Recursive LLM Editing Rapidly Saturates: Textual Relaxation in Scientific Writing

Large language models increasingly participate in scientific authorship by repeatedly revising drafts. Yet the behavior of this common writing workflow is poorly characterized: does continued revision produce continued improvement, or does it drive a manuscript toward a model-preferred textual form? We study recursive editing as a stochastic dynamical process in which repeated LLM revision drives scientific text toward a low-change region. Using GPT-5.5, we generate ten-step refinement trajectories for 50 ICML 2025 abstracts under API-default and temperature-$0$ sampling, with 15 ICML 2020 abstracts as a cross-year comparison. We characterize the trajectories using normalized edit distance, exact-repeat and conservative terminal-stability statistics, approximate convergence, sensitivity analysis, exponential relaxation, and an external LLM-as-a-judge evaluation. Across the evaluated settings, trajectories saturate rapidly. At the primary thresholds, all observed trajectories meet a sustained approximate-convergence rule; rates remain 80%, 98%, and 100%, respectively, when the edit threshold is reduced fourfold. Requiring the final two transitions to be exact yields terminal-stability rates of 24% for ICML 2025 API-default, 86% for ICML 2025 temperature-$0$, and 80% for ICML 2020 temperature-$0$. The average edit magnitude $\Delta_t=d(V_t,V_{t+1})$ follows an exponential relaxation pattern. In ten exploratory, order-visible comparisons, an external judge selects the final labeled version in all cases. The findings show sharply diminishing edits under this operator and motivate auditable pause criteria for AI-assisted writing. They do not establish scientific improvement, absorbing fixed points, cross-document homogenization, or corpus-level feedback effects.