diff --git a/include/onnxruntime/core/session/onnxruntime_cxx_api.h b/include/onnxruntime/core/session/onnxruntime_cxx_api.h
index c4a3ec047e..ff73f3d93d 100644
--- a/include/onnxruntime/core/session/onnxruntime_cxx_api.h
+++ b/include/onnxruntime/core/session/onnxruntime_cxx_api.h
@@ -179,9 +179,9 @@ struct Float16_t : onnxruntime_float16::Float16Impl<Float16_t> {
   Float16_t() = default;
 
   /// <summary>
-  /// Explicit conversion to uint16_t representation of bfloat16.
+  /// Explicit conversion to uint16_t representation of float16.
   /// </summary>
-  /// <param name="v">uint16_t bit representation of bfloat16</param>
+  /// <param name="v">uint16_t bit representation of float16</param>
   /// <returns>new instance of Float16_t</returns>
   constexpr static Float16_t FromBits(uint16_t v) noexcept { return Float16_t(v); }
 
@@ -192,9 +192,9 @@ struct Float16_t : onnxruntime_float16::Float16Impl<Float16_t> {
   explicit Float16_t(float v) noexcept { val = Base::ToUint16Impl(v); }
 
   /// <summary>
-  /// Converts bfloat16 to float
+  /// Converts float16 to float
   /// </summary>
-  /// <returns>float representation of bfloat16 value</returns>
+  /// <returns>float representation of float16 value</returns>
   float ToFloat() const noexcept { return Base::ToFloatImpl(); }
 
   /// <summary>
diff --git a/include/onnxruntime/core/session/onnxruntime_float16.h b/include/onnxruntime/core/session/onnxruntime_float16.h
index 6a16f685ca..0b066a9cc9 100644
--- a/include/onnxruntime/core/session/onnxruntime_float16.h
+++ b/include/onnxruntime/core/session/onnxruntime_float16.h
@@ -310,7 +310,16 @@ inline float Float16Impl<Derived>::ToFloatImpl() const noexcept {
     o.f -= magic.f;           // re-normalize
   }
 
-  o.u |= (val & 0x8000) << 16;  // sign bit
+  // Attempt to workaround the Internal Compiler Error on ARM64
+  // for bitwise | operator, including std::bitset
+#if (defined _MSC_VER) && (defined _M_ARM || defined _M_ARM64 || defined _M_ARM64EC)
+  if (IsNegative()) {
+    return -o.f;
+  }
+#else
+  // original code:
+  o.u |= (val & 0x8000U) << 16U;  // sign bit
+#endif
   return o.f;
 }