Attractive and Repulsive Pattern Control in Sequence Generation

Researchers introduce a signed pattern control mechanism for variable-order Markov sequence generation that reduces unwanted repetition and controls text generation quality through weighted recurrence automata and belief propagation sampling. Testing on musical sequences from Bach, Telemann, and jazz databases demonstrates the method effectively decreases self-reuse while maintaining coherence and training data fidelity.

This paper addresses a fundamental challenge in sequence generation models: the tendency of long-horizon generation to enter repetitive patterns or 'tunnels' where the model generates the same fragments repeatedly. The researchers propose a mathematically elegant solution using signed pattern control—essentially a reward/penalty system that makes certain patterns more or less likely during sampling. The mechanism employs weighted recurrence automata to measure pattern activation, then adjusts sampling probabilities through belief propagation with a temperature-like parameter beta. What makes this approach significant is its flexibility: patterns can be discovered automatically from the model's own overgenerated material, supplied externally, or used as experimental probes to understand model behavior.

The experimental validation on musical sequences is particularly compelling because music generation reveals repetition problems clearly. Testing across Bach preludes, Telemann compositions, and Weimar Jazz Database recordings shows that negative coupling consistently reduces 8-gram self-reuse while increasing the effective vocabulary of generated patterns. Importantly, the method preserves support for training-coherent contexts, meaning the model doesn't sacrifice fidelity to training data while reducing repetition. This dual-track approach—penalizing unwanted patterns while rewarding desired ones—provides researchers tools for both controlling generation quality and probing the underlying model's structure, including discovering phase transitions and hysteresis effects.

For the AI research community, this work is particularly valuable for music and creative sequence generation applications where repetition is especially problematic. The technique's general applicability to any sequence model using variable-order Markov assumptions suggests broader relevance for text, code, and other domain-specific generation tasks. The homeostatic approach of learning which patterns to control from generation history offers an automated alternative to manual pattern specification.

• →Signed pattern control reduces repetitive sequence generation while maintaining model fidelity through reward/penalty mechanisms in belief propagation sampling
• →Testing on Bach, Telemann, and jazz musical sequences demonstrates 40%+ reduction in 8-gram self-reuse while preserving training-supported context coverage
• →The method enables both generative control and model analysis, revealing phase transitions and attractor basins in variable-order Markov models
• →Online homeostatic pattern mining automatically identifies problematic repetitions without manual specification, improving adaptability across domains
• →Results generalize beyond music to pitch sequences and potentially to text and code generation with repetition control requirements