SIGNAL
Depth in neural-network alternatives has a hidden prerequisite no optimizer can supply: every layer needs a path back to the original input. Most "we went deeper" claims collapse two failure modes into one, leaving readers unable to tell real scalability from a benchmark brag.
An, Kim, Lee and Joo's 19 July 2026 preprint makes that prerequisite concrete for a specific class, logic-gate-based networks, which swap the weighted sums of standard neural networks for AND/OR/NOT-style operations. The paper splits depth failure into two parts. The first is optimization collapse, where training breaks down as stacks get deeper, a familiar problem that better initialization and gradient tricks can fix. The second is topology-induced: even with stable training, the original input stops reaching the deeper layers, so added depth cannot help. An, Kim, Lee and Joo's fix is structural. Each gate keeps a direct connection to the input, restoring the information path that prior architectures lost. Fixed-width accuracy gains past 100 layers, where prior LGN topologies saturated or degraded, is evidence that depth scaling is an information-access problem wearing an optimization disguise.
The portable test: any "we went deeper" claim in this research line should be judged by whether every layer can still reach the input, not just whether the optimizer is stable.