LLM Recursive Self-Refinement Converges to Stable Textual Fixed Points
A recent study published on arXiv (2607.22653) examines the long-term behavior of recursive self-refinement in large language models (LLMs), where a model revises its own initial draft multiple times. This process is conceptualized as a dynamical system that approaches a soft fixed-point region favored by the model. Utilizing GPT-5.5, the researchers created 10-step refinement paths for 50 abstracts from ICML 2025, employing both default-temperature and deterministic decoding, and also assessed 15 abstracts from ICML 2020. They evaluated normalized edit distance, fixed points, word-count consistency, exponential relaxation, and external LLM evaluations. Results indicate that most edits occur early in the refinement process, suggesting that repeated revisions do not lead to indefinite improvements but rather stabilize, impacting iterative LLM revision workflows.
Key facts
- Study examines recursive self-refinement dynamics in LLMs
- Uses GPT-5.5 to generate 10-step refinement trajectories for 50 ICML 2025 abstracts
- Also evaluates 15 ICML 2020 abstracts
- Analyzes normalized edit distance, fixed points, word-count stability, exponential relaxation
- Refinement trajectories rapidly saturate across all settings
- Most edits occur within first few steps
- Converges toward a model-preferred soft fixed-point region
- Published on arXiv with ID 2607.22653
Entities
Institutions
- arXiv
- ICML