Faithfulness Metric

Most faithfulness metrics rely on Natural Language Inference, a task that classifies the relationship between a premise (source) and hypothesis (generated text) into three categories:

• Entailment: The hypothesis is logically supported by the premise
• Contradiction: The hypothesis directly conflicts with the premise
• Neutral: The hypothesis introduces information not addressed in the premise

A faithful output should produce high entailment scores and minimal contradictions. Modern implementations use fine-tuned NLI models like DeBERTa or T5 fine-tuned on MNLI and ANLI datasets.

Advanced faithfulness metrics like FActScore break generated text into discrete, verifiable atomic facts before evaluation. This granular approach enables precise identification of which specific claims are unsupported.

Process:

• Decompose long-form generation into individual factual statements
• Verify each atomic fact against a trusted knowledge base or source document
• Calculate the percentage of supported facts

This method aligns closely with human judgment and provides actionable diagnostics for model improvement.

While often used interchangeably, entailment and factual consistency measure different aspects:

• Entailment: Strict logical deduction — can the hypothesis be derived from the premise?
• Factual Consistency: Broader alignment — are all factual claims in the output supported by the source, even if not strictly entailed?

A response can be factually consistent without being strictly entailed. For example, a summary stating 'The CEO resigned' is factually consistent with a source stating 'The CEO stepped down,' even though the lexical forms differ.

Faithfulness metrics employ different alignment strategies to map generated claims to source evidence:

• Sentence-level alignment: Matches entire generated sentences to source sentences using cosine similarity on embeddings
• Span-level alignment: Identifies specific text spans in the source that support each generated claim
• Cross-attention analysis: Examines internal model attention weights to trace output tokens back to input tokens

Span-level alignment provides the most granular verification and is essential for Citation Precision and Citation Recall calculations in RAG evaluation.

Several specialized benchmarks exist for evaluating faithfulness detection systems:

• TruthfulQA: Tests ability to avoid false answers from common misconceptions
• HaluEval: Contains human-annotated and LLM-generated hallucinated samples across dialogue, QA, and summarization
• FaithDial: Curated dialogue dataset with corrected hallucinated responses from Wizard of Wikipedia
• RAGTruth: Word-level and passage-level hallucination annotations for RAG systems
• XSum Hallucination: Summarization dataset with faithfulness annotations on the XSum corpus

Current faithfulness metrics have known limitations that practitioners must account for:

• Extractive blindness: Metrics may flag paraphrased content as unfaithful if it uses different wording than the source
• NLI model bias: Pre-trained NLI models can inherit biases from training data, misclassifying domain-specific entailment
• Granularity mismatch: Sentence-level evaluation misses intra-sentence hallucinations where one clause is faithful and another is not
• Knowledge cutoff: Verification against static knowledge bases fails when the source contains novel or updated information

Mitigation strategies include domain-adaptive fine-tuning of NLI models and human-in-the-loop calibration.

A core evaluation metric that measures whether all factual claims in a generated text are supported by the source document. While faithfulness focuses on logical deduction, factual consistency specifically checks for alignment between the output and the provided grounding context.

• Detects contradictions between summary and source
• Often evaluated using NLI-based methods
• Critical for summarization and RAG systems

A method for assessing factual accuracy by framing the relationship between a source text (premise) and a generated hypothesis as a Natural Language Inference task. The system classifies the relationship as entailment, contradiction, or neutral.

• Forms the computational backbone of most faithfulness metrics
• Uses models fine-tuned on datasets like MNLI and ANLI
• Provides interpretable, token-level reasoning for verification decisions

The frequency at which a language model generates nonsensical, unfaithful, or factually incorrect text relative to the source material. Expressed as a percentage of total generated tokens or sentences.

• A high-level KPI for production monitoring
• Segmented by severity: critical, major, minor
• Directly correlated with user trust degradation

A quantitative measure of how well a language model's output is anchored to a specific retrieved document or knowledge base entry. Essential in Retrieval-Augmented Generation (RAG) systems to prevent contextual drift.

• Evaluates if the response stays within retrieved context bounds
• Penalizes extraneous information not present in the source
• Often combined with faithfulness for a complete accuracy picture

A zero-resource hallucination detection method that samples multiple responses from a black-box LLM and checks for factual inconsistency. Leverages the principle that hallucinated facts are stochastically unstable across sampling runs.

• Requires no external knowledge base or training data
• Compares semantic similarity across generated samples
• Effective for detecting entity-level and relation-level hallucinations

A prompting technique where an LLM first drafts a response, then generates a series of independent verification questions to fact-check its own work, and finally produces a corrected, verified answer.

• Implements a self-critique loop without external tools
• Reduces hallucination by decomposing verification into atomic checks
• Effective for long-form generation tasks