[Bug]: CPU: a Concat/ReduceMean/Unsqueeze/Tile/Reshape graph returns the right float32 values in the wrong element order #37962

OpenVINO Version

2026.3.1-22476-759c5a6ab8c; also 2025.4.1-20426-82bbf0292c5

Operating System

Other (Please specify in description)

Device used for inference

CPU

Framework

ONNX

Model used

Attached ZIP: ONNX model, inputs, expected outputs, and runner

Issue description

Environment

• OpenVINO version: 2026.3.1-22476-759c5a6ab8c-releases/2026/3 (verified 2026-09-06); first seen on 2025.4.1-20426-82bbf0292c5.
• Build / install: pip wheel openvino==2026.3.1
• OS / Python: Ubuntu 24.04.3 LTS x86_64, Python 3.10
• Reference implementations: onnx.reference.ReferenceEvaluator (onnx 1.17.0), onnxruntime 1.20.1 CPU EP
• Target / device: CPU plugin

Minimal reproducer

27-node float32 ONNX graph (Concat / ReduceMean / Unsqueeze / Tile / Slice / Reshape / Shape), two inputs v1_0: float32[4,6,4] and v0_0: float32[4,5,4], one output float32[32]. Attached as model.onnx + feed.npz (+ expected.npz) with run.py (same 18-line driver as the other reports in this batch).

Expected vs actual

Output float32[32], 24 / 32 elements differ. Full precision, first 24 shown:

• Expected (onnx.reference.ReferenceEvaluator, matched element-for-element by onnxruntime CPU EP) — each distinct value appears twice in a row:

0.27272728085517883, 0.27272728085517883, 0.9090909361839294, 0.9090909361839294,
0.7272727489471436,  0.7272727489471436,  0.6363636255264282, 0.6363636255264282,
0.4545454680919647,  0.4545454680919647,  0.27272728085517883, 0.27272728085517883,
0.9090909361839294,  0.9090909361839294,  0.7272727489471436,  0.7272727489471436,
0.6363636255264282,  0.6363636255264282,  0.4545454680919647,  0.4545454680919647,
0.27272728085517883, 0.27272728085517883, 0.9090909361839294,  0.9090909361839294

• Actual (OpenVINO CPU 2026.3.1) — the same five distinct values, but in block order instead of interleaved:

0.27272728085517883, 0.9090909361839294, 0.7272727489471436, 0.6363636255264282,
0.27272728085517883, 0.9090909361839294, 0.7272727489471436, 0.6363636255264282,
0.4545454680919647,  0.27272728085517883, 0.9090909361839294, 0.7272727489471436,
0.4545454680919647,  0.27272728085517883, 0.9090909361839294, 0.7272727489471436,
0.6363636255264282,  0.4545454680919647,  0.27272728085517883, 0.9090909361839294,
0.6363636255264282,  0.4545454680919647,  0.27272728085517883, 0.9090909361839294

• Difference: max abs 0.6363636553287506, max rel 2.3333333333333335, 24 / 32 elements; largest at index (1,) where the spec has 0.27272728085517883 and OpenVINO has 0.9090909361839294.
• The multiset of values is identical. Every individual ReduceMean result is bit-exact; what differs is which position each one lands in. So this is an index/layout defect in the Unsqueeze → Tile → Reshape flattening, not an arithmetic one.
• The output shape is correct ((32,)), which is why it is not the already-fixed Concat+Reduce shape-leak family.
• Cross-backend: the same program passes on Apache TVM and TensorFlow/XLA.

Root cause (if known)

Not localised. The Tile is not at fault on its own: the isolated patterns Unsqueeze(axis=1) → Tile([1,2]) → Reshape([8]) (expected 1,1,2,2,3,3,4,4) and Unsqueeze(axis=0) → Tile([2,1]) → Reshape([8]) (expected 1,2,3,4,1,2,3,4) are both correct on this same OpenVINO build. The defect needs this graph's combination with Concat, ReduceMean and Slice.

Why this is a bug (not tolerance / not undefined behaviour)

The element ordering of Tile-13 and Reshape-19 is fully specified (row-major, Tile repeating along each axis repeats[i] times), so there is no implementation freedom in which value goes into which position. The bit-exactness of every individual value rules out a tolerance explanation, and the correct output shape rules out a shape-inference issue. Two independent implementations — ONNX's own reference evaluator and onnxruntime — produce the interleaved order.

How found

Found by EquiAutomaton (equivalence-graph differential testing against onnxruntime / torch eager). Depth-0 (the seed graph itself fails). Signature b17301f83d40, 5 independent observations, replayed on 2026.3.1 in a fresh process.

Step-by-step reproduction

Run run.py from the attached ZIP. Full steps and exact outputs are in the issue description.

https://github.com/user-attachments/files/31877191/OV-3-reproducer.zip

Relevant log output

See the exact reference/OpenVINO output comparison in the issue description.

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.