Self-Reinforcing Repetition in Autoregressive Models

This direction is merged into another direction: neural text degeneration. Its last version stays readable and citable here.

Version 1 of 117 papers32 conceptsfrozen 2026-10-04

Large language models (LLMs) are prone to a failure mode where they repeat the same phrase or sentence indefinitely. While this is often dismissed as a "hallucination" or a lack of training data, recent research reveals it is a fundamental byproduct of the autoregressive objective. When a model generates a token, that token becomes part of its own future context, creating a feedback loop that can collapse the model's probability distribution into a cycle.

The Self-Reinforcement Effect

The core mechanism behind neural text degeneration is a "self-reinforcement effect." In a standard Transformer, the probability of a token is conditioned on the preceding sequence. Jin Xu and Jianhao Yan demonstrate in Learning to Break the Loop that this probability increases monotonically with each subsequent repetition.

Once a sentence is repeated even once, the model’s confidence in repeating it a third time spikes. This creates a "probability ceiling" where the repetitive tokens become so dominant that the model cannot mathematically justify choosing a diverse alternative. Muhammed Emrullah Ildiz formalizes this in From Self-Attention to Markov Models, showing that single-layer attention naturally behaves like a context-conditioned Markov chain that is prone to distribution collapse.

Why Maximization Fails

Common decoding strategies like greedy search or beam search exacerbate this issue. These methods prioritize the highest-probability tokens at each step. Ari Holtzman and Yejin Choi argue in The Curious Case of Neural Text Degeneration that human language is not a sequence of maximum-probability tokens.

By always picking the most "likely" word, models enter a feedback loop where the repetition of a phrase increases its predicted probability for the next step. This is why metrics like Perplexity (PPL) can be misleading; as Yequan Wang notes in Perplexity from PLM Is Unreliable, a model may assign a very "good" (low) perplexity score to a repetitive loop simply because the repetition makes the next token highly predictable.

Repetition as a Double-Edged Sword

Interestingly, the same mechanism that causes loops also enables the "magic" of in-context learning (ICL). In Understanding In-Context Learning from Repetitions, Jianhao Yan suggests that ICL is essentially a sophisticated form of the model's urge to repeat itself. The model treats the patterns in a prompt as templates to be mirrored. When this works, the model follows instructions; when it fails, it enters an infinite loop.

Breaking the Cycle

Researchers are exploring several ways to mitigate this:

  • Dynamic Truncation: Nucleus Sampling avoids the "long tail" of low-probability noise while preventing the model from getting stuck in the high-probability peaks of repetitive loops.
  • Training Interventions: The DITTO method proposed by Jin Xu uses pseudo-repetitive data to teach the model to recognize and penalize its own looping behavior during training.
  • Preference Optimization: In multimodal contexts, Fei Zuo uses multi-conditional preference optimization to correct the internal probability distributions that lead to repetition hallucinations.

Despite these fixes, the problem persists in advanced reasoning models. Charilaos Pipis observes in Wait, Wait, Wait... Why Do Reasoning Models Loop? that even models like DeepSeek-R1 can get stuck when faced with difficult reasoning tasks, suggesting that repetition remains a primary "escape hatch" for models when they encounter high-uncertainty scenarios.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_self_reinforcing_repetition_in_autoregre_1401c496,
  title = {Self-Reinforcing Repetition in Autoregressive Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/self-reinforcing-repetition-in-autoregressive-models/rdr_1401c4964dcd4daeb0f52407852b0dac},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_self_reinforcing_repetition_in_autoregre_1401c496_v1,
  title = {Self-Reinforcing Repetition in Autoregressive Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/self-reinforcing-repetition-in-autoregressive-models/rdr_1401c4964dcd4daeb0f52407852b0dac/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}