Phase-Slip: Latent Perturbation for LLMs

The Concept

Neural text generation often collapses into repetitive loops ("degenerate repetition") because the model falls into a deep local minimum of probability—it becomes too confident in its own redundant output. Standard penalties operate on the output (logits), often leading to grammatical fracturing.

Phase-Slip is a stochastic intervention sampler that operates on the internal memory of the model.

1. Monitor: It tracks the Shannon Entropy of the model in real-time.
2. Detect: It identifies "Stagnation States"—periods where entropy drops below a threshold for too long, indicating a loop.
3. Perturb: It injects non-destructive Gaussian noise directly into the Key-Value (KV) Cache.

This "Latent Shock" effectively shakes the model's memory, forcing it to hallucinate a new context and break out of the local minimum.

Installation

For Users

You can install the package directly from PyPI:

pip install phase-slip-sampler

For Developers

If you are cloning this repository for local development or research:

git clone https://github.com/Mmorgan-ML/phase-slip-sampler.git
cd phase-slip-sampler
pip install -r requirements.txt

Note: While the package name is phase-slip-sampler, the Python module is named phase_slip.

Usage

Python Import

To use the sampler in your own code:

import torch
from transformers import GPT2LMHeadModel, GPT2Tokenizer
from phase_slip.sampler import PhaseSlipSampler

model = GPT2LMHeadModel.from_pretrained("gpt2").cuda()
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Initialize the Sampler
# stagnation_threshold: How confident is "too confident"? (Lower = stricter)
# noise_scale: Magnitude of the memory shock (Higher = more chaotic)
sampler = PhaseSlipSampler(
    model, 
    tokenizer, 
    stagnation_threshold=0.6, 
    patience=3,
    noise_scale=0.1
)

text = sampler.generate("The scientific method is", max_new_tokens=50)
print(text)

Running the Demo

To see the "Thermal Shock" in action and visualize the divergence from greedy decoding:

python demo.py

Benchmarking

To statistically compare Phase-Slip against Greedy Decoding and Standard Sampling:

python benchmark.py

Empirical Evidence

Benchmarks performed on gpt2 (Small) demonstrate that Phase-Slip effectively shatters repetition loops, achieving higher vocabulary diversity than even standard temperature sampling.

1. The "Loop Breaker" Test

Prompt: "The research paper described the finding that the"

2. Vocabulary Diversity Score (n=5 rounds)

Score based on unique word count ratio. Higher is better.

Data collected via benchmark.py on 2025.12.03.

Calibration & Limitations

This method balances Repetition against Coherence.

• Noise Scale: Controls the magnitude of the "Shock."
  • Low (0.05): Subtle nudges. Keeps grammar intact but might not break strong loops.
  • High (0.15+): Strong divergences. Can lead to "dream-like" or nonsensical transitions (e.g., breaking syntax).
• Stagnation Threshold: Controls the trigger sensitivity.
  • 0.6 is a recommended starting point for GPT-2/Llama.
  • Setting this too high (e.g., 0.8) will trigger shocks constantly, leading to incoherence.

Project Structure

• phase_slip/: The source code for the sampler.
  • sampler.py: Contains the latent_perturbation logic.
• demo.py: A visual comparison script.
• benchmark.py: A statistical tool to measure vocabulary diversity.

License

MIT