| name | stable-self-modulating-quantum-fwps |
| description | Self-Modulating Quantum Fast-Weight Programmers with bounded memory gates for stable sequence processing via variational circuit parameters. |
| category | ai_collection |
| trigger_words | quantum fast-weight programmer, QFWP, bounded memory gate, quantum sequence modeling |
| arxiv_id | 2607.02363 |
Stable Self-Modulating Quantum Fast-Weight Programmers with Bounded Memory Gates
Background
Quantum Fast-Weight Programmers (QFWPs) store temporal information in dynamically programmed variational circuit parameters rather than nonlinear recurrent hidden states.
arXiv: 2607.02363 (July 2026)
Core Concept
Traditional RNNs maintain temporal memory through hidden state vectors. QFWPs encode sequence history in variational circuit parameters:
- Fast-weight programming: Each input dynamically programs variational parameters
- Self-modulation: Input-dependent gates control updates and accumulated state
- Bounded memory: Sign-preserving tanh gate on recurrent memory branch prevents divergence
Architecture
- Update Gate: Controls new information flow into variational parameters
- Forget Gate (Bounded): Sign-preserving tanh prevents unbounded growth
- Variational Circuit: Parameters encode temporal history
Key Findings
- Unbounded multipliers cause divergence in long sequences
- Bounded tanh gating solves the stability problem
- Evaluated on CUDA-Q quantum dynamics forecasting and Milan SMS prediction
Pitfalls
- Unbounded Multiplier Divergence: Old-state multiplier grows without bounds in long sequences
- Circuit Depth: Deep circuits may suffer from barren plateaus
- Parameter Scaling: Parameters scale with sequence length
Applications
- Quantum sequence modeling and time series forecasting
- Quantum-enhanced NLP
- Temporal pattern recognition on quantum hardware
