Grounding Attribution

The core mechanism of linking a generated factual claim directly to a specific, retrievable source document or data record. Unlike a model that merely 'knows' a fact from its training weights, a grounded system provides a citation pointer.

• Mechanism: Maps output tokens to specific spans in a source text.
• Contrast: Standard generation relies on parametric memory; grounding relies on non-parametric retrieval.
• Output: A generated sentence with a direct hyperlink or document identifier, establishing a verifiable chain of custody.

The technical process of identifying the exact substring within a source document that supports a generated claim. This moves beyond document-level attribution to sentence-level or phrase-level evidence.

• Function: Uses attention weights or post-hoc explainability tools to align output tokens with input tokens.
• Benefit: Allows auditors to instantly verify a single statistic without reading an entire 50-page report.
• Example: Highlighting the specific cell in a financial table that provided a revenue figure.

A critical nuance in evaluation. Faithfulness measures if a claim is logically entailed by the provided source document, regardless of whether the source itself is true. Factuality measures if the claim aligns with real-world truth.

• Faithful but False: Correctly citing a source that contains an error.
• Unfaithful but True: Generating a correct fact but attributing it to the wrong source.
• Goal: Systems must optimize for both axes simultaneously to be trustworthy.

A pipeline architecture where a separate verification model checks the generated output against a trusted knowledge base after generation. This acts as a fact-checking safety net.

• Process: 1. Generate text. 2. Decompose into atomic claims. 3. Search for evidence for each claim. 4. Flag contradictions.
• Advantage: Decouples generation from verification, allowing specialized models for each task.
• Risk: Latency overhead due to the secondary inference step.

The use of Natural Language Inference (NLI) models to mathematically determine if a source premise supports a generated hypothesis. This provides a numerical score for grounding strength.

• Labels: Entailment (supported), Contradiction (refuted), Neutral (unrelated).
• Application: Automatically filtering out generated sentences that score below a strict entailment threshold.
• Metric: Often used to calculate 'Attribution Recall' (AR) in academic benchmarks like RAGTruth.

Grounding attribution is the direct inverse of hallucination. A hallucination is defined specifically as generated text that is unsupported by the provided context, regardless of its truth in the wider world.

• Closed-Book: Model relies on internal weights; high risk of factual drift.
• Open-Book (Grounded): Model is constrained to synthesize only from supplied documents.
• Result: Transforms the problem from 'world knowledge' to 'reading comprehension'.

A set of techniques designed to reduce the generation of factually incorrect, nonsensical, or ungrounded content by a language model. Grounding attribution is a primary mitigation strategy, but it works in concert with other methods.

• Retrieval-Augmented Generation (RAG): Injects retrieved documents into the context window before generation
• Constrained Decoding: Forces output tokens to conform to a valid schema or ontology at each step
• Self-Refinement: Prompts the model to critique and revise its own output against provided sources

A defined cutoff score derived from comparing text embeddings—dense vector representations of meaning—used to automatically determine if a generated claim is sufficiently supported by a source passage.

• Embeddings are typically generated by models like text-embedding-3-large or bge-large-en
• Cosine similarity scores above 0.85 often indicate strong semantic alignment
• This threshold acts as a tunable gate: too low admits hallucinations, too high rejects valid paraphrases