diff --git a/src/lighteval/metrics/utils/extractive_match_utils.py b/src/lighteval/metrics/utils/extractive_match_utils.py
index da0d718d7..519369a0f 100644
--- a/src/lighteval/metrics/utils/extractive_match_utils.py
+++ b/src/lighteval/metrics/utils/extractive_match_utils.py
@@ -103,9 +103,10 @@ def lazy_expr_regex(expr_config: ExprExtractionConfig, language: Language) -> li
 
     # Basic number patterns (no LaTeX)
     number_re = (
-        # Format 1: Numbers with thousand separators (e.g., "1,234.56" or "1 234.56")
+        # Format 1: Numbers with thousand separators (e.g., "1,234.56" or "1 234.56"). The space can also be a no-break,
+        # narrow no-break or thin space (U+00A0, U+202F, U+2009). The last group must not be followed by another digit.
         r"(?:"
-        r"(?P<integer1>-?\d{1,3}(?:[ ,]\d{3})+)(?P<decimal1>\.\d+)?|"
+        r"(?P<integer1>-?\d{1,3}(?:[ ,\u00a0\u202f\u2009]\d{3})+(?!\d))(?P<decimal1>\.\d+)?|"
         # Format 2: Simple numbers with decimal point or comma (e.g., "123.45" or "123,45")
         r"(?P<integer2>-?\d+)(?P<decimal2>[.,]\d+)|"
         # Format 3: Decimal part only (e.g., ".123")
@@ -430,7 +431,7 @@ def extract_expr(match: re.Match, timeout_seconds: int) -> tuple[str | sympy.Exp
 
     if integer or decimal:
         # This makes sure we can convert numbers like 0001 to 1. Do note that this can convert 0 to '', so we assume an empty string was 0 and convert it back afterwards.
-        integer = integer.translate(str.maketrans("", "", ", ")).lstrip("0")
+        integer = integer.translate(str.maketrans("", "", ", \u00a0\u202f\u2009")).lstrip("0")
         if len(integer) == 0:
             integer = "0"
 
diff --git a/tests/unit/metrics/test_extractive_match.py b/tests/unit/metrics/test_extractive_match.py
index d2fd71606..709e5333f 100644
--- a/tests/unit/metrics/test_extractive_match.py
+++ b/tests/unit/metrics/test_extractive_match.py
@@ -1146,3 +1146,26 @@ def test_math_extraction_additional_cases(gold, pred, expected):
 )
 def test_math_numina_cases(gold, pred, expected):
     assert compare_strings(gold, pred, match_types=["latex", "expr"]) == expected
+
+
+@pytest.mark.parametrize(
+    "gold,pred,expected",
+    [
+        # No-break, narrow no-break and thin spaces are digit-group separators, like the ASCII space and the comma
+        ("2500", "ANSWER: 2\u202f500", 1),
+        ("2", "ANSWER: 2\u202f500", 0),
+        ("2500", "ANSWER: 2\u00a0500", 1),
+        ("2", "ANSWER: 2\u00a0500", 0),
+        ("2500", "The total is 2\u2009500.", 1),
+        ("500", "The total is 2\u2009500.", 0),
+        ("1250000", "ANSWER: 1\u202f250\u202f000", 1),
+        ("1234.5", "ANSWER: 1\u202f234.5", 1),
+        ("2500", "ANSWER: 2 500", 1),
+        # A group that is not exactly three digits is not joined
+        ("7", "ANSWER: 7\u202f1234", 1),
+        ("7", "ANSWER: 7 1234", 1),
+        ("12", "ANSWER: 12\u202fapples", 1),
+    ],
+)
+def test_typographic_digit_group_spaces(gold, pred, expected):
+    assert compare_strings(gold, pred, match_types=["latex", "expr"]) == expected