Exact-match accuracy counts a prediction as correct only when its normalized answer matches the reference answer exactly. It is therefore an all-or-nothing measure: one differing answer string receives no credit. section 4.0 section 4.5.1
A strict answer check
For “When did humans first land on the moon?”, producing “1969” passes. A different answer fails, even when it reflects partial knowledge. In the Pacific Northwest example, “Seattle” is counted correct after retrieval combines population and descriptive evidence. section 4.5.1
Use exact-match accuracy to read several distinct evaluation signals:
- Compare canonical-answer performance rather than partial similarity; F1 supplies partial credit for close or paraphrased answers. section 5.3
- Interpret gains as more frequent production of the expected answer string, not merely vaguely improved wording. section 5.3
- Use it to compare retrieval and generation systems on knowledge-intensive questions, including systems that synthesize multiple passages. section 4.0 section 4.5.1
The metric is demanding but operationally clear. A system can retrieve relevant information yet fail exact match if it cannot synthesize the canonical answer; conversely, a correct paraphrase may receive no credit. section 4.5.1 section 5.3







