Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference
PLGA is framed as a strict superset of standard attention, but its claimed inference behavior collapses back to a constant-operator form.
The paper says scaled dot-product attention appears exactly when the learned PLGA operator is the identity. Its main inference-collapse theorem says exact input invariance of deductive outputs reduces inference to generalized SDPA with a constant operator. The author reports measured invariance at relative fluctuations of `10^-6` and below, while noting the perturbation bounds do not certify cached inference. Some proof cores are machine-checked in Lean 4. HF Daily Papers' note
The paper says scaled dot-product attention appears exactly when the learned PLGA operator is the identity. Its main inference-collapse theorem says exact input invariance of deductive outputs reduces inference to generalized SDPA with a constant operator. The author reports measured invariance at relative fluctuations of `10^-6` and below, while noting the perturbation bounds do not certify cached inference. Some proof cores are machine-checked in Lean 4. HF Daily Papers' note
score 4