Bounded memory gates stabilize self-modulating quantum fast-weight programmers

Researchers propose a bounded old-state modulation rule for Self-Modulating Quantum Fast-Weight Programmers (QFWPs) to address divergence issues in long-sequence regimes. The method applies a sign-preserving tanh gate exclusively to the recurrent memory branch while leaving additive and new-update modulations unchanged.

• Evaluated on two CUDA-Q quantum-dynamics forecasting tasks and Milan SMS telecommunication activity prediction.
• Old-state modulation identified as the primary source of improvement over Standard QFWP.
• Bounding the old-state gate removes long-sequence divergence and improves aggregate robustness.
• Original unbounded Self-Modulating QFWP shows clear gains at longer input windows on Milan SMS data.

The findings identify accumulated-memory modulation as the key mechanism of Self-Modulating QFWP and bounded gating as a targeted stabilization strategy.