🤖 AI

Attractive and Repulsive Pattern Control in Sequence Generation

This paper introduces a signed pattern control mechanism for variable-order Markov models that uses belief propagation sampling to either suppress or promote specific recurring patterns, thereby preventing unwanted repetition in sequence generation and enabling the exploration of attractor dynamics.

Original authors: Francois Pachet

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francois Pachet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to write a story or compose a piece of music. You give it a massive library of existing books or songs, and it learns the rules of grammar and melody. When you ask it to continue a sentence or a melody, it usually does a great job. But sometimes, the robot gets stuck in a "tunnel."

Think of this tunnel like a hamster running on a wheel. The robot starts repeating the same phrase, the same musical riff, or the same pattern over and over again. It gets trapped in a loop because, in its own history, that pattern became very popular. Once it starts, it keeps going, creating a boring, repetitive mess.

This paper introduces a clever way to fix that, or even to use that repetition on purpose. The author, François Pachet, calls this "Signed Pattern Control."

Here is how it works, broken down into simple concepts:

1. The "Traffic Cop" with a Scorecard

Usually, the robot just picks the next note or word based on what it has seen before. This paper adds a special "Traffic Cop" (called an automaton) that watches what the robot is doing.

This Traffic Cop doesn't just say "Yes" or "No." Instead, it keeps a scorecard for specific patterns.

The Negative Sign (The "Stop" Signal): If the robot starts repeating a pattern too much (like a hamster on a wheel), the Traffic Cop gives that pattern a negative score. It doesn't ban the pattern entirely, but it makes it "expensive" for the robot to choose. The robot thinks, "Oh, I've used that riff five times already; let's try something else." This stops the robot from getting stuck in a tunnel.
The Positive Sign (The "Go" Signal): Conversely, if you want the robot to repeat a specific pattern (maybe you want a catchy chorus to come back often), you give that pattern a positive score. The Traffic Cop says, "Great! Do that again!" This turns the pattern into a magnet, pulling the robot toward it.

2. The "Homeostatic" Thermostat

The paper tests a specific version of this called Homeostatic Control. Imagine a thermostat in a house. If the room gets too hot, the AC turns on to cool it down. If it gets too cold, the heater turns on.

In the experiment, the robot acts as its own thermostat.

As the robot generates music, it constantly checks: "Am I repeating a specific 8-note melody too much?"
If the answer is "Yes," the system automatically applies the Negative Sign. It gently pushes the robot away from that overused melody.
The result? The robot keeps playing music that sounds like the original style (it still uses the right chords and rhythms), but it stops getting stuck in boring loops. It explores more of the library of possibilities.

3. The "Attractor" Probe

The paper also shows that you can use the Positive Sign as a scientific tool. Instead of just fixing a problem, you can use it to test the robot's limits.

You can tell the robot, "I want you to repeat this specific sad melody as much as possible."
You can watch how the robot reacts. Does it get "obsessed" with that melody? Does it stop playing anything else?
This helps researchers understand how the robot's "mind" works. It's like turning a knob to see how much gravity is pulling the robot toward a specific idea.

4. What They Actually Found

The researchers tested this on classical music (like Bach and Telemann) and Jazz solos.

Without the fix: The robots would often get stuck repeating long, boring chunks of music.
With the "Negative Sign" fix: The robots stopped repeating those long chunks. They used a wider variety of musical ideas. They sounded more natural and less like a broken record.
The Trade-off: To stop the repetition, the robot had to be slightly less perfect at predicting the very next note (a tiny drop in "lower-order support"), but the overall result was much more interesting and less repetitive.

Summary

Think of this paper as giving a creative AI a volume knob for repetition.

Turn the knob to negative, and the AI becomes less likely to get stuck in a loop, keeping the music fresh and varied.
Turn the knob to positive, and the AI becomes obsessed with a specific idea, allowing researchers to study how strong those "habits" can become.

The paper proves that you can control these habits exactly, without breaking the AI's ability to understand the style of the music it is creating. It's a way to keep the robot creative and prevent it from getting bored with its own ideas.

Technical Summary: Attractive and Repulsive Pattern Control in Sequence Generation

Problem Statement
Variable-order Markov (VOM) models are effective for symbolic sequence generation because they preserve local syntax and style while adapting context length. However, over long continuations, these models frequently enter "tunnels": attractor-like corridors characterized by recurring high-order contexts, repeated suffixes, or locally periodic passages. This self-reuse often exceeds the model's formal Markov order, leading to a collapse in diversity. While reducing memory or adding noise can break these repetitions, it simultaneously degrades the structural coherence that makes the generator useful. Previous work (BP-Regular) demonstrated that hard regular constraints (e.g., forbidding chunks longer than a specific length) could be compiled into automata and sampled exactly using Belief Propagation (BP). However, this approach was binary (accept/reject) and lacked the ability to softly modulate the probability of specific patterns.

