Answers are grounded in this paper and cite their sources.
Demonstrates better grokking in quantum circuits than classical models, suggesting unique generalization mechanisms.
Grokking---delayed generalization long after memorization---has been observed almost exclusively in classical networks trained with weight decay.We show that 9-layer parameterized quantum circuits (PQCs, 463 parameters) grok modular addition (a+b 23) without weight decay, achieving a 25% grok rate across 20 seeds.A classical transformer (14{,}304 parameters) requires weight decay (λ=1.0) for 100% grokking and fails without it. Weight decay has opposite effects on the two architectures: removing it improves quantum test accuracy ($68.7%$ vs.\ $42.3%$; Mann--Whitney p=0.002) while eliminating classical grokking entirely.Tracking the bipartite entanglement entropy (EE) of the circuit state, we find that successful grokking occurs in a sweet spot (EE ∈ [3.1, 4.1]), while EE overshoot above 4.2 triggers a ``grok-then-ungrok'' phenomenon---test accuracy rises to $93.7%$ then collapses to $11.3%$ as entanglement saturates. Ablations confirm that entanglement is necessary (product-state circuits reach ~0.2%) and that PQCs are more parameter-efficient than size-matched classical models.These results connect quantum unitarity to implicit regularization and entanglement dynamics to generalization transitions.
No takes yet. Share an insight, caveat, or question.
liang wang (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: