When mixing basic slicing with an advanced index, NumPy moves the advanced index's subspace to the front, so in the example, A[0, :, B] produces a shape of (4, 2) rather than (2, 4).<p><pre><code> import numpy as np
A = np.random.rand(1, 2, 2)
B = np.array([0, 1, 0, 1])
C = A[0, :, B]
print("C.shape:", C.shape)
</code></pre>
So far every LLM I've tried (Grok 3, o1, Gemini Pro) all predict (2, 4) and can't be persuaded otherwise.