Methodology
This paper introduces signed pattern control, a mechanism that extends the BP-Regular interface to support soft, adaptive, and symmetric control over recurrence.

Signed Recurrence Field: Instead of a binary automaton that accepts or rejects a continuation, the system employs a weighted finite-state recognizer that computes a non-negative recurrence activation $R(x)$ for a candidate sequence $x$ .
Reweighted Sampling: The sampler draws from a reweighted distribution defined as:
$P_\beta(x) \propto P_0(x) \exp(\beta R(x))$
where $P_0(x)$ $P_{0} (x)$ is the original Markov distribution and $\beta$ $β$ is a coupling scalar.
- Negative Coupling ( $\beta < 0$ ): Penalizes recognized patterns, creating a "homeostatic" regime that discourages overactive repetitions without strictly forbidding them.
- Positive Coupling ( $\beta > 0$ ): Rewards recognized patterns, turning them into controlled attractors.
- Zero Coupling ( $\beta = 0$ ): Recovers the unmodified generator.
Adaptive Homeostasis: In the primary experiments, the target patterns are not fixed but are mined adaptively from the generation history. The system tracks projected $k$ -grams (e.g., pitch classes) across multiple orders ( $k \in \{2, 3, 4, 6, 8\}$ ). Patterns that become overactive in a recent window or over the lifetime of the generation are assigned activation strengths based on a combination of recent frequency and lifetime frequency.
Exact Sampling: The weighted automaton is integrated into the BP factor graph. Because the exponential weighting $\exp(\beta R)$ remains positive, the sum-product BP algorithm can still compute exact normalizing masses and conditional marginals, ensuring the sampling remains mathematically exact despite the soft constraints.

Key Contributions

Unified Signed Control: The paper demonstrates that the same automaton/BP interface used for hard constraints can be repurposed for soft, signed control. This allows a single mechanism to function as both a repulsive force (anti-collapse) and an attractive force (attractor probe).
Homeostatic Anti-Collapse: By dynamically identifying and penalizing overactive patterns, the system reduces long-horizon self-reuse while maintaining the underlying variable-order model's structure.
Attractor Probing: The positive branch provides a controlled method to create and measure attractor basins, phase transitions, and hysteresis in variable-order models.
Separation of Concerns: The approach separates pattern selection (what to control) from the sampling mechanism (how to control), allowing the target patterns to be mined, hand-authored, or externally supplied.

Experimental Results
The mechanism was evaluated on monophonic symbolic music, a domain where recurrence is audible and structurally meaningful. Experiments covered six duration-bearing sources (Bach and Telemann) and five pitch-only jazz solos (Weimar Jazz Database).

Negative Branch (Homeostatic):
- Reduced Self-Reuse: The penalty branch significantly reduced the reuse rate of generated 8-grams (e.g., dropping from 0.140 to 0.064 in Bach core data at 4096 events).
- Increased Diversity: The effective number of generated 8-grams increased (e.g., from 3269 to 3728), and the longest repeated suffix decreased.
- Preserved Support: Lower-order support (2- and 3-grams) remained high, with only a minor drop (1–3 percentage points) in coverage of training-supported contexts.
- Cost: The mechanism incurred a measurable increase in static model loss and computational overhead (roughly 4x slower per event due to adaptive recognizer rebuilding), but maintained exact sampling.
Positive Branch (Attractor Probe):
- Fixed-motif probes demonstrated that specific patterns could be made attractive or repulsive at will.
- Phase-transition sweeps showed that increasing positive $\beta$ leads to a sharp increase in target activation and a corresponding drop in diversity, allowing for the measurement of critical coupling points.
- "Addiction and withdrawal" scenarios confirmed the existence of hysteresis, where the generated context retains attractor properties even after the reward signal is removed.

Significance and Claims
The paper claims to provide a general tool for steering and probing the recurrence landscape of symbolic sequence generators.

For Stabilization: The negative branch offers a way to prevent "tunneling" and collapse in high-order models without sacrificing the local style or requiring a complete retraining of the model. It acts as a homeostatic regulator.
For Exploration: The positive branch transforms the generator into an experimental instrument, allowing researchers to programmatically create attractors, study phase transitions, and test recovery dynamics.
Scope: The authors explicitly limit their claims to the tested variable-order/BP setting with monophonic sequences. They do not claim to solve general musical value or composition quality but rather to provide a precise, measurable mechanism for controlling recurrence pressure. The work connects to broader computational creativity criteria by balancing novelty against typicality through signed energy control.