Alexander Doria proposes training data shape as a cause of superposition in Qwen-3 and Kimi

Alexander Doria says interpretability analyses of modern models overlook data shape as a likely cause of superposition, the phenomenon where a model encodes multiple features in overlapping directions.

In an X post, he argues that Qwen-3 and Kimi, which were trained on large amounts of clean synthetic data and environments, may have their features and superposition shaped by that training data.

The claim is a hypothesis from one source, and the post does not include experiments or measurements that test it.