diff --git a/tests/math_test.cpp b/tests/math_test.cpp
index 084ea0cf1..773401856 100644
--- a/tests/math_test.cpp
+++ b/tests/math_test.cpp
@@ -1224,36 +1224,23 @@ TEST(math, frexpl) {
 
 TEST(math, modf) {
   double di;
-  double df = modf(123.456, &di);
+  double df = modf(123.75, &di);
   ASSERT_DOUBLE_EQ(123.0, di);
-  // ASSERT_DOUBLE uses more decimals than the double precision when performing
-  // the comparison which can result in false failures. And it seems that modf
-  // results are not 100% precise as expected but within the acceptable delta.
-  // Work around this by tweaking the expected value (taken) from the result of
-  // glibc modf).
-  ASSERT_DOUBLE_EQ(0.45600000000000307, df);
+  ASSERT_DOUBLE_EQ(0.75, df);
 }
 
 TEST(math, modff) {
   float fi;
-  float ff = modff(123.456f, &fi);
+  float ff = modff(123.75f, &fi);
   ASSERT_FLOAT_EQ(123.0f, fi);
-  // See modf comment on why we don't use 0.456f as an excepted value.
-  ASSERT_FLOAT_EQ(0.45600128f, ff);
+  ASSERT_FLOAT_EQ(0.75f, ff);
 }
 
 TEST(math, modfl) {
   long double ldi;
-  long double ldf = modfl(123.456l, &ldi);
-  ASSERT_DOUBLE_EQ(123.0l, ldi);
-  // See modf comment on why we don't use 0.456l as an excepted value when the
-  // modf == modfl. For LP64, where long double != double, modfl algorithm
-  // gives precise results and thus we don't need to tweak the expected value.
-#if defined(__LP64__) || !defined(__BIONIC__)
-  ASSERT_DOUBLE_EQ(0.456l, ldf);
-#else
-  ASSERT_DOUBLE_EQ(0.45600000000000307, ldf);
-#endif // __LP64__ || !__BIONIC__
+  long double ldf = modfl(123.75L, &ldi);
+  ASSERT_DOUBLE_EQ(123.0L, ldi);
+  ASSERT_DOUBLE_EQ(0.75L, ldf);
 }
 
 TEST(math, remquo) {