Follow the incentive: the assertion that gating networks fundamentally oppose basic methods is a false dichotomy.
Who benefits from presenting these approaches as opposing rather than complementary, if not those selling the "novelty"?
These networks can only be superior if the signals from the basic methods they aggregate are relevant and of quality, otherwise gating merely propagates biases.
For example, a gating network cannot generate better predictions if the underlying LLM produces low-quality textual rationales about a startup, regardless of its dynamic motive.