# The Anti-Ouroboros Effect: Emergent Resilience in Large Language Models from Recursive Selective Feedback

arXiv: https://arxiv.org/abs/2509.10509 · HTML: https://mlsanity.com/paper/2509.10509/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 22 of its 31 references that are in the map.
- First posted on arXiv 2 Sep 2025.

## Builds on (most cited first, 22 of 22)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2307.15217/) (2023)
- [The Curse of Recursion: Training on Generated Data Makes Models Forget](https://mlsanity.com/paper/2305.17493/) (2023)
- [The False Promise of Imitating Proprietary LLMs](https://mlsanity.com/paper/2305.15717/) (2023)
- [Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data](https://mlsanity.com/paper/2404.01413/) (2024)
- [Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task](https://mlsanity.com/paper/2506.08872/) (2025)
- [Reward Shaping to Mitigate Reward Hacking in RLHF](https://mlsanity.com/paper/2502.18770/) (2025)
- [How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse](https://mlsanity.com/paper/2404.05090/) (2024)
- [Scaling Laws of Synthetic Data for Language Models](https://mlsanity.com/paper/2503.19551/) (2025)
- [Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification](https://mlsanity.com/paper/2406.07515/) (2024)
- [Position: Model Collapse Does Not Mean What You Think](https://mlsanity.com/paper/2503.03150/) (2025)
- [Cross-Entropy Is All You Need To Invert the Data Generating Process](https://mlsanity.com/paper/2410.21869/) (2024)
- [Rate of Model Collapse in Recursive Training](https://mlsanity.com/paper/2412.17646/) (2024)
- [BiMark: Unbiased Multilayer Watermarking for Large Language Models](https://mlsanity.com/paper/2506.21602/) (2025)
- [The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking](https://mlsanity.com/paper/2501.19358/) (2025)
- [Improving the Scaling Laws of Synthetic Data with Deliberate Practice](https://mlsanity.com/paper/2502.15588/) (2025)
- [Protecting Human Cognition in the Age of AI](https://mlsanity.com/paper/2502.12447/) (2025)
- [LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs](https://mlsanity.com/paper/2506.15690/) (2025)
- [An Entropy-Based Model for Hierarchical Learning](https://mlsanity.com/paper/2212.14681/) (2022)
- [Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?](https://mlsanity.com/paper/2504.03814/) (2025)

## Abstract (arXiv)

The stability of recursively trained large language models (LLMs) is a foundational problem for AI safety. Prevailing theory predicts model collapse, a progressive degradation when models are trained on their own output. We challenge this narrative by introducing a selective feedback mechanism. Contrary to expectation, instead of merely slowing decay, our experiments provide strong evidence that this pressure reverses it, inducing a statistically significant performance improvement in a Gemma 2B model on a complex summarization task. We name this phenomenon the Anti-Ouroboros Effect. We contrast this with a foundational experiment using a simple classifier, where the theoretical degenerative loop was validated, highlighting the unique dynamics of high-dimensional models. Our findings establish that systemic resilience can be an emergent property of LLMs under simple selection pressure, suggesting a powerful and scalable principle for developing safer and more robust AI systems. Across five generations, a quality-filtered condition improved by 6.6% in ROUGE-L F1 score, whereas an unfiltered control degraded by 3.5% and a random-filter control degraded by 4.2%