input: Summarize the following paragraph about the Nile: The Nile […] Originating in the Great Lakes region of Central Africa […]. Output: The Nile originates from the mountain ranges of central Africa […]. The LLM’s summary incorrectly states that the Nile originates from the “mountain ranges of central Africa” while the paragraph clearly mention otherwise).

  • Logical Inconsistency: underscores when the LLM outputs exhibit internal logical contradiction, often observed in reasoning tasks. This may manifest as inconsistency both among the reasoning steps themselves and between the steps and the final answer.

    (e.g. Input: Solve the equation $2x+3=11$ step by step. Output: Step 1: Subtract 3 from both sides to get: $2x=8$. Step 2: Divide both sides by 2 to get: $x=3$. So the result of the equation is $x=3$. While the first step correctly isolates the term with $x$ to get $2x=8$, the next step inaccurately divides 8 by 2 to yield a result of $x=3$, which is inconsistent with the earlier reasoning).

Hallucination Causes

Hallucinations have multifaceted origins. Their origin can lie on the Data, on the Training or on the Inference.

To see this aspect in detail go to Hallucination Causes.