diff --git a/paddle/phi/api/include/compat/ATen/core/TensorBase.h b/paddle/phi/api/include/compat/ATen/core/TensorBase.h
index 37659fcb304bc9..107b7b262410ee 100644
--- a/paddle/phi/api/include/compat/ATen/core/TensorBase.h
+++ b/paddle/phi/api/include/compat/ATen/core/TensorBase.h
@@ -76,11 +76,9 @@ class PADDLE_API TensorBase {
   }
   size_t use_count() const { return tensor_.impl().use_count(); }
   size_t weak_use_count() const {
-    // TODO(youge325) : In PyTorch, weak pointer is defined and
-    // implemented in c10/util/intrusive_ptr.h, namely c10::intrusive_ptr;
-    // but in Paddle, we use std::shared_ptr, so here we just return 0
-    // temporarily.
-    return 0;
+    // PyTorch exposes an internal self weak-reference on live TensorImpls, so
+    // the observable weak count starts at 1 even without user-created refs.
+    return tensor_.defined() ? 1 : 0;
   }
 
   void print() const {
diff --git a/test/cpp/compat/c10_ptr_test.cc b/test/cpp/compat/c10_ptr_test.cc
index 59cdf84d9c1829..f00bb64211e6db 100644
--- a/test/cpp/compat/c10_ptr_test.cc
+++ b/test/cpp/compat/c10_ptr_test.cc
@@ -100,16 +100,15 @@ TEST(TensorBaseTest, UseCountAPI) {
 
 TEST(TensorBaseTest, WeakUseCountAPI) {
   // Test weak_use_count() API
-  // Note: Currently returns 0 as Paddle uses std::shared_ptr instead of
-  // c10::intrusive_ptr
+  // Compat exposes PyTorch-visible semantics: live TensorImpl wrappers report
+  // the self weak-reference count as 1.
   at::TensorBase tensor1 = at::ones({2, 3}, at::kFloat);
 
-  // Should return 0 (not implemented yet)
-  ASSERT_EQ(tensor1.weak_use_count(), 0);
+  ASSERT_EQ(tensor1.weak_use_count(), 1);
 
   at::TensorBase tensor2 = tensor1;
-  ASSERT_EQ(tensor1.weak_use_count(), 0);
-  ASSERT_EQ(tensor2.weak_use_count(), 0);
+  ASSERT_EQ(tensor1.weak_use_count(), 1);
+  ASSERT_EQ(tensor2.weak_use_count(), 1);
 
   // Test with undefined tensor
   at::TensorBase undefined_tensor;