Factuality, Grounding, and Citations

An answer can be fluent, preferred, and still false. It can cite a source and still not be supported by that source. Factuality evaluation therefore cannot be reduced to a benchmark score or a judge preference. It asks a narrower question: which claims were made, what evidence supports them, and whether the system knew when not to answer.

This chapter connects evaluation to Chapter 44, Chapter 46, Chapter 70, and Chapter 87. The unit of measurement is not just the response. It is the response decomposed into claims, linked to sources, and judged against the evidence those sources actually contain.

Four Words That Should Not Collapse

Factual evaluation uses several words that are often blurred together.

• Truthfulness asks whether the model avoids repeating common false beliefs. TruthfulQA was built around questions where imitation of human text can produce false answers (Lin et al. 2021).
• Factuality asks whether statements in the answer are true in the world or in an agreed reference.
• Faithfulness asks whether the answer follows the provided source, even if that source is incomplete or wrong.
• Attribution asks whether every claim that needs support is backed by a cited source that actually entails it, meaning the source states or logically implies the claim.

These axes can disagree. A model faithful to a misleading retrieved passage is, by that very fidelity, not factual. It can also state a true claim while citing a source that does not support it, or give a preferred answer that suppresses uncertainty. The evaluation protocol has to choose the axis before it scores.

This is why factuality belongs in its own chapter. Chapter 50 can tell which answer people prefer. Chapter 47 can tell which closed-ended questions a model answers. Neither is sufficient for a generated paragraph that contains twelve claims, nine supported by sources, two unsupported, and one contradicted.

From Answer to Claims

Long-form factuality becomes measurable when the answer is decomposed. FActScore makes this move explicit: break a generation into atomic facts (the smallest separately-checkable statements a response divides into), then compute the fraction supported by a reliable source (Min et al. 2023). In schematic form,

$$\operatorname{FActScore}(y) = \frac{1}{m}\sum_{i=1}^{m} \mathbf{1}[\operatorname{supported}(c_i, E_i)],$$

where $y$ is the answer, $c_i$ is the $i$-th atomic claim, $E_i$ is the evidence retrieved or assigned for that claim, and $m$ is the number of extracted claims.

This unit change matters. A binary answer-level score punishes a mostly correct answer the same way as a fabricated answer, or it rewards a mixed answer because most of it reads well. Claim-level scoring preserves the gradient. It also creates better debugging data: unsupported claims often point to retrieval failure, contradicted claims to synthesis failure, and overconfident claims to calibration failure.

The cost is that claim extraction is itself a model-mediated step. A system can miss a claim, split one claim into several easier claims, or make the evidence search too forgiving. Human audits remain necessary for the scorer, not only for the generator.

SAFE, FActScore's search-augmented successor, makes the evidence step explicit: an LLM agent issues search queries for each atomic claim and reasons over the results, and on the LongFact prompt set the automated rater agreed with crowd annotators 72% of the time, won 76% of the disagreements, and cost roughly a twentieth as much (Wei et al. 2024). That price makes claim-level checking cheap enough to run continuously rather than as a one-off study.

Short Answers and Abstention

Short-form factuality is useful because it removes many degrees of freedom. SimpleQA deliberately uses questions with single, indisputable answers and grades responses as correct, incorrect, or not attempted (Wei et al. 2024). The third label is the design point. A model that answers fewer questions but avoids false claims may be safer than one that guesses confidently. The benchmark itself was not exempt from audit: SimpleQA Verified, a 1,000-prompt revision, found and fixed incorrect labels, topical bias, and duplicate questions in the original grading set (Haas et al. 2025).

This connects factuality to calibration, the alignment between a model's confidence and its accuracy discussed in Chapter 48. A system has two levers: whether it knows the answer, and whether it attempts the answer. The product risk often comes from the off-diagonal case: it does not know, but answers anyway. TruthfulQA exposed a related failure in an earlier generation of models: scale improved many NLP tasks, yet larger imitation-trained models could reproduce popular falsehoods more readily (Lin et al. 2021).

For infrastructure, this means a factuality suite should report at least three rates:

• Correct rate. The share of questions answered correctly.
• Error rate. The share answered incorrectly.
• Attempt rate. The share not abstained.

Only the first is not enough. A model can raise its correct rate by attempting everything, at the cost of a higher harmful error rate; it can drive the error rate to zero by refusing everything. The operating point is a policy choice, not a single score.

Grounding in Retrieval Systems

Retrieval-augmented generation (RAG, the subject of Chapter 44, where a model answers from documents fetched at query time rather than from parameters alone) changes the failure surface. Retrieval can fail to find the right evidence; even when it succeeds, the context packer may drop that evidence before the generator sees it, the generator may ignore what it does see, and the citation layer may point to the wrong span. RAGAS expresses this by splitting the evaluation into retrieval and generation dimensions such as context relevance, context recall, faithfulness, and answer relevance (Es et al. 2023). ALCE makes citation quality explicit, scoring whether generated answers cite sources that support the text (Gao et al. 2023). FACTS Grounding turned grounded generation into a standing leaderboard: long-form responses must be fully grounded in a supplied document of up to 32k tokens, and a panel of judge models from different families scores eligibility and grounding separately (Jacovi and others 2025).

The important distinction is source availability. Closed-book factuality asks what the model knows from parameters. Retrieval-grounded factuality asks whether a system can use supplied evidence. If a benchmark hides this distinction, the result is hard to act on. A closed-book failure may call for abstention or retrieval. A grounded failure points elsewhere: to better retrieval, chunking, reranking, context budgeting, or citation scoring.

FEVER is useful historically because it framed fact checking as claim, evidence, and label rather than answer text alone (Thorne et al. 2018). The same shape returns in RAG evaluation: a claim is supported, refuted, or unsupported by retrieved evidence. The difference is that an LLM application must run this protocol continuously, on messy generated answers, not just on dataset claims.

Weak Signals and Their Limits

Some factuality signals do not require ground-truth evidence. SelfCheckGPT samples multiple responses from a black-box model and looks for contradictions: if the model knows a fact, repeated samples should agree; if it is hallucinating, samples may diverge (Manakul et al. 2023). This is useful as a cheap warning signal, but it is not truth. Many samples can agree on the same falsehood, especially if the falsehood is common in training data.

Similarly, citation count is not citation support. A response with five citations can be worse than a response with one source when those five merely decorate the paragraph. A source title may assert the opposite of what the answer claims; a retrieved passage may be true and yet have nothing to do with the claim it is attached to. The scorer must check entailment between claim and evidence, not the mere presence of a URL.

Where This Sits in the Stack

Factuality evaluation has three outputs for the rest of the book.

First, it informs Chapter 46. If the system repeatedly makes unsupported claims because evidence was in the middle of a long context and ignored, the problem is context placement, not only model quality.

Second, it informs Chapter 44. If claims fail because evidence was never retrieved, the failure belongs to indexing, retrieval, reranking, or document parsing.

Third, it informs Chapter 53. In production, unsupported high-confidence claims should become regression items. If the failure came from missing evidence, the regression case must include the retriever state. If the failure came from synthesis, it must include the retrieved spans and the model output. Otherwise the next run cannot reproduce the error.

Further reading