The Anti-Ouroboros Effect: Empirical Evidence for Resilience in AI Feedback Loops

This repository contains the official code, data, and results for the paper, "The Anti-Ouroboros Effect: Empirical Evidence for Resilience in AI Feedback Loops." The research began by validating the "Ouroboros Effect" in a simple classifier and culminated in the discovery of the "Anti-Ouroboros Effect" in a large language model (LLM).

📂 Experiments

This repository contains two primary experiments:

1. Large Language Model (LLM) Summarization Experiment (Gemma 2B)

This is the main experiment detailed in the paper. It uses the google/gemma-2b-it model to test recursive feedback loops on a scientific summarization task (ccdv/arxiv-summarization) over five generations.

• Key Finding: Contrary to the original hypothesis, a simple automated quality filter did not accelerate model collapse. Instead, it induced an "Anti-Ouroboros Effect," leading to a robust improvement in model performance (ROUGE-L score), while the unfiltered control arm showed degradation.
• Location: /llm_gemma_experiment/

2. Simple Classifier Experiment (Digits Dataset)

This initial study simulated the Ouroboros Effect using an ODE model and a simple classifier on the digits dataset.

• Key Finding: This experiment successfully validated the original Ouroboros hypothesis under controlled conditions, showing a coupled decline in model and feedback quality.
• Location: /simple_digits_experiment/

---

🚀 Running the Experiments

LLM Summarization Experiment (Main)

The entire experiment can be replicated by running the Jupyter Notebook. This experiment was conducted on a Tesla P100 GPU.

1. Navigate to the experiment directory:

   cd llm_gemma_experiment

2. Create a virtual environment and install dependencies:

   pip install -r requirements.txt

3. Run the notebook: Open ouroboros_llm_gemma.ipynb in a Jupyter environment and execute the cells. The notebook handles all data loading, model fine-tuning, and evaluation.

Simple Classifier Experiment (Initial Study)

(Instructions for the digits experiment follow here...)

---

🔁 Reproducibility

This project is committed to full reproducibility. All necessary components are publicly available:

• Code: All scripts and notebooks used for the experiments are provided in this repository.
• Dataset: The ccdv/arxiv-summarization dataset is publicly available on the Hugging Face Hub.
• Model: The base model, google/gemma-2b-it, is an open-weight model available on Hugging Face.
• Environment: The exact library versions are specified in the requirements.txt file within each experiment's folder. The main experiment was run using a Tesla P100-16GB GPU.

---

📈 Key Results

The primary result is the discovery of the Anti-Ouroboros Effect in the LLM experiment. The quality filter created a positive feedback loop, improving the model with each generation.

Generation	Control_QM	Ouroboros_QM
0	0.1638	0.1638
1	0.1737	0.1641
2	0.1716	0.1658
3	0.1724	0.1681
4	0.1741	0.1716
5	0.1684	0.1746

(Table data extracted from the notebook run)

---

📜 Citation

If you use the code or findings from this research, please cite the paper:

@article{Adapala2025AntiOuroboros,
  author  = {Adapala, Sai Teja Reddy},
  title   = {The Ouroboros Effect and its Inverse: Empirical
Studies of AI Feedback Loop Dynamics},
}