diff --git a/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp b/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp
index 99dc8a89293..46f59bba89f 100644
--- a/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp
+++ b/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp
@@ -9,6 +9,8 @@
 #include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
+#include <cstring>
+
 namespace torch {
 namespace executor {
 namespace native {
@@ -171,6 +173,13 @@ Tensor& max_pool2d_with_indices_backward_out(
 
   static constexpr auto name = "max_pool2d_with_indices_backward.grad_input";
 
+  // max_pool_backward_impl scatter-adds into grad_input (`grad_input_ptr[maxindex] += ...`), writing
+  // only the argmax positions. resize_tensor does not clear the buffer and the memory planner recycles
+  // arena allocations across ops and iterations, so every other element would accumulate onto stale
+  // data. ATen zeroes gradInput before dispatching the identical loop
+  // (aten/src/ATen/native/DilatedMaxPool2d.cpp).
+  memset(grad_input.mutable_data_ptr(), 0, grad_input.nbytes());
+
   ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
     max_pool_backward_impl<CTYPE, false>(grad_input, grad_output, indices);
   });