A frozen neural net reads out how much of Atari is really there

An unlearned feature extractor, with a policy learner trained on top of it via gradient descent, can read out how much of a task is really there. That is the awkward, specific claim inside arXiv:2607.26059, and it lands with numbers concrete enough to test.

A convolutional stack is frozen at initialization, and a policy learner is trained on top. On Pong, the learner ends up driving decisions through as few as five of sixty-four internal neurons. On Breakout, seventeen to twenty-five. On Space Invaders, around forty-two. Train the convolutional stack instead, and the same sixty-four neurons light up across nearly all of them. The frozen stack is forced to express the problem, and the count of dimensions it uses is the dimensionality of the problem itself.

That mechanism may be portable — anywhere raw input dwarfs the structure a task actually has, gradient descent on an unlearned projection may act as a cheap diagnostic for effective rank, with no sparsity penalty required. The arXiv paper's own +18.4 expert plateau versus +14 for a five-neuron Pong seed, with a participation-ratio gap showing the frozen representation is concentrated while the trainable one is diffuse, pins the case to Atari. Whether the same lens holds beyond three games and three Pong seeds is the question the next preprints will have to settle.