What does FActScore measure and how does it decompose long-form factual evaluation?

**FActScore** is a way to measure how factually accurate a long piece of generated text is, like an LLM-written biography. The problem it solves: when a model writes a paragraph with twenty claims, a single thumbs up or thumbs down throws away most of the signal. You cannot tell whether the model got nineteen facts right and one wrong, or only half. The core idea is decomposition. FActScore breaks the generated text into **atomic facts**, the smallest claims that can each be checked on their own. "She was born in 1945 in Detroit" becomes two atomic facts: born in 1945, and born in Detroit. Next it verifies each atomic fact against a trusted knowledge source, usually Wikipedia. Each fact is labeled supported or not supported. The final score is simple: the fraction of atomic facts that the source supports. If 18 of 20 facts check out, the score is 0.9. This gives a precise, per-claim measure of factual precision instead of one fuzzy holistic rating, and it pinpoints exactly which claims were wrong.

**FActScore** measures long-form factual precision. It decomposes a generation into atomic facts, verifies each against a knowledge source, and reports the fraction supported. The workflow has three steps. First, decompose: split the text into atomic facts, the smallest independently checkable claims. A single sentence often yields several. Second, verify: check each atomic fact against a reference like Wikipedia, labeling it supported or not. Third, aggregate: the score is supported facts divided by total facts. The score is: $$\text{FActScore} = \frac{1}{|A|} \sum_{a \in A} \mathbb{1}[a \text{ supported by } S]$$ The payoff is granularity. A holistic rating collapses twenty claims into one number and hides which ones failed. Atomic decomposition gives per-claim resolution. You see exactly which facts the model invented. Two caveats matter. FActScore is a precision metric, not recall. It scores the facts the model stated, not the ones it omitted. And it only judges claims the knowledge source covers. Facts outside the reference get marked unsupported even when true, so source coverage bounds what you can fairly evaluate.